Inter-area Oscillations in Power Systems
Power Electronics and Power Systems
Series Editors:Series Editors: M. A. PaiUniversity of Illinois at Urbana-ChampaignUrbana, Illinois
Alex StankovicNortheastern UniversityBoston, Massachusetts
Inter-area Oscillations in Power Systems: A Nonlinear and Nonstationary PerspectiveArturo Roman Messina, ed.ISBN 978-0-387-89529-1
Robust Power System Frequency ControlHassan BevraniISBN 978-0-387-84877-8
Synchronized Phasor Measurements and Their ApplicationsA.G. Phadke and J.S. ThorpISBN 978-0-387-76535-8
Digital Control of Electical DrivesSlobodan N. VukosavicISBN 978-0-387-48598-0
Three-Phase Diode Rectifiers with Low HarmonicsPredrag PejovicISBN 978-0-387-29310-3
Computational Techniques for Voltage Stability Assessment and ControlVenkataramana AjjarapuISBN 978-0-387-26080-8
Real-Time Stability in Power Systems: Techniques for Early Detection of the Risk of BlackoutSavu C. Savulesco, ed.ISBN 978-0-387-25626-9
Robust Control in Power SystemsBikash Pal and Balarko ChaudhuriISBN 978-0-387-25949-9
Applied Mathematics for Restructured Electric Power Systems: Optimization, Control, and Com-putational IntelligenceJoe H. Chow, Felix F. Wu, and James A. Momoh, eds.ISBN 978-0-387-23470-0
HVDC and FACTS Controllers: Applications of Static Converters in Power SystemsVijay K. SoodISBN 978-1-4020-7890-3
Power Quality Enhancement Using Custom Power DevicesArindam Ghosh and Gerard LedwichISBN 978-1-4020-7180-5
Computational Methods for Large Sparse Power Systems Analysis: An Object Oriented ApproachS.A. Soman, S.A. Khaparde, and Shubha PanditISBN 978-0-7923-7591-3
Continued after Index
Arturo Roman MessinaEditor
Inter-area Oscillationsin Power Systems
ANonlinear andNonstationary Perspective
1 3
Editor
Arturo Roman MessinaCentro de Investigacion yde Estudios Avanzadosdel IPN
Guadalajara, [email protected]
ISBN 978-0-387-89529-1 e-ISBN 978-0-387-89530-7DOI 10.1007/978-0-387-89530-7
Library of Congress Control Number: 2008939222
# Springer ScienceþBusiness Media, LLC 2009All rights reserved. This workmay not be translated or copied in whole or in part without the writtenpermission of the publisher (Springer ScienceþBusinessMedia, LLC, 233 Spring Street, NewYork,NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use inconnection with any form of information storage and retrieval, electronic adaptation, computersoftware, or by similar or dissimilar methodology now known or hereafter developed is forbidden.The use in this publication of trade names, trademarks, service marks, and similar terms, even if theyare not identified as such, is not to be taken as an expression of opinion as to whether or not they aresubject to proprietary rights.
Printed on acid-free paper
springer.com
Preface
The study of complex dynamic processes governed by nonlinear and
nonstationary characteristics is a problem of great importance in the
analysis and control of power system oscillatory behavior. Power system
dynamic processes are highly random, nonlinear to some extent, and
intrinsically nonstationary even over short time intervals as in the case of
severe transient oscillations in which switching events and control actions
interact in a complex manner.Phenomena observed in power system oscillatory dynamics are diverse and
complex. Measured ambient data are known to exhibit noisy, nonstationary
fluctuations resulting primarily from small magnitude, random changes in load,
driven by low-scale motions or nonlinear trends originating from slow control
actions or changes in operating conditions. Forced oscillations resulting from
major cascading events, on the other hand, may contain motions with a broad
range of scales and can be highly nonlinear and time-varying.Prediction of temporal dynamics, with the ultimate application to real-time
system monitoring, protection and control, remains a major research challenge
due to the complexity of the driving dynamic and control processes operating
on various temporal scales that can become dynamically involved. An
understanding of system dynamics is critical for reliable inference of the
underlying mechanisms in the observed oscillations and is needed for the
development of effective wide-area measurement and control systems, and for
improved operational reliability.Complex power system response data can contain nonlinear and possibly
strong local trends, noise, and may exhibit sudden variations and other
nonlinear effects associated with large and abrupt changes in system topology
or operating conditions that make the extraction of salient features difficult.
Accounting for nonlinear and time-varying features can not only provide a
better description of the data but can also reveal crucial information on
system’s oscillatory behavior such as modal properties and moving patterns.
By tracking the evolving dynamics of the underlying oscillations, the onset of
system instability can be determined and the critical stages for analysis and
control can be identified.
v
Recent years have seen a flourishing of activity in various techniques for theanalysis of power system dynamic behavior. Foremost among linear analysistools, Prony’s method has been widely applied to estimate small-signal dynamicproperties frommeasured and simulated data. Applications of linear techniquesin the context of power oscillations include, for example, modal extraction fromringdowns, the analysis of dynamic tests, and the identification of transferfunctions. Ongoing research into the study of modal behavior in the presenceof high noise levels and possibly nonstationary situations has resulted invariations to these approaches that extend their practical use to the realm ofnear-real-time stability assessment and control, and has stimulated thedevelopment of enhanced monitoring systems. This in turn, has sparked aresurgence of interest in the development of new algorithms that use theavailable online information to estimate modal properties.
Advances in signal processing algorithms, along with continuously growingcomputational resources and monitoring systems are beginning to makefeasible the analysis and characterization of transient processes using real-time information. Much of the recent work has been driven by interest in nearreal-time estimation of electromechanical modal properties from measuredambient data. This effort has resulted in various signal processing methodswith the capability of tracking the evolving dynamics of critical system modes.
Complementary, time–frequency analysis techniques that explicitlyacknowledge and incorporate nonlinearity or nonstationarity in both the timeand frequency domain are emerging as subjects of research and application inengineering investigations. Adaptive, nonlinear time-varying methods with theability to capture the temporal evolution of critical modal parameters, promise toenhance our understanding of the physical mechanisms that underlie systemoscillatory dynamics and have the potential to be applied to more generaltransient oscillations, governed by multiscale, time-varying processes.
A significant element of this major thrust is the development of wide-areameasurement systems. Extracting the salient features of interest from a widelydispersed and usually large number of system observations is a complexproblem. In the analysis of large models, where a significant amount ofobservational data is available, the development of data-based statisticalmodels with the capacity to process the vast wealth of information andextract relevant, physically independent patterns is appealing. For many ofthe above developments, a complete framework for temporal characterizationof system behavior, however, is still evolving.
The combined utilization of temporal, modal information and advancedmeasurement and control techniques holds also enormous potential to providecritical information for early detection, mitigation, and avoidance of large-scalecascading failures and could form the basis of smart, wide-area automatedanalysis and control systems. Analysis and characterization of time-synchronized system measurements requires mathematical tools that areadaptable to the varying system conditions, accurate and fast, while reducingthe complexity of the data to make them comprehensible and useful for control
vi Preface
and real-time decisions. Experience with the analysis of complex inter-areaoscillations from measured data, shows that issues such as noise, time-varyingbehavior, data measurement errors, and nonlinear effects have to be addressedif these tools are to be of practical use. Further, the applicability of thesetechniques to both, ambient from online system measurements and large-scaletransient oscillations has to be fully investigated because some techniques arebetter suited for a specific type of behavior.
This book deals with the development and application of advancedmeasurement-based signal processing techniques to the study, characterization,and control of complex transient processes in power systems. Recent advances inunderstanding, modeling and controlling system oscillations are reviewed.Specific attention is given to the modeling and control of complex time-varying(and possibly nonlinear) power system transient processes which have not beenpresent in previous work. Techniques that explicitly address and treatnonlinearity and nonstationarity are given and efficient methods to generatetime-varying system approximations from both measured and simulated dataare discussed. Attention is also given to the vital new ideas of dynamic securityassessment in real-time implementations and the development of smart, wide-area measurement and control systems incorporating FACTS (flexible ACtransmission system) technology. Application examples include the analysis ofreal data collected on grids in western North America, Australia, Italy, andMexico. These studies are expected to stimulate the interest of otherresearchers, toward the investigation of complex nonstationary power systemoscillations andmay form the basis of more advanced computational algorithms.
The book is organized into eight chapters written by leading researchers whoare major contributors to knowledge in this field.
Chapter 1 demonstrates and examines the performance of several methods forestimating small-signal dynamic properties from measured responses. Thetheoretical basis for these methods is described as well as application,properties, and performance. Examples include computer simulations andactual system experiments from the western North American power grid.Analysis goals center on estimating the modal properties of the systemincluding modal frequency, damping, and shape.
Chapter 2 revisits some of the fundamental assumptions of the recentlyintroduced Hilbert–Huang transform. The ability of empirical modedecomposition (EMD) to yield monocomponent intrinsic mode functions isexamined in the context of power system oscillations. Some enhancements tothe EMD are proposed to enhance its ability to better discriminate betweenclosely spaced frequency components. Additionally, frequency demodulation issuggested, to extract physically relevant instantaneous frequency from theHilbert transform. Synthetic data as well as real life data are used todemonstrate the validity of the enhancements.
Chapter 3 discusses some refinements to the Hilbert–Huang technique toanalyze time-varying multicomponent oscillations. Improved masking signaltechniques for the EMD are proposed and tested on measured data of a real
Preface vii
event in northern Mexico. Based on this framework, a novel approach to thecomputation of instantaneous damping is suggested and a local implementationof the Hilbert transform is also described. The accuracy of the method isdemonstrated by comparisons to Prony and Fourier analysis.
Chapter 4 investigates the applicability of Hilbert–Huang analysis techniqueto extract modal information in the presence of noise and possibly nonstationarysituations. Application of Hilbert analysis is examined relative to the moreestablished Prony analysis, with particular reference to the considerablestructural differences which exist between the two methods. Factors affectingthe performance of the techniques including noise tolerance, performance in thecase of closely spaced frequency components and changes in the underlyingsystem dynamics are discussed and investigated using synthetic and measureddata.
In Chapter 5 a real-time centralized controller for addressing small-signalinstability related events in large electric power systems is proposed. Using wide-area monitoring schemes to identify the emergence of growing or undampedoscillations related to interarea and/or local modes, rules are developed forincreasing multi-Prony method’s observability and dependability. Thisinformation is then utilized to initiate static VAR compensation controls toenhance the damping of a critical mode; the algorithms are tested in a two-areapower system and in a large-scale simulation example.
Chapter 6 discuses the use of multivariate data analysis techniques to extractand identify dynamically independent spatiotemporal patterns from time-synchronized data. By seeing the snapshots of system data as a realizationof random fields generated by some kind of stochastic process, a statisticalapproach to investigate propagating phenomena of different spatial scales andtemporal frequencies is proposed and tested on real noisy measurements fromtheMexican system. Themethod provides accurate estimation of nonstationaryeffects, modal frequency, time-varying shapes, and time instants of intermittenttransient behavior.
Chapter 7 proposes new techniques for detection and estimation ofnonstationary power transients. Attention is focused on two aspects of smallsignal models: the detection of change in the system and the identification of thenew operating parameters. Techniques to detect significant changes in systemdynamics by analyzing the dynamic response to continual load changes basedon detection theory are proposed. Approaches based on time–frequencyanalysis techniques are then used to yield improved modal estimates innonstationary environments. Applications to measurement data from theAustralian connected system are presented.
Finally, Chapter 8 discusses the development of advanced monitoring andcontrol approaches for enhancing power system security. The monitoringstructure is based on wavelet analysis of wide-area measurements systemstargeted to extract the critical damping of critical oscillation modes. Ahierarchical response-based control strategy that may incorporate FACTStechnologies and special protection systems is developed and tested on a
viii Preface
dynamic model of the Italian interconnected system to provide effectivestabilization of critical modes.
The book is the first comprehensive, systematic account of current analysismethods in power system oscillatory dynamics in both time and frequencydomains ranging from modal analysis, to data-driven time-series models andstatistical approaches. The procedures can be used in various disciplines otherthan power engineering, including signal and time analysis, processidentification and control, and data compression and has wide applications tomany important problems covering engineering, biomedical, physical,geophysical, and climate data.
This is a book intended for advanced undergraduate and graduate courses,as well as for researchers, utility engineers, and advanced teaching in the fieldsof power engineering, signal processing, and identification and applied control.
Guadalajara, Mexico A.R. Messina
Preface ix
Acknowledgments
The editor is grateful to the contributing authors. He is also thankful to anumber of colleagues who provided the thrust for this work. Among them areJuan J. Sanchez-Gasca (GE, Schenectady, NY), Mike Gibbard (University ofAdelaide, Australia), and Brian Cory (Imperial College, UK). Their numerouscomments and observations throughout the years are a highly appreciatedcontribution.
It is also a pleasure to acknowledge the support of Ms Katelyn Stanne,Springer US, who proofread different versions of the manuscript and guidedthe editor during the editorial work.
xi
Contents
1 Signal Processing Methods for Estimating Small-Signal Dynamic
Properties from Measured Responses . . . . . . . . . . . . . . . . . . . . . . . . . 1Daniel Trudnowski and John Pierre
2 Enhancements to the Hilbert–Huang Transform for Application to
Power System Oscillations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37Nilanjan Senroy
3 Variants of Hilbert–Huang Transform with Applications to Power
Systems’ Oscillatory Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63Dina Shona Laila, Arturo Roman Messina, and Bikash Chandra Pal
4 Practical Application of Hilbert Transform Techniques in Identifying
Inter-area Oscillations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101T. J. Browne, V. Vittal, G.T. Heydt, and Arturo Roman Messina
5 A Real-Time Wide-Area Controller for Mitigating Small-Signal
Instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127Jaime Quintero and Vaithianathan (Mani) Venkatasubramanian
6 Complex Empirical Orthogonal Function Analysis of Power System
Oscillatory Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159P. Esquivel, E. Barocio, M.A. Andrade, and F. Lezama
7 Detection and Estimation of Nonstationary Power Transients . . . . . . 189Gerard Ledwich, Ed Palmer, and Arindam Ghosh
8 Advanced Monitoring and Control Approaches for Enhancing Power
System Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231Sergio Bruno, Michele De Benedictis, and Massimo La Scala
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
xiii
Contributors
M.A. Andrade Department of Electrical Engineering, Universidad Autonoma
de Nuevo Leon, Monterrey, Mexico, [email protected]
E. Barocio Department of Electrical Engineering, University of Guadalajara,
Guadalajara, Mexico, [email protected]
Michele De Benedictis Dipartimento di Elettrotecnica ed Elettronica (DEE),
Politecnico di Bari, Bari, Italy, [email protected]
T.J. Browne Ira A. Fulton School of Engineering, Department of Electrical
Engineering, Arizona State University, Tempe, AZ, USA, [email protected]
Sergio Bruno Dipartimento di Elettrotecnica ed Elettronica (DEE),
Politecnico di Bari, Bari, Italy, [email protected]
P. Esquivel Department of Electrical and Computer Engineering, The Center
for Research and Advanced Studies, Cinvestav, Mexico
Arindam Ghosh Faculty of Built Environment and Engineering, Queensland
University of Technology, Brisbane, Australia, [email protected]
G.T. Heydt Ira A. Fulton School of Engineering, Department of Electrical
Engineering, Arizona State University, Tempe, AZ, USA, [email protected]
Dina Shona Laila Department of Electrical and Electronic Engineering,
Imperial College, London, UK, [email protected]
Massimo La Scala Dipartimento di Elettrotecnica ed Elettronica (DEE),
Politecnico di Bari, Bari, Italy, [email protected]
Gerard Ledwich Faculty of Built Environment and Engineering, Queensland
University of Technology, Brisbane, Australia, [email protected]
F. Lezama Department of Electrical and Computer Engineering, The Center for
research and Advanced Studies, Cinvestav, Mexico, [email protected]
xv
Arturo Roman Messina Department of Electrical and Computer Engineering,The Center for Research and Advanced Studies, Cinvestav, Guadalajara,Mexico, [email protected]
Bikash Chandra Pal Department of Electrical and Electronic Engineering,Imperial College, London, UK, [email protected]
Ed Palmer Faculty of Built Environment and Engineering, QueenslandUniversity of Technology, Brisbane, Australia, [email protected]
John Pierre Electrical and Computer Engineering Department, Universityof Wyoming, Laramie, WY, USA, [email protected].
Jaime Quintero Faculty of Engineering, Universidad Autonoma de Occidente,Cali-Valle, Colombia, [email protected]
Nilanjan Senroy Department of Electrical Engineering, Indian Instituteof Technology, New Delhi, India, [email protected]
Daniel Trudnowski Electrical Engineering Department, Montana Tech of theUniversity of Montana, Butte, MT, USA, [email protected]
V. Vittal Ira A. Fulton School of Engineering, Department of ElectricalEngineering, Arizona State University, Tempe, AZ, USA, [email protected]
Vaithianathan (Mani) Venkatasubramanian School of Electrical Engineeringand Computer Science, Washington State University, Pullman, WA 99164USA, [email protected]
xvi Contributors
Chapter 1
Signal Processing Methods for Estimating
Small-Signal Dynamic Properties
from Measured Responses
Daniel Trudnowski and John Pierre
Abstract Power system small-signal electromechanical dynamic properties areoften described using linear system concepts. The underlying hypothesis is thatsmall motions of the system can be described by a set of ordinary differentialequations. Modal analysis of these governing equations provides considerableinsight into the stability properties of the system. Over the past two decades,many signal processing techniques have been developed to conduct modalanalysis using only time-synchronized actual system measurements. Some tech-niques are appropriate for transient signals, others are for ambient signalconditions, and some are for conditions where a known probing signal isexciting the system. In this chapter, an overview of many of the more successfulanalysis techniques is presented. The theoretical basis for these methods isdescribed as well as application properties and performance. Examples includecomputer simulations and actual system experiments from the western NorthAmerican power system. Analysis goals center on estimating the modal proper-ties of the system including modal frequency, damping, and shape.
1.1 Introduction
Time-synchronized measurements provide rich information for estimating apower system’s electromechanical modal properties via advanced signal proces-sing. This information is becoming critical for the improved operational relia-bility of interconnected grids. A given mode’s properties are described by itsfrequency, damping, and shape. Modal frequencies and damping are usefulindicators of power system stress, usually declining with increased load orreduced grid capacity. Mode shape provides critical information for opera-tional control actions. Over the past two decades, many signal processing
D. Trudnowski (*)Electrical Engineering Department, Montana Tech of the University of Montana,Butte, MT, USAe-mail: [email protected]
A.R. Messina, Inter-area Oscillations in Power Systems,Power Electronics and Power Systems, DOI 10.1007/978-0-387-89530-7_1,� Springer ScienceþBusiness Media, LLC 2009
1
techniques have been developed and tested to conduct modal analysis usingonly time-synchronized actual system measurements. Some techniques areappropriate for transient signals while others are for ambient signal conditions.
Many of the signal processing algorithms described in this chapter are thebasis for several evolving software tools. The majority of these tools are used toconduct engineering analysis of the grid in an off-line or post disturbancesetting [1]. More recently, online real-time software tools and applications areevolving [2] and will likely continue to be a research focus area for the powersystem community.
Near-real-time operational knowledge of a power system’s modal propertiesmay provide critical information for control decisions and thus enable reliablegrid operation at higher loading levels. For example, modal shapemay somedaybe used to optimally determine generator and/or load-tripping schemes toimprove the damping of a dangerously low damped mode. The optimizationinvolves minimizing load shedding and maximizing improved damping. Thetwo enabling technologies for such real-time applications are a reliable real-time-synchronized measurement system and accurate modal analysis signalprocessing algorithms.
In this chapter, an overview of many of the more successful analysis techni-ques is presented. The theoretical basis for these methods is described as well asapplication and performance properties. Examples include computer simula-tions and actual system experiments from the western North American powersystem (wNAPS). Analysis goals center on estimating the modal properties ofthe system including modal frequency, damping, and shape.
The chapter is organized as follows. Section 1.2 discusses system basics. Anoverview of mode estimation algorithms is provided in Section 1.3. Section 1.4discusses the use of probing signals to improve mode estimates. Section 1.5provides some examples. Model validation and estimation assessment is dis-cussed in Section 1.6, and Section 1.7 covers mode-shape estimation. Finally,conclusions are discussed in Section 1.8.
1.2 System Basics
Analyzing and estimating power system electromechanical dynamic effects area challenging problem because the system:
1. is nonlinear, high order, and time varying;2. contains many electromechanical modes of oscillation close in frequency;
and3. is primarily stochastic in nature.
Design of signal processing algorithms requires that one address each of theseissues. Fortunately, the system behaves relatively linear when at a steady-stateoperating point [3].
2 D. Trudnowski and J. Pierre
As has been established in one of the many excellent books that address theproperties and nature of electromechanical dynamics in power systems (e.g., see[4, 5]), electromechanical modes are typically classified as either local or inter-area in nature. Local modes occur when a single generator or plant swingsagainst the system while an inter-area mode occurs when several generators inan area swing against generators in another area. Because local modes arecharacterized by larger inertias and lower impedance paths, their frequenciestend to be higher. In general, local modes tend to be in the 1–2 Hz range whileinter-area modes tend to be in the 0.2–1.0 Hz range. Typically, the inter-areamodes are more troublesome.
Consistent with power system dynamic theory, we assume that a powersystem can be linearized about an operating point [4, 5]. The underlyingassumption is that small motions of the power system can be described by aset of ordinary differential equations of the form
_x tð Þ ¼ Ax tð Þ þ BLq tð Þ þ BEuE tð Þ
y tð Þ ¼ Cx tð Þ þDLq tð Þ þDEuE tð Þ þ m tð Þ(1:1)
where q is a hypothetical random vector perturbing the system, vector x con-tains all system states including generator angles and speeds, and t is time.Control actions that can be described as smooth functions of the state x areembedded in the system A matrix, and all other actions are represented by theexogenous input vector uE. These include set-point changes, low-level probingsignals (e.g., a low-level probing signal into a DC converter), and load pulsesthat are applied to examine system dynamics. Measurable signals are repre-sented by y which contains measurement noise m that includes effects frominstruments, communication channels, recording systems, and similar devices.In general, measurement noise has a relatively small amplitude when qualityinstrumentation is employed. Changes which are breaker actuatedmay producesystem topology changes that alter the system A matrix to various degrees.
The assumption for q is that it is a vector of small-amplitude randomperturbations typically conceptualized as noise-produced load switching. Ithas been hypothesized that the load switching is primarily integrated stationaryGaussian white noise with each element of q independent [6]. This assumption iscertainly open to more research.
An expanded perspective of the system is shown in Fig. 1.1 where yi is the ithelement of y [7]. Multiple-input and multiple-output (MIMO) system G isassumed linear. Network topology changes are represented by switches indynamic gain matrices K and K0, which may or may not be deliberate.
We classify the response of the system in Fig. 1.1 as one of two types:transient (sometimes termed a ringdown) and ambient. The basic assumptionfor the ambient case is that the system is excited by low-amplitude variations atq and uE and that the variations are typically random or pseudorandom innature. This results in a response at y that is colored by the dynamics G.
1 Signal Processing Methods for Small-Signal Dynamic Properties 3
A transient response is typically larger in amplitude and is caused by a sudden
switch at s or s0, or a sudden step or pulse input at uE. The resulting time-domain
response is a multimodal oscillation superimposed with the underlying ambient
response.The different types of responses are shown in Fig. 1.2, which shows a widely
published plot of the real power flowing on a major transmission line during a
breakup of the wNAPS in 1996. Prior to the transient at the 400 s point, the
system is in an ambient condition. After the ringdown at the the 400 s point, the
system returns to an ambient condition. The next event in the system causes an
unstable oscillation.
Fig. 1.2 Real power flowing on a major transmission line during the western North Americanpower system breakup of 1996
G
K '
K
unknown dynamics
known dynamics
s' = unknown topology changes
s = known topology changes
µ
yi(t)
i(t)
+
+q(t)
uE(t))(ˆ tyi
Fig. 1.1 A structure for information sources in process identification
4 D. Trudnowski and J. Pierre
In developing and applying measurement-based modal analysis algorithms,
it is imperative that one considers the stochastic nature of the system. Power
systems are continually excited by random inputs with high-order indepen-
dence. This is modeled by q(t) in our formulation. Because of this stochastic
nature, no algorithm can exactly estimate the modal properties of the system
from finite-time measurements. There will always be an error associated with
the estimate. When evaluating estimation algorithms, one must address these
error properties. This includes the bias error as well as the variance of the
estimate.In terms of application, we classifymodal frequency and damping estimation
algorithms into two categories: (1) ringdown analyzers and (2) mode meters. A
ringdown analysis tool operates specifically on the ringdown portion of the
response; typically the first several cycles of the oscillation (5–20 s). Alterna-
tively, a mode meter is applied to any portion of the response: ambient,
transient, or combined ambient/transient. Ultimately, a mode meter is an
automated tool that estimates modal properties continuously and without
reference to any exogenous system input.
1.3 Signal Processing Methods for Estimating Modes
Many parametric methods have been applied to estimate power system electro-
mechanical modes. As stated above, we classify these methods into two cate-
gories: ringdown analyzers and mode meters. In this section, we provide an
overview of some of the algorithms that have been used to solve these problems.
1.3.1 Ringdown Algorithms
Ringdown analysis for power system modal analysis is a relatively mature
science. The underlying assumed signal model for these algorithms is a sum of
damped sinusoids. The most widely studied ringdown analysis algorithm is
termed Prony analysis. The pioneering paper by Hauer, Demeure, and Scharf
[8] was the first to establish Prony analysis [9] as a tool for power system
ringdown analysis. Expansion to transfer function applications, multiple out-
puts, and improved numerics were progressively established in [10–17]. Other
ringdown analysis algorithms have been successfully applied to power system
applications. These include the minimal realization algorithm first introduced
in [18], the eigenvalue ralization algorithm (ERA) in [19], the matrix pencil
method [20], and the Hankel total least squares (HTLS) [20]. The conclusions
and discussions in [21] point to the vast similarities between Prony analysis and
the ERA. A comparative analysis between matrix pencil, HTLS, and Prony
analysis in [20] conclude that HTLS and matrix pencil estimate the mode
1 Signal Processing Methods for Small-Signal Dynamic Properties 5
damping more accurately. These conclusions are certainly subject to the exam-ple case and the parameters chosen for the analysis.
It is beyond the scope of this chapter to provide the equations for all theringdown methods. As an overview, we provide the basic equations for Pronyanalysis. The reader is directed to the above references for more details.
While ignoring noise content and assuming nonrepeated poles, if one appliesan impulse input to the system in (1.1), the response at the ith output can bewritten as
yj tð Þ ¼Xn
i¼1Bie
lit (1:2)
where li is the ith pole (mode). If we let t = kT, where T is the constant sampleperiod, this equation can be converted to discrete-time form as
yj kTð Þ ¼Xn
i¼1Biz
ki ; for k ¼ 0; 1; . . . ;m (1:3)
where zi ¼ eliT is the discrete-time pole. Equation (1.3) is expanded into matrixform as
y 0ð Þy Tð Þ
..
.
y mTð Þ
266664
377775¼
1 1 � � � 1
z1 z2 � � � zn
..
. ... . .
. ...
zm1 zm2 � � � zmn
266664
377775
B1
B2
..
.
Bn
266664
377775
(1:4)
It is relatively easy to show [9] that
y nTð Þy ðnþ 1ÞTð Þ
..
.
y mTð Þ
2
66664
3
77775¼
y ðn� 1ÞTð Þ y ðn� 2ÞTð Þ � � � y 0ð Þy nTð Þ y ðn� 1ÞTð Þ � � � y Tð Þ
..
. ... . .
. ...
y ðm� 1ÞTð Þ y ðm� 1ÞTð Þ � � � y ðm� nÞTð Þ
2
66664
3
77775
a1
a2
..
.
an
2
66664
3
77775(1:5)
where the ai’s are the coefficients of the characteristic equation
zn � a1zn�1 þ a2z
n�2 þ � � � þ an� �
¼ 0 (1:6)
The solution of (1.6) are the zi poles.Prony analysis involves solving (1.5) for the ai characteristic equation coeffi-
cients. Then (1.6) is rooted to obtain the zi discrete-time poles. Lastly, (1.4) issolved for the Bi’s. As described in the above references, these equations can beextended to the multioutput case. Selection of model order n, sample period T,and number of data points mþ1 are also addressed throughout the literature.
6 D. Trudnowski and J. Pierre
1.3.2 Mode-Meter Algorithms
Ambient analysis of power system data estimates the modes when the primary
excitation to the system is random load variations, which results in a low-
amplitude stochastic time series (ambient noise). A good place to begin ambient
analysis is with nonparametric spectral estimation methods, which are very
robust as they make very few assumptions. The most widely used nonpara-
metric method is the Welch periodogram [22, 23] spectrum which provides
an estimate of a signal’s strength as a function of frequency. Thus, usually
the dominate modes are clearly visible as peaks in the spectral estimate. The
estimates of the mode frequencies are identifiable in the locations of the peaks.
The narrower the peaks, the lighter is the damping.Welch spectral estimates are
also used in estimating mode shape as will be discussed in Section 1.7. While
robust and insightful, nonparametric methods do not provide direct numerical
estimates of a mode’s damping ratio and frequency. Therefore, to obtain
further information parametric methods are applied.Ambient-based mode estimation can be conducted in the time domain or
frequency domain. Time-domain algorithms operate directly on the sampled
data while frequency-domain methods require the estimation of the power
spectral density (PSD) function (usually using Welch’s method). The first
available ambient-based mode estimation work [6] used a frequency-domain
strategy. The method described in [6] was applied to actual system measure-
ments. With this approach, Welch periodogram averaging is used to estimate
the PSD of a signal. Frequency-domain identification is then used to estimate
the system modes. A disadvantage of the approach in [6] is that the frequency-
domain identification process used requires an initial estimate of the system
modes prior to analysis which is difficult to automate.There are two basic types of parametric mode estimation algorithms: block
processing and recursive. With block processing algorithms, the modes are
estimated from a window of data. For each new window of data, a new estimate
is calculated. For example, assume one is using a 5min window length. For each
window of data, a single set of modes is calculated. All data in the 5 min block
are equally weighted. A new mode estimate can be calculated as often as
required, but each calculation requires 5 min of the most recent data. The
first application of block processing is contained in [24] where the Yule–Walker
(YW) algorithm is used to estimate modes using an autoregressive (AR) model.
The method is extended to the overdetermined modified YW method [35] to
estimate an autoregressive moving average (ARMA) model in [25]. The
approach is further extended to multiple signals in [26], which can improve
the performance. Block processing methods using subspace methods CVA
(canonical variate algorithm) and N4SID (numerical algorithm for subspace
state–space system identification) were first introduced in [27] and [34], respec-
tively. A variation of the YW approach that estimates the autocorrelation
function using a frequency-domain calculation is introduced in [28]; this
1 Signal Processing Methods for Small-Signal Dynamic Properties 7
method is termed the Yule–Walker spectrum (YWS) method. Also in [28], theYW, YWS, and N4SID algorithms are compared. Another frequency-domainmethod is the frequency-domain decomposition (FDD) method described in[29], which decomposes the signals’ estimated power spectrum.
For recursive methods, the estimated modes are updated for each newsample of the data. The new estimate is obtained using a combination of thenew data point and the previous mode estimate. A forgetting factor is used todiscount information based on previous data; therefore, new data is weightedmore in each calculation. Similar to the block processing methods, all recursivemethods tested to date require many minutes of data to converge to a steady-state solution. Published results include the least-mean squares (LMS) method[30] and the regularized robust recursive least-squares (R3LS) method [31, 32].
The R3LS method described in [32] offers several advances to previousalgorithms. First, it accommodates an autoregressive moving average exogen-ous (ARMAX) model to account for ambient noise as well as a known input,which can enhance performance during probing. Second, it has a robust objec-tive function to reduce the impact of missing or outlier data, and third, it canincorporate a priori knowledge of the modes. The full impact of these advancesis the subject of current and future research.
An important component of a mode meter is the automated application ofthe algorithm. With all algorithms, several modes are estimated and many ofthem are ‘‘numerical artifacts.’’ Typically, ‘‘modal energy’’ methods are used todetermine which of the modes in the frequency range of the inter-area modeshave the largest energy in the signal [28]. It is then assumed that this is the modeof most interest.
It is beyond the scope of this chapter to provide the equations for all themode-meter methods described above. The reader is directed to the abovereferences for more details and for information on preprocessing the databefore application of the mode-meter algorithms.
1.4 Power System Identification Using Known Probing Signals
It is absolutely imperative to understand that because of the stochastic nature ofthe system, the accuracy of any mode estimation is limited. It is possible tosignificantly improve the estimation by exciting the system with a probingsignal. A signal may be injected into the power system using a number ofdifferent actuators such as resistive brakes, generator excitation, or modulationof DC intertie signals. For example, operators of the wNAPS use both the1,400 MW Chief Joseph dynamic brake and modulation of the Pacific DCintertie (PDCI) to inject known probing signals into the system. The wNAPSis shown in Fig. 1.3 with the PDCI being the DC line flowing from Oregon tosouthern California. The PDCI has been modulated with a number of differentsignals including short duration mid-level probing resulting in transient
8 D. Trudnowski and J. Pierre
PHOENIXMOJAVE
CRANBROOK
SAN FRANCISCO
CORNERSNAVAJO
PINTO
SHASTA
BUCKLEY
SUNDANCE
FT. PECK
KEMANO
PEACE CANYON
MICA
VANCOUVER
SEATTLE
PRINCE RUPERT
AREA
AREA
COLSTRIP
BOISE
PORTLANDAREA
MALIN
TABLE MTN
ROUND MTN
SALT LAKECITY AREA
MEXICO
EL PASOAREA
PALO
LUGO
MIDPOINT
AREA
LOS ANGELESAREA
ALBUQUERQUEAREA
VERDE
DENVERAREA
HOOVER
AREA
LANGDON
HOT SPRINGS
HELLSCANYON
CHIEF JOSEPH
GRAND
BURNS
FOUR
HVDC TERMINAL
COULEE
DEVERS
Fig. 1.3 Major buses and lines in the western North American power system
1 Signal Processing Methods for Small-Signal Dynamic Properties 9
responses and long-duration low-level probing that result in measured signalsonly slightly above the system ambient noise floor. Low-level probing should becarried out at a level low enough to not be a significant disturbance.
The wNAPS has a long history in the use of probing signals for electromecha-nical mode identification [7, 33]. During the 1980s and 1990s the Chief Josephbrake was frequently used to benchmark system characteristics. In the late 1990s,with synchronized wide-area measurements becoming readily available, modu-lating the PDCI became more common. In 1999, mid-level probing signals wereused to characterize the mode damping. In 2000, low-level pseudorandom noisewas injected into the system. The application of system identification methods tothe input and output data from that test showed great promise for mode estima-tion [34]. In 2005, 2006, and 2008, a number of extensive tests were carried outusing low-level multisine probing signals modulate at the PDCI. The synchro-nized measurements of the system response to those tests proved to be rich ininformation about the system’s dynamic characteristics.
With known input signals, not only can the electromechanical modes beidentified with improved performance, but complete input/output system mod-els, such as transfer functions and state–space models, can be estimated from theinput location to the measured output locations. Many different system identifi-cation methods can be used. This includes extending the R3LS [32] and N4SID[34] methods described previously. There is a tremendous amount of literature onsystem identification given measured inputs and outputs. Some of these algo-rithms work on the time-domain data while other algorithms utilize the fre-quency-domain data. The literature is too extensive to review here; the reader isreferred to one of many textbooks (e.g., see [36]). Classical nonparametric meth-ods such as ETFE (empirical transfer function estimation) and spectral methods[36] may be used to estimate the system magnitude and phase response. Theadvantage of the nonparametric methods is that theymake very few assumptionsabout the underlying systemmodel. Thus, they play an important role in validat-ing parametric system models where one looks for consistency from the fre-quency response identified from a parametric method and the nonparametricmethods. The parametric methods provide much more information about thesystem such as a state–space model or a transfer function equation. It is impor-tant that the parametric algorithm chosen matches well with the underlyingcondition. For example, if an algorithm designed to analyze a transient response(i.e., a sum of damped sinusoids) is applied to ambient data, which is not the sumof damped sinusoids, then poor results are expected.
1.4.1 Probing Signal Selection
In choosing a low-level probing signal to inject into the system, many factorscome into play. The objective in probing signal design is to create an input thatwill result in accurate estimates of the electromechanical inter-area modes andpossibly other system dynamic characteristics while maintaining safe operation
10 D. Trudnowski and J. Pierre
of the power system. The choice of probing signals has a very substantialinfluence on the observed measured data. The protection of the power systemis of the utmost importance. Other important considerations include the shape,amplitude, duration, and repetition of the injected signal. Some identificationtechniques were developed for specific input signals.
A few limitations on the input design are specific to the power systemapplication. It is desirable not to have too many sharp transitions in themodulated signal on a DC intertie. Thus, this rules out many common systemidentification probing signals, which typically transition from rail to rail. Also,the signal should begin and end near a value of zero creating smooth transitionswhen injected into the system. Second, it is desirable to keep the peak probingamplitude small when probing for a long duration. For example, with thePDCI, the maximum input magnitude has been limited to �20 MW for long-duration probing. Another constraint is that the probing input should not looklike a single sinusoidal component as it could be mistaken for a sustainedoscillation. A pseudorandom signal is preferred.
System identification theory gives much guidance for input design. It isimportant to keep in mind that when probing, the measured outputs are acombination of the response to the probing signal and the ambient signal,which is always present in the measured outputs. The ambient signal is stochas-tic (random) in nature. Thus, when the probing signal is present, only a portionof the measured output is the system response to the probing signal, and theother portion is the ambient noise process. When it comes to the quality of theestimated parameters, it is the spectrum of the probing signal which is mostimportant, not the particular time-domain wave shape. The general idea is toplace the content of the probing signal in the frequency band of interest, in thiscase the frequency range of the inter-area electromechanical modes.
The amplitude and time duration of the low-level injected signal are critical.Clearly, the amplitude needs to be small enough not to interfere with the normaloperation of the power system. Yet, there is a well-known trade-off in systemidentification between the observation time and the signal strength. Perfor-mance of system identification algorithms improves with signal-to-noise ratio(SNR) and with observation time. The repeatability of the pseudonoise isimportant to fully take advantage of the repetition of the injected signal.Also, knowing the specific frequency content is critical.
An important quality of a probing signal is its crest factor. The crest factor ofa zero mean waveform u[n] is defined as
Cr¼�
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffimaxn u2½n�
1=Nð ÞPN
n¼1 u2½n�
s(1:7)
whereN is the number of samples in the waveform and n is the nth time sample.The crest factor is the ratio of the maximummagnitude of the signal to the rootmean square (RMS) value. It is desirable to have a probing signal with as large
1 Signal Processing Methods for Small-Signal Dynamic Properties 11
an RMS value as possible for a given maximum peak magnitude. Thus, a good
waveform design should have a small crest factor while maintaining the desired
spectrum and power carried by the waveform. The minimum crest factor is
unity and this only occurs in signals which transition from rail to rail. Because
these types of signal are undesirable in this application, the minimum crest
factor cannot be achieved.There are some important advantages to using a periodic input signal.
Output waveforms can be averaged over the periods giving an effective
increase in the SNR by the number of periods averaged. This increase is
known as the processing gain. A similar gain in SNR can be seen in the
frequency domain at the frequency bins of the harmonics of the periodic
input. It is very important to inject an exact integer number of cycles so that
there is no leakage effect in the frequency domain. Also, periodic inputs allow
for methods to estimate the noise signal. The signal period, T, is important as
it determines the frequency resolution as �f=1/T when conducting fre-
quency-domain analysis. Because the inter-area electromechanical modes
are usually in the frequency range from approximately 0.1 to 1.0 Hz, a
frequency resolution in the neighborhood of 0.01 Hz should be adequate.
Note, there is a trade-off between the frequency resolution and the number of
averages. For a given input signal duration, the larger the period, the better
the frequency resolution, but fewer periods of the signal are available for
averaging, so the processing gain is less.For the system tests carried out in the wNAPS in 2005, 2006, and 2008,
a multisine input signal was used because of its favorable characteristics
relative to the above discussion. The bottom line is that the injected signal
should be chosen to not disrupt the normal operation of the power system,
yet to provide an accurate system model given a specific identification
method.
1.5 Mode Estimation Examples
Many papers have been published demonstrating the power system applica-
tion of signal processing methods for estimating modal frequencies and
damping. These papers include ringdown analysis and mode-meter applica-
tions. Many of these papers are referenced in the previous sections. In this
section, we provide a few examples to emphasize some of the significant
challenges. We refer the reader to the references for a complete view of the
application issues.Two types of examples are considered. With the first, a simulated system
with known properties is employed. The advantage of this system is that the
exact solution is known; therefore, algorithm properties can be evaluated. The
second type of examples uses actual system cases from the wNAPS.
12 D. Trudnowski and J. Pierre
1.5.1 Simulation System
The simulation test system is shown in Fig. 1.4. A modified version of the
system was originally developed as a simplified model of the western North
American power grid in [37]; detailed information is presented in the appendix
of [38]. It has been used in many publications as a research demonstration
model for stability-limited issues and mode estimation analysis.The system consists of major generation buses 17 through 24 and 45, and load
buses 31 through 41. Each generator is represented using a detailed two-axis
transient model equipped with a fast-acting voltage regulator, a power system
stabilizer (PSS) unit, and a turbine governor. Two identical generators are
attached to buses 17 through 24. Overall, the system order is 203. Each load is
split into a portion consisting of constant impedance, constant current, constant
power, and random. The random portion of both the real and reactive loads is
obtained by passing independent Gaussian white noise through a 1/f filter. It has
been hypothesized that such a filter is appropriate for load modeling [7].
23
24
1920
21
25
17
18
22
26
500230
230500
28
29
500230
2775, 35
70, 30
82, 40
44, 2644, 26
30, 12
14, 4
54, 20
Real, reactive load (pu)
62, 28
55, 25
3130, 1230
34
33
3235
37
36
38
39
4041
500/110
45
Rectifier
Inverter
42
500 kVHVDC
500/110
4330
Fig. 1.4 Simulation test system
1 Signal Processing Methods for Small-Signal Dynamic Properties 13
Two primary system operating conditions are used with the simulations thatfollow. With the first condition, termed the 17-machine system, all generators areconnected to the system. Under this condition, the most dominant inter-areamodes are shown in Table 1.1. With the second condition, generator bus 45 isdisconnected from the system; this condition is termed the 16-machine system.Under this condition, the dominant inter-areamodes are shown in Table 1.2. Themodes shown in Tables 1.1 and 1.2 were calculated by conducting an eigenana-lysis of the entire system’s small-signal model under nominal steady-state operat-ing conditions. The eigenanalysis was conducted using the methodology in [4].
For the examples that follow, a typical time-domain simulation consists ofdriving the systemwith independentGaussian load variations tomimic ambientconditions. The system’s response consists of small random variations in thesystem states. As an example, the top plot of Fig. 1.5 shows the resultingrandom variations of bus 22 frequency for a 10 min simulation. The frequencyis calculated using the derivative of the bus phase angle.
To mimic a transient condition, a 0.5 s long load pulse is applied to bus 35.The bottom plot of Fig. 1.5 shows the system’s response to a 700 MW loadpulse. Figure 1.6 shows the response to a 1,400 MW pulse.
1.5.2 Ringdown Analysis Performance
As described in Section 1.3, ringdown analysis is used to estimate the modalproperties from a transient. One important property we wish to emphasize isthat the accuracy of the estimate is strongly related to the SNR. That is, howlarge the ringdown is compared to the ambient noise.
As an example, consider the ringdowns in Figs. 1.5 and 1.6. For the ring-down portion, the SNR in Fig. 1.6 is four times as large as that in Fig. 1.5.Table 1.3 compares the Prony analysis results for these two responses. For eachcase, the Prony analysis was conducted from 31 to 50 s into the simulation. Asseen in Table 1.3, the higher SNR signal provides a more accurate mode
Table 1.1 Inter-area modes of 17-machine system
Frequency (Hz) Damping (%) Buses vs. Buses
0.318 10.74 North half vs. Southern half
0.422 3.63 North half vs. Southern half + bus 45
0.635 3.94 18 vs. Rest of system
0.673 7.63 20,21 vs. 24
Table 1.2 Inter-area modes of 16-machine system
Frequency (Hz) Damping (%) Buses vs. Buses
0.361 6.59 North half vs. Southern half
0.618 3.57 18 vs. Rest of system
0.673 7.66 20,21 vs. 24
14 D. Trudnowski and J. Pierre
estimate of the dominant mode. To exactly quantify the accuracy, a MonteCarlo simulation must be conducted.
1.5.3 Mode-Meter Performance
In this example, the performance of mode-meter algorithms is demonstrated.Specifically, we consider estimation accuracy in the context of the mode damp-ing, the analysis window size, and transient versus ambient conditions. A moreextensive overview of this comparison is contained in [32, 28]. Time-domain
0 2 4 6 8 10
59.98
60
60.02
Time (min.)
Hz
0 0.5 1 1.5 2
59.98
60
60.02
Hz
Ringdown
Zoom
Fig. 1.6 Bus 22 frequency for 16-machine system under a transient simulation response to a1,400 MW 0.5 s load pulse at bus 35. Pulse is applied 30 s into simulation
Ambient
Ringdown
0 2 4 6 8 10
59.98
60
60.02
Hz
0 2 4 6 8 10
59.98
60
60.02
Hz
Ringdown
Ambient
Time (min.)
Fig. 1.5 Bus 22 frequency for 16-machine system. Top plot, ambient condition. Bottom plot,transient simulation response to a 700MW 0.5 s load pulse at bus 35. Pulse is applied 30 s intosimulation
1 Signal Processing Methods for Small-Signal Dynamic Properties 15
simulation of the Fig. 1.4 system is employed to investigate properties. Because
of the random nature of the data, Monte Carlo simulations are employed to
fully evaluate the properties. In each simulation case, the measurement noise
terms (�i(t) in Fig. 1.1) are represented by passing white noise through a low-
pass first-order filter with a corner at 5 Hz. In each case, �i(t) is scaled such that
the SNR between yiðtÞ and �i(t) is 12 dB. This represents a relatively high
measurement noise condition.Figure 1.7 shows the estimates of the 0.361 Hz mode from the 16-machine
system under varying damping conditions and ambient operation using the YW
Table 1.3 Prony analysis estimates for 16-machine modes
Actual
Estimated Estimated
(700 MW pulse) (1400 MW pulse)
Frequency(Hz)
Damping(%)
Frequency(Hz)
Damping(%)
Frequency(Hz)
Damping(%)
0.361 6.59 0.362 7.2 0.361 6.8
0.618 3.57 0.619 3.9 0.618 3.6
–0.4 –0.2 00.2
0.3
0.4
0.5
Real (1/sec)
Imag
. (H
z)
YW Estimated
Actual
–0.4 –0.2 00.2
0.3
0.4
0.5
Real (1/sec)
Imag
. (H
z)
–0.4 –0.2 00.2
0.3
0.4
0.5
Real (1/sec)
Imag
. (H
z)
1%D 1%D
1%D
20%D20%D
20%D
Actual damping:upper left plot = 6.59% upper right plot = 3% lower plot = 1%
Fig. 1.7 Mode estimates for 16-machine system for estimating the 0.361 Hz mode from 100ambient Monte Carlo simulations using the YW algorithm. 1%D and 20%D are constantdamping lines
16 D. Trudnowski and J. Pierre
algorithm. The figure shows three s-plane plots representing varying damping
conditions. Mode conditions in the system can be varied by modifying system
operating conditions (e.g., steady-state loading) resulting in new small-signal prop-
erties. For each plot, 100 independent simulationswere conductedusing an analysis
window length of 10 min. The relative voltage angles between buses 25–21, 20–22,
19–23, and 26–17 are used as the inputs to the algorithm. For each case, the
estimated mode with the largest pseudoenergy in the region of the s-plane bound
by 0.2, 0.5 Hz, and 20% damping is plotted for each simulation. This tests the
algorithms capability to automatically estimate a mode without human software
interaction. As the mode becomes more lightly damped, it is estimated with con-
siderably more accuracy. This is a property of all mode estimation algorithms.A more challenging situation is that of estimating two closely spaced modes
as with the 17-machine system. Figure 1.8 shows plots for the YW, YWS, and
N4SID algorithms using an analysis window size of 10 min. The two modes
with the largest pseudoenergy terms in the region of the s-plane bound by 0.2,
0.5 Hz, and 20% damping are estimated with a mode-meter algorithm for each
simulation. The YW and YWS provide more accurate estimates than the
N4SID algorithm. Further investigation in [39] shows that the poor perfor-
mance of the N4SID is related to the measurement noise. That is, the
–0.4 –0.2 00.2
0.3
0.4
0.5
Real (1/sec)
Imag
./(2p
i) (H
z)
Estimated
Actual
–0.4 –0.2 00.2
0.3
0.4
0.5
Real (1/sec)
Imag
./(2p
i) (H
z)
–0.4 –0.2 00.2
0.3
0.4
0.5
Real (1/sec)
Imag
./(2p
i) (H
z)
N4SID
YWS
YW
1%D20%D
1%D1%D20%D20%D
Fig. 1.8 Mode estimates for 17-machine system for estimating the 0.318 and 0.422 Hz modesfrom 100 ambient Monte Carlo simulations
1 Signal Processing Methods for Small-Signal Dynamic Properties 17
performance of the N4SID degrades significantly as measurement noiseincreases while the YW and YWS algorithms continue to perform well withmeasurement noise. Note that the frequency estimates are much more accuratethan the damping estimates, and the accuracy of estimating the less dampedmode is better than the more highly damped mode.
The analysis window size is the amount of historical data required to obtainan estimate. Certainly, one would expect better performance from the mode-meter algorithms as total data used for analysis increases. The length of analysisdata is indicated byN samples in the algorithms or byTtotal¼TN seconds whereT is the sample period.
As before, a series of Monte Carlo simulations are conducted to demonstratethe effect of different window sizes on the performance of the algorithms. In eachof the simulations, the system is operated under the 17-machine condition. Thegoal is to estimate both the 0.318 Hz mode and the 0.422 Hz mode. For eachsimulation, the two modes with the largest pseudoenergy terms in the region ofthe s-plane bound by 0.2, 0.5 Hz, and 20% damping are estimated with a mode-meter algorithm. The results for the YW algorithm are summarized in Fig. 1.9.One can certainly see the improved performance as the window size is increased.
The basic assumption for the mode-meter algorithms is that the system isexcited by random inputs. In real-world automated conditions, data may oftencontain transient ringdowns and nonstationary conditions. A ringdown occurswhen the system is excited by a sudden input such as a fault, generator trip, or loadtrip. Often, the post-transient steady-state condition of the system will change.That is, the mode damping and frequencies will change following the transient. Ingeneral, the mode estimates actually improve if a ringdown is present in the data.
A condition often encountered is a transient which causes the mode tobecome less damped. As an example, a 700 MW, 0.5 s pulse is added to bus35 of the 16-machine system. Prior to the transient, the mode is at 0.361 Hz, 6%damping. During and after the transient, the mode shifts to 0.25 Hz, 1%
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0 5 10 15 20 25
Window size (min.)
Fre
q. (
Hz)
0
2
4
6
8
10
12
14
16
18
20
0 5 10 15 20 25Window size (min.)
Dam
ping
(%
)
Fig. 1.9 Accuracy of the YW algorithm for 0.318 and 0.422 Hz modes for varying Ttotal.¼mean � standard deviation, ¼median. Ambient condition
18 D. Trudnowski and J. Pierre
damping. Figure 1.10 shows the system’s response. The resulting mode esti-mates using the 10 min data window in the top plot of Fig. 1.10 are shown inFig. 1.11 using three different algorithms. The corresponding estimates usingthe 2 min data window in the bottom plot of Fig. 1.10 are shown in Fig. 1.12.The modes estimates converge to the more lightly damped post-transient con-dition. This is certainly the desired result.
1.5.4 Field Measured Data
As described earlier, operators of the wNAPS periodically conduct extensivedynamic tests. These tests typically involve 0.5 s insertion of the Chief Joseph1,400 MW braking resistor in Washington; and probing of the power referenceof the PDCI (see Fig. 1.3). The resulting system response provides rich data fortesting mode estimation algorithms. This section presents a few of these results.
Figure 1.13 shows the system response from a brake insertion along withseveral minutes of ambient data. The signal shown is the detrended real powerflowing on a major transmission line.
Two recursive mode estimation algorithms are applied to the data: the RLSand RRLS [32] algorithms. The resulting mode estimates are shown inFigs. 1.14 and 1.15. The damping estimates for the 0.39 Hz mode are shownas this is the most lightly damped dominant mode. The results are compared toa Prony analysis of the ringdown. More detailed results are shown in [32]. TheRRLS algorithm provides a more accurate mode-damping estimate and theaccuracy improves after the ringdown.
0 2 4 6 8 10–0.04
–0.02
0
0.02
0.04
Rad
ians
Time (min.)
8 8.5 9 9.5 10–0.04
–0.02
0
0.02
0.04
Time (min.)
Rad
ians
zoom
Fig. 1.10 Angle between buses 25 and 21 voltages for the 16-machine system. At 540 spoint in simulation, a 700 MW, 0.5 s pulse is added to bus 35. Pre- and post-transientmodes differ
1 Signal Processing Methods for Small-Signal Dynamic Properties 19
1.5.5 Probing Test Results
As discussed in Section 1.4, exciting the system with a probing signal improves
the mode estimates. This is demonstrated in this section using results from
wNAPS probing test results from August 2006.Figure 1.16 summarizes the responses from a typical wNAPS probing test.
The pseudorandom probing shown in the blue region is a multisine signal with
the phases adjusted to optimize the signal crest factor. The frequency content
is 0.1–0.9 Hz; after 1.0 Hz, the signal rolls off at 120 dB/decade and drops to
zero at 2 Hz. Note the probing signal is scaled to�20MWon the DC line. The
resulting response on the AC system is just above the ambient noise.Figure 1.17 shows the mode estimates comparing probing versus ambient
for a mode known to be near 0.38 Hz. In both the ambient case and the
probing case, 20 min of data are analyzed. The ambient and probing occur
at the same hour of the test and the system remained in the same relative
operating point; therefore, the actual systemmode likely remain constant. The
black circles are the mode estimates from a 680 s window sliding over every
60 s in the ambient data. The crosses show the same for the probing data.
Estimated
Pre-transient
Post-transient
–0.5 –0.4 –0.3 –0.2 –0.10.2
0.3
0.4
0.5
Real (1/sec)
Imag
./(2p
i) (H
z)
YW 1%D
20%D
–0.5 –0.4 –0.3 –0.2 –0.10.2
0.3
0.4
0.5
Real (1/sec)
Imag
./(2p
i) (H
z)
YWS 1%D
20%D
–0.5 –0.4 –0.3 –0.2 –0.10.2
0.3
0.4
0.5
Real (1/sec)
Imag
./(2p
i) (H
z)
N4SID 1%D
20%D
Fig. 1.11 Mode estimates for 16-machine system with a transient from 100 Monte Carlosimulations for each algorithm. Pre- and post-transient damping differ. Ttotal¼ 10 min
20 D. Trudnowski and J. Pierre
The green markers show the corresponding estimates using the entire 20 min
of data. As can be seen in the plot, the variance of the estimated damping
during probing is more than twice as small as the variance during the ambient
condition.
Estimated
Pre-transient
Post-transient
–0.5 –0.4 –0.3 –0.2 –0.10.2
0.3
0.4
0.5
Real (1/sec)
Imag
./(2p
i) (H
z)
YWS 1%D
20%D
–0.5 –0.4 –0.3 –0.2 –0.10.2
0.3
0.4
0.5
Real (1/sec)
Imag
./(2p
i) (H
z)
YW1%D
20%D
–0.5 –0.4 –0.3 –0.2 –0.10.2
0.3
0.4
0.5
Real (1/sec)
Imag
./(2p
i) (H
z)
N4SID 1%D
20%D
Fig. 1.12 Mode estimates for 16-machine system with a transient from 100 Monte Carlosimulations for each algorithm. Pre- and post-transient damping differ. Ttotal¼ 2 min
Fig. 1.13 Brake response ofwestern North Americanpower system. Brakeinserted at the 300 s point.Combined ambient andringdown data from fieldmeasurements. Detrendedpower flowing on a majortransmission line
1 Signal Processing Methods for Small-Signal Dynamic Properties 21
Figure 1.18 shows similar results for varying window sizes. The plot shows
that as the window size is increased, the accuracy of the mode estimate
increases. This is seen in the graph by the decrease in the standard deviation
of the estimate as the window size increases. A bootstrap technique [40] was
used to estimated the standard deviation. Also, in this particular case, estimates
during probing are more than twice as accurate as during ambient conditions
illustrating the improvement in mode estimation performance resulting from
low-level probing.
Fig. 1.14 Frequency estimation of the major modes using the RRLS algorithm
Fig. 1.15 Damping ratio (DR) estimation of the major mode around 0.39 Hz
22 D. Trudnowski and J. Pierre
–20 –15 –10 –5 00.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Real (1/sec)
Imag
. (ra
d/se
c)
YW
AmbientProbingAll AmbientAll Probing
Fig. 1.17 Mode plots for wNAPS probing test series C of the August 2006 probing test.Window¼ 680 s. Window sliding over in one minute sections. Ambient¼ test series C1;Probing¼ test series C2; All ambient¼mode estimates using all of C1; All probing¼modeestimates using all of C2
0 10 20 30 40 50 60500
600
700
800
900
1000
1100
Time from 22-Aug-2006 20:00 GMT (minutes)
Pow
er le
vel (
MW
)
Fig. 1.16 wNAPS response to probing tests. Bottom signal is DC power flow. Top signal is realpower flowing on a major transmission line. 10 min to 17 min region contains brake responses,22 min to 42 min region contains PDCI pseudorandom probing response, 44 min to 47 minregion contains mid-level probing response, and all other regions are ambient responses
1 Signal Processing Methods for Small-Signal Dynamic Properties 23
1.6 Model Validation and Performance Assessment
As seen in the previous examples, there are limits to the accuracy of any
estimator. To fully solve the problem, estimates need to be validated and the
accuracy of the estimates needs to be assessed. The stochastic nature of the
problem cannot be overemphasized. This stochastic nature of the measured
data results in statistical variability of the estimates. The validity of the esti-
mates needs to be tested with asmany techniques as possible and the accuracy of
the estimates need to be determined. This section discusses numerous methods
to validate the mode estimates and to assess the accuracy of those estimates for
both ambient and probing. The resulting estimates of the modes are just that –
estimates (best guesses). A key question is the following: ‘‘is there adequate
agreement between the observed data and the estimated modes?’’
1.6.1 Model Validation
In real-time operation, there are a limited number of ways to validate mode
estimates from ambient data. In this case, one of the most direct ways to
validate the mode estimates is to compare the measured signal’s nonparametric
spectrum with the parametric spectrum corresponding to the estimated modes.
Fig. 1.18 Mode-damping estimate standard deviation for wNAPS probing test in August2006. Probing versus ambient. One multisine cycle¼ 136 s
24 D. Trudnowski and J. Pierre
For example, if the YW method is being applied to estimate the modes, thecorresponding YW parametric spectrum estimate [35] should be compared forconsistency with the nonparametric Welch periodogram spectrum estimate. Ifthere is not a strong similarity in the spectrums, then one may question thevalidity of the mode estimates. Another approach would be to find comparablemode estimation results using different output channels where the modes arehighly observable. Consistency in results between different mode identificationalgorithms would also be evidence of a valid model. For some algorithms,residual analysis [36] can also be employed. Residual analysis studies the portionof the data which the estimated model cannot reproduce and provides a goodindicator of the validity of the estimated models. Another method of validatingthe ambient mode estimates is by comparing those estimates with the estimatesfrom the analysis of a transient response occurring immediately before or after theambient data. This would assume nomajor system configuration change occurredat the transient event and that the true mode values were nearly the same.
When a known probing signal is used, the same validation methods as in theambient case can be used but additional validation methods are possible. Asdiscussed earlier, injecting known probing signals can improve the accuracy ofthe mode estimates. Comparisons can be made between nonparametric fre-quency response estimates and parametric methods. Moreover, validation datamay be used. Validation data are data not used in the estimation of the modes,but used to validate the estimated model. For example, in the wNAPS probingtests carried out in 2000, 2005, 2006, and 2008, short-duration mid-level probingpulses with a peak amplitude of 125 MWwere applied to the PDCI immediatelyfollowing the long-duration low-level probing. The input and output data fromthe low-level probing is used to estimate system transfer functions. The actualsystem response from the mid-level probing is then compared to the responsefrom applying the same mid-level probing signal to the estimated transfer func-tion. Consistency between the actual response and estimated response is a posi-tive indicator of the validity of the estimated transfer function. There are twosources for differences – errors in the transfer function estimate and system noise.Thus the mid-level probing needs to be performed such that a reasonable SNR isachieved so that the difference in the responses coming from the noise is small.The mid-level probing can be applied multiple times and the responses averaged,which increases the SNR by the number of averages.
1.6.2 Performance Assessment
Because the measured time series of data is stochastic (random) in nature, thereis always some variability in the mode estimates around the true value of themode. Thus, the estimates of the modes are essentially a best guess given theobserved time series of data. It is important to try to quantify the quality ofthe estimate. One way to do this is to not only provide a point estimate of themode frequency and damping but also provide an estimate of the mean square
1 Signal Processing Methods for Small-Signal Dynamic Properties 25
error in those quantities, or to provide a confidence interval as the estimateinstead of just a point estimate. For example, instead of just stating that themode-damping ratio estimate is 6%, state that the mode-damping ratio esti-mate is 6% with an RMS error of 1%. This gives some indication of theaccuracy of the estimate. Achieving this measure of performance is a difficulttask. In [26], a bootstrapping approach was first applied to electromechanicalmode estimation. The central idea of the bootstrap method is re-sampling andthe method was first introduced in [41]. Performance assessment is certainly anarea of current and future research.
1.7 Estimating Mode Shape
Similar to the modal damping and frequency information, near-real-timeoperational knowledge of a power system’s mode-shape properties may providecritical information for control decisions. For example, modal shape maysomeday be used to optimally determine generator and/or load trippingschemes to improve the damping of a dangerously low damped mode. Theoptimization involves minimizing load shedding and maximizing improveddamping. This section describes how mode shape can be estimated from time-synchronized measurements.
Results published in [42, 43] demonstrate how one can use spectral analysisto estimate the mode shape from synchronized measurements. The followingsummarizes these results.
1.7.1 Defining Mode Shape
The eigenvalues and eigenvectors for (1.1) are defined from the equations
liI� Aj j ¼ 0; Aui ¼ liui; viA ¼ livi (1:8)
where li is the ith eigenvalue (i= 1. . .n), ui (order n � 1) is the ith righteigenvector, and vi (order 1 � n) is the ith left eigenvector, and I is the n � nidentity matrix. As shown in [4] when considering the ambient case, each systemstate can be written as
xðtÞ ¼Xn
i¼1ziðtÞui (1:9)
where
ziðtÞ ¼ vixðtÞ (1:10)
_ziðtÞ ¼ liziðtÞ þ viBLqðtÞ (1:11)
for i= 1. . .n.
26 D. Trudnowski and J. Pierre
The solution of (1.11) results in zi(t), which is the ith mode’s response to q(t).Equation (1.9) provides information on how the modes are combined to createthe system states. Examination of (1.9) reveals that element ui,k (the kth elementof ui) provides the critical information on the ith mode in the kth state. Theamplitude of ui,k provides the information on the magnitude of mode zi in statexk. It is a direct measure of the observability of the mode in the state. The angleof ui,k provides the information on the phasing of zi in state xk. By comparingthe ffui;k for a common generator state (such as the speed), one can determinephasing of the oscillations for the ith mode. As such, ui has been termed the‘‘mode shape’’ vector [4]. Knowledge of ui provides all the required informationto completely determine the pattern of generator swings for the ith mode [4].
1.7.2 Estimating Mode Shape
As described above, the right eigenvector ui completely describes the modeshape ofmode zi. The question here is: How can the properties of ui be estimatedfrom direct power system measurements without the dependence on the lineardifferential model (1.1)? This section summarizes how spectral analysis providesthe required information; see [42] for more details.
Begin by defining two spectral functions.
Skl !ð Þ ¼ limT!1
1
TE Y�k !ð ÞYl !ð Þ� �
(1:12a)
Skk !ð Þ ¼ limT!1
1
TE Y�k !ð ÞYk !ð Þ� �
(1:12b)
where Sk,l(!) is the cross-spectral density (CSD) function between generalsignals yk(t) and yl(t), Sk,k(!) is the PSD of signal yk(t), Yk(!) is the discreteFourier transform of signal yk(t) at frequency !,Y
�k !ð Þ is the complex conjugate
of Yk(!), and E{} is the expectation operator. These definitions are found inmany signal processing textbooks such as [44].
Now assume that li is a lightly damped mode with
li ¼ �i þ j!i (1:13)
where �i << !i. As shown in [42], the following relationships result
ffSkl !ið Þ ffi ffui;l � ffui;k (1:14)
Skk !ið Þ ffi ui;k�� ��2Knoise (1:15)
where Knoise is an unknown constant.
1 Signal Processing Methods for Small-Signal Dynamic Properties 27
Equations (1.14) and (1.15) are used to estimate the mode shape. Assume forthe moment that all generator speed signals are time-synchronized sampled.Also assume that the frequency of the oscillation mode !i is known. The PSD iscalculated for each generator speed signal. From (1.15), the PSD of each signalis scaled by ui;k
�� ��2; therefore, the PSD is a direct measure of the observability ofthe mode at that generator.
The phasing of the mode among the generators is directly estimated from theangle of the CSD by (1.14). A reference generator with high mode observabilityis chosen as the reference generator k. The angle of the CSD is calculated for allother generators at mode frequency !i. From (1.14), the angle of the CSD forgenerator l represents the phasing of the oscillation.
1.7.2.1 The Coherency
The squared coherency function is defined as
�2k;l !ð Þ ¼Sk;l !ð Þ�� ��2
Sk;k !ð ÞSl;l !ð Þ(1:16)
It represents a measure of the correlation between two signals as a func-tion of frequency [44]. As the two signals become uncorrelated, the coherencyconverges to zero. Similarly, as the signals become totally correlated, thecoherency converges to unity. Basically, it is a measure of percentcorrelation.
As shown in [45], the coherency function can be used to determine if a modeof oscillation is due to one mode or multiple modes at the same frequency. Forexample, given two signals y1(t) and y2(t), if both S1,1(!) and S2,2(!) have peaksat frequency !i, this indicates that the system contains one ormoremodes at thisfrequency. If �21;2 !ið Þ is near unity, then the same mode is contained in both y1and y2. Alternatively, if �21;2 !ið Þ is near zero, then the system contains at leasttwo different modes at frequency !i.
1.7.2.2 Calculating Spectral Terms
Using traditional periodogram averaging methods [44], calculating the PSD,CSD, and coherency requires several minutes of time-synchronized measure-ments. In theory, the measurement should be either the generator angle orspeed; but, in most cases, these signals are rarely time synchronously measured.An excellent approximation to a speed signal is the frequency of the generatorbus voltage (or a nearby bus). Such a signal is obtained from standard synchro-phasor measurements (PMUs, phasor measurement units). More efficientmethods that calculate the required terms from less data are being researchedwith initial results found in [43].
28 D. Trudnowski and J. Pierre
1.7.3 16-Machine Example
This example demonstrates the application of the spectral approach for esti-
mating the mode-shape properties of the 0.36 Hz mode of the 16-machine
system. For comparison, Table 1.4 shows the right eigenvector (ui) terms for
the 0.36 Hz mode and the speed state variable for one generator at each
generation bus. Note that ui has been normalized by the largest term (i.e., the
ui,k for bus 19). Examination of Table 1.4 shows that the mode is primarily a
north–south mode with buses 19 and 20 swinging against buses 18, 22, 17, and
23. Buses 24 and 21 are in the middle of the shape and have very low participa-
tion in the mode. This is indicated by the small value for |ui,k| for these two
buses.The goal is to estimate the mode-shape information in Table 1.4 using only
time-synchronized frequency measurements at the buses. It is assumed that the
mode frequency is known and the system is simulated for 10 min. To compare
with the eigenanalysis results in Table 1.4, bus 19 is selected as the reference bus
for the correlation analysis.Figures 1.19, 1.20, 1.21 show the results for the spectral analysis. The
significant peaks in the PSD and coherency estimates at 0.36 Hz for buses 17,
18, 19, 20, 22, and 23 indicate that the generators connected to these buses all
significantly participate in a single mode at that frequency. The lack of signifi-
cant peaks for buses 21 and 24 indicate that the generators at these buses do not
significantly participate in the mode. The angle of the CSD shown in Fig. 1.21
provides the information for the mode phasing. Bus 20 generators swing in
phase with bus 19 generators, while generators at buses 17, 18, 22, and 23 swing
against bus 19 generators.Using Eqs. (1.14) and (1.15), the spectral analysis results can be directly
compared to the eigenanalysis. The normalized estimated ui;k�� �� is calculated by
taking the inverse decibel from Fig. 1.19 and normalizing by the value for bus
19. Similarly, the estimated ffui;k is directly taken from Fig. 1.21. The results are
shown in Table 1.5. As seen in the table, the spectral analysis very accurately
estimates the eigenvector solution.
Table 1.4 Eigenvector results for 16-machine, 0.361 Hz mode
Bus Angle (ui;k) (degrees) jui;kj (relative)19 0 1.00
20 1 0.76
18 154 0.64
22 151 0.39
17 151 0.38
23 146 0.34
24 61 0.17
21 –7 0.13
1 Signal Processing Methods for Small-Signal Dynamic Properties 29
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8–65
–60
–55
–50
–45
–40
–35
Frequency (Hz)
PS
D (
dB) Bus 17
Bus 18
Bus 19
Bus 20
PS
D (
dB)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8–65
–60
–55
–50
–45
–40
–35
Frequency (Hz)
Bus 21
Bus 22
Bus 23
Bus 24
Fig. 1.19 PSD estimates for 16-machine example. * indicates the 0.361 Hz mode
Coh
eren
cy2
Bus 17
Bus 18
Bus 20
Bus 21
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80
0.2
0.4
0.6
0.8
1
Frequency (Hz)
Coh
eren
cy2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80
0.2
0.4
0.6
0.8
1
Frequency (Hz)
Bus 22
Bus 23
Bus 24
Fig. 1.20 Coherency estimates for 16-machine example. Frequency at bus 19 is the referencesignal. * indicates the 0.361 Hz mode
1.7.4 Field Measured Data
Now consider the actual system data from the wNAPS in Fig. 1.3. The modal
properties of the system are routinely investigated through system probing
testing that employs a synchronized measurement system. The system contains
several inter-area modes including significant ones near 0.25 and 0.37 Hz.
The mode’s frequency and damping are estimated using previously described
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8–200
–100
0
100
200
Frequency (Hz)
CS
D a
ngle
(de
gree
s)
Bus 17
Bus 18Bus 20
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8–200
–100
0
100
200
Frequency (Hz)
CS
D a
ngle
(de
gree
s)
Bus 22
Bus 23
Fig. 1.21 CSD angle estimates for 16-machine example. Frequency at bus 19 is the referencesignal. * indicates the 0.361 Hz mode
Table 1.5 Comparison of eigenvector and spectral analysis for the 0.362Hz mode of the16-machine example
Bus
Eigenvector Spectral analysis
Angle (ui;k)(degrees)
jui;kj(relative)
Angle (ui;k)(degrees)
jui;kj(relative)
19 0 1.00 0 1.00
20 1 0.76 3 0.74
18 154 0.64 161 0.66
22 151 0.39 157 0.41
17 151 0.38 158 0.40
23 146 0.34 157 0.38
1 Signal Processing Methods for Small-Signal Dynamic Properties 31
mode-meter methods. The goal of this section is to demonstrate the application
of spectral analysis to actual systemmeasurement using the 0.25 Hz mode as anexample.
For demonstration purposes, six buses near major generation sitesspread across the system are selected. The actual locations of the sitesare not revealed in order to protect data confidentiality. The buses are
termed bus A through bus F. Nearly 20 min of phasor data was collectedduring a recent probing test from a PMU at each bus. The frequency errorat each bus is estimated from the phasor angle using a forward differen-
cing calculation [33].Results from the spectral analysis are shown in Figs. 1.22–1.24. The
relatively large peaks at 0.25 Hz for buses B through F in the PSD andcoherency indicates that these buses participate in a single mode at thisfrequency. The relatively small peak for bus A indicates that this bus does
not significantly observe this mode. Bus D is selected as the reference busfor the analysis as it has a large peak in the PSD. It should be noted thatany signal with a large peak may be selected as the reference bus in theanalysis.
The CSD in Fig. 1.24 shows the mode phasing. The plot indicates that buses
D and C swing together against buses B, E, and F.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8–65
–60
–55
–50
–45
Frequency (Hz)
PS
D (
dB)
A
BC
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8–60
–55
–50
–45
–40
Frequency (Hz)
PS
D (
dB)
D
EF
Fig. 1.22 Estimated PSD for wNAPS mode-shape example
32 D. Trudnowski and J. Pierre
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8–200
–100
0
100
200
Frequency (Hz)
CS
D a
ngle
(de
gree
s)
B
C
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8–200
–100
0
100
200
Frequency (Hz)
CS
D a
ngle
(de
gree
s)
E
F
Fig. 1.24 Estimated CSD angle for wNAPS example. Bus D is the reference signal. *indicates the 0.25 Hz mode
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80
0.2
0.4
0.6
0.8
1
Frequency (Hz)
Coh
eren
cy2
Coh
eren
cy2
A
BC
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80
0.2
0.4
0.6
0.8
1
Frequency (Hz)
E
F
Fig. 1.23 Estimated coherency for wNAPS example. BusD is the reference signal.* indicatesthe 0.25 Hz mode
1.8 Conclusion
This chapter has presented an overview of many of the more successful analysistechniques and challenges to electromechanical mode estimation from time-synchronized data. The theoretical basis for these methods is described as wellas application and performance properties. Several examples are used todemonstrate some of the more challenging issues. A long list of references isprovided for further details.
Acknowledgments The authors wish to acknowledge the contribution of the many graduatestudents over the years. Also, the technical leadership of Dr. JohnHauer of Pacific NorthwestNational Laboratory (retired) and Mr. Bill Mittelstadt of the Bonneville Power Administra-tion (retired) are acknowledged. Much of this work was supported by the US Department ofEnergy; the authors wish to thank Mr. Phil Overholt for his support.
References
1. DSIToolbox. Posted as freeware at ftp.bpa.gov/pub/WAMS%20Information.2. M. Parashar, J. Mo,‘‘Real Time Dynamics Monitoring System (RTDMSTM): Phasor
applications for the control room’’, 42nd Annual Hawaii International Conference onSystem Sciences (HICSS-42), Big Island, Hawaii, January 5–8, 2009 (accepted).
3. D. J. Trudnowski, J. E. Dagle, ‘‘Effects of generator and static-load nonlinearities onelectromechanical oscillations,’’ IEEE Transactions on Power Systems, vol. 12, no. 3,pp. 1283–1289, Aug. 1997.
4. G. Rogers, Power System Oscillations, Kluwer Academic Publishers, Boston, 2000.5. P. Kundur, Power System Stability and Control, New York: McGraw-Hill, Inc., 1994.6. J. F. Hauer, R. L. Cresap, ‘‘Measurement and modeling of pacific AC intertie response to
random load switchings,’’ IEEE Transactions on PAS, vol. PAS-100, no. 1, Jan. 1981.7. J. Hauer, D. Trudnowski, J. DeSteese, ‘‘A perspective on WAMS analysis tools for
tracking of oscillatory dynamics,’’ IEEE Power Engineering Society General Meeting,paper no. PESGM2007-001391, June 2007.
8. J. F. Hauer, C. J. Demeure, L. L. Scharf, ’’Initial results in Prony analysis of power systemresponse signals,’’ IEEE Transactions on Power Systems, vol. 5, no. 1, pp. 80–89, Feb.1990.
9. L. Scharf, Statistical Signal Processing: Detection, Estimation, and Time Series Analysis,Addison Wesley Publishing Company, Inc., 1991.
10. D. A. Pierre, D. J. Trudnowski, J. F. Hauer, ‘‘Identifying reduced-order models for largenonlinear systems with arbitrary initial conditions and multiple outputs using Pronysignal analysis,’’ Proceedings of the 1990 American Control Conference, vol. 1,pp. 149–154, May 1990.
11. J. F. Hauer, ‘‘Application of Prony analysis to the determination of modal content andequivalent models for measured power system response,’’ IEEE Transactions on PowerSystems, vol. 6, no. 3, pp. 1062–1068, Aug. 1991.
12. D. A. Pierre, D. J. Trudnowski, J. F. Hauer, ‘‘Identifying linear reduced-order models forsystems with arbitrary initial conditions using Prony signal analysis,’’ IEEE Transactionson Automatic Control, vol. 37, no. 6, pp. 831–835, June 1992.
13. D. J. Trudnowski, M. K. Donnelly, J. F. Hauer, ‘‘Advances in the identification oftransfer function models using Prony analysis,’’Proceedings of the 1993 American ControlConference, Vol. 2, pp. 1561–1562, June 1993.
34 D. Trudnowski and J. Pierre
14. J. R. Smith, F. Fatehi, C. S. Woods, J. F. Hauer, D. J. Trudnowski, ‘‘Transfer functionidentification in power system applications,’’ IEEETransactions on Power Systems, vol. 8,no. 3, pp. 1282–1290, Aug. 1993.
15. D. J. Trudnowski, ‘‘Order reduction of large-scale linear oscillatory system models,’’IEEE Transactions on Power Systems, vol. 9, no. 1, pp. 451–458, Feb. 1994.
16. D. Trudnowski, J. Johnson, J. Hauer, ‘‘SIMO system identification from measured ring-downs,’’ Proceedings of the 1998 American Control Conference, pp. 2968–2972, June 1998.
17. D. J. Trudnowski, J. M. Johnson, J. F. Hauer, ‘‘Making Prony analysis more accurateusing multiple signals,’’ IEEE Transactions on Power Systems, vol. 14, no.1, pp. 226–231,Feb. 1999.
18. I. Kamwa, R. Grondin, E. J. Dickinson, S. Fortin, ‘‘A minimal realization approach toreduced-order modelling and modal analysis for power system response signals,’’ IEEETransactions on Power Systems, vol. 8, no. 3, pp. 1020–1029, Aug. 1993.
19. J. Sanchez-Gasca, J. Chow, ‘‘Computation of power system low-order models from timedomain simulations using a Hankel matrix,’’ IEEE Transactions on Power Systems, vol.12, no. 4, pp. 1461–1467, Nov. 1997.
20. L. Guoping, J. Quintero, V. Venkatasubramaniar, ‘‘Oscillation monitoring system basedon wide area synchrophasors in power systems,’’ Bulk Power System Dynamics andControl – VII. Revitalizing Operational Reliability, 2007 iREP Symposium, pp. 1-13,IEEE Digital Object Identifier 10.1145/1176254.1176261, Aug. 19–24, 2007.
21. J. Sanchez-Gasca, J. Chow, ‘‘Performance comparison of three identificationmethods forthe analysis of electromechanical oscillations,’’ IEEE Transactions on Power Systems, vol.14, no. 3, pp. 995–1002, Aug. 1999.
22. J. W. Pierre, R. F. Kubichek, ‘‘Spectral Analysis: Analyzing a Signal Spectrum,’’Tektronix Application Note, http://www.tek.com/Measurement/App_Notes/55_15429/eng/, 2002.
23. J. G. Proakis, D. G. Manolakis, Digital Signal Processing Principles, Algorithms, andApplications, 4th ed., Prentice Hall, 2007.
24. J. W. Pierre, D. J. Trudnowski, M. K. Donnelly, ‘‘Initial results in electromechanicalmode identification from ambient data,’’ IEEE Transactions on Power Systems, vol. 12,no. 3, pp. 1245–1251, Aug. 1997.
25. R. W. Wies, J. W. Pierre, D. J. Trudnowski, ‘‘Use of ARMA block processing forestimating stationary low-frequency electromechanical modes of power systems,’’ IEEETransactions on Power Systems, vol. 18, no. 1, pp. 167–173, Feb. 2003.
26. M.G. Anderson, N. Zhou, J.W. Pierre, R.W.Wies, ‘‘Boostrap-based confidence intervalestimates for electromechanical modes from multiple output analysis of measured ambi-ent data,’’ IEEE Transactions on Power Systems, vol. 20, no. 2, pp. 943–950, May 2005.
27. N. Zhou, J. W. Pierre, R. W. Wies, ‘‘Estimation of low-frequency electromechanicalmodes of power systems from ambient measurements using a subspace method,’’ Pro-ceedings of the North American Power Symposium, Oct. 2003.
28. D. Trudnowski, J. Pierre, N. Zhou, J. Hauer, M. Parashar, ‘‘Performance of three mode-meter block-processing algorithms for automated dynamic stability assessment,’’ IEEETransactions on Power Systems, vol. 23, no. 2, pp. 680–690, May 2008.
29. L. Guoping, V. Venkatasubramaniar, ‘‘Oscillation monitoring from ambient PMU mea-surements by frequency domain decomposition,’’ Proceedings of IEEE InternationalSymposium on Circuits and Systems, pp. 2821–2824, May 2008.
30. R. W. Wies, J. W. Pierre, D. J. Trudnowski, ‘‘Use of least-mean squares (LMS) adaptivefiltering technique for estimating low-frequency electromechanical modes in power sys-tems,’’ Proceedings of the IEEE Power Engineering Society General Meeting, vol. 2, pp.1863–1870, June 2004.
31. N. Zhou, J. W. Pierre, D. J. Trudnowski, R. T. Guttromson, ‘‘Robust RLS methods foronline estimation of power system electromechanical modes,’’ IEEE Transactions onPower Systems, vol. 22, no. 3, pp. 1240–1249, Aug. 2007.
1 Signal Processing Methods for Small-Signal Dynamic Properties 35
32. N. Zhou, D. Trudnowski, J. Pierre, W. Mittelstadt, ‘‘Electromechanical mode on-lineestimation using regularized robust RLS methods,’’ IEEE Transactions on PowerSystems, vol. 23, no. 4, pp.1670–1680, Nov. 2008.
33. J. F. Hauer, W. A.Mittelstadt, K. E.Martin, J. W. Burns, H. Lee, In association with theDisturbance Monitoring Work Group of the Western Electricity Coordinating Council,‘‘Integrated dynamic information for the western power system: WAMS analysis in2005,’’ Chapter 14 in the Power System Stability and Control volume of The ElectricPower Engineering Handbook, edition 2, L. L. Grigsby ed., CRC Press, Boca Raton,FL, 2007.
34. N. Zhou, J.W. Pierre, and J. F. Hauer, ‘‘Initial results in power system identification frominjected probing signals using a subspace method,’’ IEEE Transactions on Power Systems,vol. 21, no. 3, pp. 1296–1302, Aug. 2006.
35. P. Stoica and R. Moses, Introduction to Spectral Analysis, Prentice Hall, New Jersey,1997.
36. L. Ljung, System Identification Theory for the User, 2nd Ed., Prentice Hall, Upper SaddleRiver, NJ, 1999.
37. D. J. Trudnowski, J. R. Smith, T. A. Short, and D. A. Pierre, ‘‘An application of Pronymethods in PSS design for multimachine systems,’’ IEEE Transactions on Power Systems,vol. 6, no. 2, pp. 118–126, Feb. 1991.
38. D. Trudnowski, M. Donnelly, and E. Lightner, ‘‘Power-system frequency and stabilitycontrol using decentralized intelligent loads,’’ Proceedings of the 2005/2006 IEEE PEST&D Conference and Exposition, pp. 1453–1459, May 2006.
39. D. J. Trudnowski, J. W. Pierre, and N. Zhou, ‘‘Performance and properties of ambient-data swing-mode estimation algorithms, version 1.0,’’ Report no. ENGR 2006-1, Engi-neering Dept., Montana Tech of the University of Montana, Butte, MT, USA, 2006.
40. F.K. Tuffner, ‘‘Computationally efficient weighted updating of statistical parameter esti-mates for time varying signals with application to power system identification,’’ Ph.D.dissertation, Department of Electrical andComputer Engineering,University ofWyoming,Laramie, WY, USA, 2008.
41. B. Efron, R. Tibshirani, ‘‘Bootstrap methods: another look at the jackknife,’’ The Annalsof Statistics, vol. 7, no. 1, pp. 1–26, 1979.
42. D. Trudnowski, ‘‘Estimating electromechanical mode shape from synchrophasor mea-surements,’’ IEEE Transactions on Power Systems, vol. 23, no. 3, pp. 1188–1195, Aug.2008.
43. L. Dosiek, D. Trudnowski, J. Pierre, ‘‘New algorithms for mode shape estimation usingmeasured data,’’ IEEEPower & Energy Society GeneralMeeting, paper no. PESGM2008-001014, July 2008.
44. J. S. Bendat, A. G. Piersol, Engineering Applications of Correlation and Spectral Analysis,2nd Ed., John Wiley & Sons, New York, 1993.
45. D. Trudnowski, J. Hauer, J. Pierre, W. Litzenberger, D. Maratukulam, ‘‘Using thecoherency function to detect large-scale dynamic system modal observability,’’ Proceed-ings of the 1999 American Control Conference, pp. 2886–2890, June 1999.
36 D. Trudnowski and J. Pierre
Chapter 2
Enhancements to the Hilbert–Huang Transform
for Application to Power System Oscillations
Nilanjan Senroy
Abstract TheHilbert–Huang transform is introduced for time–frequency analysisof oscillatory signals representing power system dynamic behavior. Fundamen-tal assumptions of the Hilbert–Huang transform are revisited, particularly theability of empirical mode decomposition to yield monocomponent intrinsicmode functions. In the context of the specific application, some enhancementsto the original algorithms are discussed. A wide variety of application examplesare employed to demonstrate the efficacy of the improved Hilbert–Huangtransform.
2.1 Introduction
Modern interconnected power system dynamics is characterized by oscillatorybehavior. These oscillations are produced as a result of a variety of disturbancessuch as changes in loads, tripping of lines, faults, and other discrete events. Theoscillations manifest themselves as variations in line flows and generator angleexcursions. A detailed study of these oscillations is necessary to gain a thoroughunderstanding of the system dynamics.
Modern computer processing speeds and memory have allowed engineers toapply analysis techniques that were previously considered impractical due totheir computational burden. Increasing usage of strategically located sensorsand measuring devices is leading to vast amounts of data, whose analysiscan lead to a deeper insight into the underlying processes. In this context, it isworthwhile to mention the advent of wide-area measurement systems(WAMSs) in modern power systems. Using such data, utilities are increasinglyrelying on distributed measurement and control of power systems. An effectiveway to process such real-time measurement data is to analyze the constituent
N. Senroy (*)Department of Electrical Engineering, Indian Institute of Technology, New Delhi, Indiae-mail: [email protected]
A.R. Messina, Inter-area Oscillations in Power Systems,Power Electronics and Power Systems, DOI 10.1007/978-0-387-89530-7_2,� Springer ScienceþBusiness Media, LLC 2009
37
modes. The frequency and damping of these modes indicate the nature andseverity of a disturbance, reveal the health of the system, and predict the out-come of an evolving contingency. Such information can be further used to plancontrol actions, instill situational awareness, and enable health monitoring andpreventive maintenance procedures.
Spectral analysis of power system oscillatory signal data is a challenging task.As power systems are inherently nonlinear, the signals representing their behaviorcontain possible time-varying waveform distortions. Further, modal interactionbetween closely spaced frequency components may lead to nonlinearities in thesignal. Thus, the spectral analysis of these signals is not a trivial task.
A popular tool for spectral analysis is the fast Fourier transform (FFT). TheFFT basically transforms a distorted signal from time domain to frequencydomain by resolving it into stationary equivalents, with frequency and ampli-tude information. Therefore, in the event of nonstationary components in theoriginal distorted signal, FFT results are suspect. The challenge of estimatingnonstationary distortion components is met by computing the FFT in a finitetime window. The signal components are assumed to be stationary within thistime window. The complete signal is analyzed by sliding the FFT window alongthe entire duration of the signal. The size of the analysis window must be largeenough to allow all the constituent modes to manifest completely, yet smallenough to satisfy assumptions of linearity and stationarity. This is a dauntingtask particularly in the case of power system oscillatory signals which arefrequently damped or ‘ringdown’ signals, containing possible nonlinearinteractions.
An alternative approach to the spectral analysis of power system oscillatorysignals is to repose the problem as an instantaneous frequency/amplitudetracking problem. Such an approach implies retaining the time-domain infor-mation of the signal without compromising on its frequency-domain analysis.The Hilbert–Huang technique is one such technique, recently proposed, totrack the temporal variations in the frequency and amplitude of the variouscomponents within a time-varying distorted signal. It was initially proposed forgeophysics applications, but has been widely applied with success in problemsin biomedical engineering, image processing, and structural safety. It has alsobeen applied in power systems in the area of power quality, subsynchronousresonance, and to analyze inter-area oscillations. The original Hilbert–Huangtechnique consists of the empirical mode decomposition (EMD) followed bythe Hilbert transform. While the Hilbert transform is a well-known mathema-tical technique widely applied in signal processing and communications, theEMDwas developed byNorden Huang [1]. As the name suggests, the EMD is adata-driven technique, empirical in nature with limited analytical justificationas of date.
There are some shortcomings of the original Hilbert–Huang technique, parti-cularly in the EMD algorithm, that limit its application to signals of the typeencountered in power system stability studies. The intention of this chapter is tohighlight these shortcomings and further resolve them using come enhancements
38 N. Senroy
to the EMD technique. The application of the enhanced Hilbert–Huang techni-que is demonstrated for a wide variety of problems in power system monitoringand control.
2.2 Hilbert–Huang Transform
Before going into the modified Hilbert–Huang transform, a brief exposition ofthe fundamental Hilbert–Huang transform is in order. TheHilbert–Huang trans-form consists of the EMD followed by the Hilbert transform. The original EMDalgorithm as provided in [1] is referred to as the standard EMD in this chapter.
2.2.1 Empirical Mode Decomposition
The underlying philosophy behind EMD is the concept of instantaneous fre-quency. From a Fourier analysis point of view, the frequency of a signal wouldbe derived from its time period, which is the time taken to complete onestationary time period. Therefore, for a nonstationary waveform whereinthere exist incomplete time periods, the frequency would be hard to define.Essentially, a full wave is required to be present in the signal, to be recognized asa legitimate frequency. However, the frequency is also equivalent to the angularvelocity which can be defined as the rate of change of phase. Hence, if a uniquephase can be defined for a real-valued signal, it would be easy to compute itsrate of change of phase and thereby its frequency. This point will be furtherelaborated in the next subsection.
The frequency obtained in this manner is unique at any instant in time, andhence termed as instantaneous frequency. In other words, it is possible to defineonly one instantaneous frequency for a signal at any point in time. This poses aproblem for multicomponent signals which have more than one frequencycomponents existing at a given time. The instantaneous frequency obtainedfor such a signal would be meaningless, unless the individual components areisolated before applying Hilbert transform on them. EMD is a method pro-posed by Norden Huang to decompose a multicomponent waveform intointrinsic mode functions (IMFs) that have well-defined Hilbert transforms.
The EMD technique essentially involves identifying a baseline signal existingin a signal. Such a baseline signal, when subtracted from the original signal,leaves behind a monocomponent signal, i.e., whose Hilbert transform is welldefined. This is called an IMF, which is characterized by the followingproperties:
� local uniformity around zero,� all the maxima (minima) are greater (less) than zero,� the numberof extremaand thenumberof zeros are equalordiffer byatmost one.
2 Enhanced Hilbert–Huang Technique for Power System Oscillations 39
The baseline signal is identified as the local mean of an upper signal envelopeand lower signal envelope. The upper signal envelope arises from the interpola-tion between the local maxima points, while the lower envelope is similarlyobtained from the local minima points. The steps to obtain an IMF from adistorted signal are as follows:
A1. identify local extrema of the signal, s(t),A2. interpolate between maxima (minima) to obtain the upper (lower) envel-
ope, eM(t) (em(t)),A3. compute the baseline signal, m(t) = (eM(t) + em(t))/2A4. extract c(t) = s(t) – m(t),A5. if c(t) is not an IMF then improve it by applying steps A1–A4 replacing s(t)
by c(t), until the new c(t) obtained satisfies the conditions of being an IMF,A6. if c(t) is an IMF, then compute the residue, r(t) = s(t) – c(t),A7. if r(t) is not monotonic, then repeat steps A1–A6 replacing s(t) with r(t), to
obtain the next IMF.
In practice, an appropriate stopping criterion must be used in step A5 toavoid overimproving the IMF, as that can potentially lead to loss of modalinformation. Once all the IMFs have been extracted, the final residue is mono-tonic in nature, i.e., lack of extrema points. The IMFs ‘sifted’ in this mannerfrom a distorted signal are orthogonal in nature, and appear in decreasing orderof frequency. In other words, the first IMF extracted contains the highestfrequency component, while the last IMF extracted contains the lowest fre-quency component. If n IMFs are extracted in this manner, the original dis-torted signal can be recovered from the IMFs as follows:
sðtÞ ¼X
n
ciðtÞ þ rðtÞ (2:1)
2.2.2 Hilbert Transform
In 1946, Gabor developed the concept of an analytic signal to provide anunambiguous definition of the phase of a signal [2]. If the instantaneousfrequency is defined as the rate of change of phase, the instantaneous phase ofthe signal is required. For a real-valued signal, its instantaneous phase cannotbe computed without defining its imaginary counterpart. While there are infi-nite ways of arriving at an imaginary counterpart of a given real signal, theappropriate method, as defined by Gabor, is as follows. In the Fourier trans-form of the original real-valued signal, the negative frequencies are suppressed,and an inverse Fourier transform applied to the resultant spectrum. The signalthus obtained is complex in nature and is referred to as the analytic form of theoriginal signal. The real part of this analytic signal is the original real-valuedsignal, while its imaginary part equals the Hilbert transform of the original
40 N. Senroy
signal. Hence, the first step toward the calculation of the instantaneous fre-quency (and amplitude) of a signal, is the application of the Hilbert transform.
The Hilbert transform of a real-valued signal, s(t), is defined as
sH tð Þ ¼ H x tð Þf g ¼ 1
p}
Z1
�1
x tð Þt� y
dt (2:2)
where } indicates the Cauchy’s principal value of the integral. Accordingly, theintegral is evaluated in the range [–1, y– e] and [y+ e,1]. The analytic form ofs(t) can be defined from its Hilbert transform as sA(t) = s(t) + jsH(t) , such that
= sA tð Þf g ¼ Re = s tð Þf gf g (2:3)
where = �f g refers to the Fourier transform operator. The instantaneous phaseangle of sA(t) is accordingly defined as
� tð Þ ¼ tan�1Im sA tð Þf gRe sA tð Þf g ¼ tan�1
sH tð Þs tð Þ (2:4)
The phase, thus obtained, must be unwrapped to be meaningful.
2.3 Modified Hilbert–Huang Transform
In this section, the limitations of the standard EMD are highlighted, and theenhancements to the EMD are presented. The modified Hilbert–Huang trans-form was initially suggested as an analysis technique for signals encountered inpower quality [3]. These signals are characterized by the presence of a strongfundamental frequency (50 or 60 Hz), distorted by the presence of higherharmonic and inter-harmonic frequency components that are relatively weakerin magnitude. Additionally, the frequencies of these components lie within anoctave posing a challenge in separation. Encouraged by the results, theenhanced algorithms were applied to signals representing power system oscilla-tory behavior. While there are significant differences between signals encoun-tered in power quality studies and those that are used in this chapter, thephilosophy behind the enhanced Hilbert–Huang transform remains the samefor both kind of signals. The nature of the signals that are the focus of thischapter is (a) signals are nonstationary and possibly nonlinear, (b) signals are oflow frequency in the range of 0–2 Hz, and (c) individual component frequencieslie within an octave.
Before proceeding further, an explanation of a frequency octave is provided.An octave is the frequency range between one frequency and its double or half-frequency. Examples of octaves are 7.5–15 kHz and 0.5–1.0 Hz. Two
2 Enhanced Hilbert–Huang Technique for Power System Oscillations 41
frequencies are said to share an octave if their ratio lies between 0.5 and 2.Hence, 0.45 and 0.6 Hz share an octave, while 0.35 and 0.8 Hz lie in differentoctaves.
2.3.1 Limitations of EMD
It is interesting to note here the similarity between an IMF and a real-zero (RZ)signal as defined by [4]. RZ signals are bandpass signals whose zeros are all realand distinct. A subclass of bandpass signals that share only real simple zeroswith their Hilbert transform, is completely described by zero crossings only ifthe bandwidth is less than an octave. An IMF would fall in this subclass ofsignals. Extensive experiments have already established the ability of EMD todecompose white noise into IMFs whose frequency spectrum comprises anoctave. The mean frequencies of the extracted IMFs show a doubling phenom-enon. Thus, the implication is that if a signal contains two modes whosefrequencies lie within an octave, standard application of EMD is unable toseparate the modes.
Figure 2.1 shows the result of application of EMD to separate the constitu-ent frequencies in a distorted signal. When the signal is of the form sin pt +0.7 sin 0.3pt, the standard application of EMD yields two monocomponentIMFs corresponding to the two frequency components present in the signal.These two frequencies are 0.5 and 0.15 Hz, both of which clearly lie in differentoctaves. However, when the signal is of the form sin pt + 0.7 sin 0.75pt, stan-dard application of EMD fails to yield the two components. Rather, the firstIMF comes out to be identical to the original signal as shown in Fig. 2.1. This isbecause the two frequencies, 0.5 and 0.375 Hz, lie within the same octave.
0 10 20–2
0
2 s1 = sinπt + 0.7sin0.3πt
time (s)
0 10 20–2
0
2 IMFs extracted from s1
time (s)
0 10 20–2
0
2 s2 = sinπt+0.7sin0.75πt
time (s)
0 10 20–2
0
2 IMFs extracted from s2
time (s)
Fig. 2.1 Standardapplication of EMD toextract IMFs from amulticomponent signal.When the signalcomponents havefrequencies 1 and 0.3 Hz, thetwo IMFs obtained aremonocomponent. When thesignal constituentfrequencies are 1 and0.75 Hz, the first IMF isidentical to the signal
42 N. Senroy
Based on the above example, a 30 s synthetic signal sampled at 100 Hz,
s(t)=10 sin 2pt + m sin 2pft where 5 � m�40, 0 � f� 1, was analyzed using
the standard EMD. The signal consists of two components. The first compo-
nent was held steady at 1 Hz, with amplitude of 10. The second component’s
frequency and amplitude varied from 0 to 1 Hz and 5 to 40, respectively. The
standard EMD was applied to the signal, and the spectral content of the first
IMF was analyzed using FFT. Note, the signal is a stationary signal, and FFT
is expected to yield the spectral content accurately. Normally, the first IMF is
expected to contain the highest frequency component, which is this case is
10 sin 2pt. The multicomponent character of the first IMF arises from the
inclusion of the second component (whose frequency and magnitude are
being varied), by EMD. Hence, the percentage of the second component
included in the first IMF was analyzed, and plotted in Fig. 2.2. In Fig. 2.2,
the x-axis shows the variation of the frequency of the second component, while
the y-axis shows the variation of the ratio of the amplitudes of the two
components.
01
2
3
4
prop
ortio
n of
IInd
com
pone
ntex
trac
ted
in Is
t IM
F
0.5
1
1.5
0.2
frequency of IInd component (Hz)
amplitude ratio of IInd
component toIst component
0.40.6
0.8
Fig. 2.2 Percentage of lower frequency component included in the first IMF by the standardapplication of EMD. The second component frequency was varied from 0 to 1 Hz (x-axis),while its amplitude was varied from half to four times the amplitude of the first component(y-axis)
2 Enhanced Hilbert–Huang Technique for Power System Oscillations 43
FromFig. 2.2, some observationsmay bemade. As the frequency of the second
component approaches the frequency of the first component, the likelihood of it
being included in the first IMF increases. For relatively small amplitudes of the
second component, a frequency of more than 0.5 Hz means that the standard
EMDwill not be able to separate the two components. Additionally, as the second
component amplitude increases, the ‘cutoff’ frequency for the first IMFdecreases.
In Fig. 2.2, if the second component amplitude is four times the first component
(i.e., the 1 Hz component), it will be included in the first IMF even if its frequency
is less than 0.4 Hz. It may thus be concluded that the first IMF obtained by
applying the standard EMD to a distorted signal, may be multicomponent underthe following conditions:
1. if the highest two frequencies in the distorted signal fall in an octave,2. if the distorted signal consists of a weak high-frequency component along
with a dominating lower frequency component.
2.3.2 Masking Signal-Based EMD [5]
The use of masking signals has been proposed to enhance the discriminatingcapability of EMD [5]. A masking signal of frequency higher than the highest
frequency component present in the original distorted signal is artificially added
and subtracted from the signal to obtain two new signals. EMD is performed on
these two new signals to obtain the first IMF only. The average of the two IMFs
is then computed to yield the correct IMF for the original signal. This ‘improved’
IMF would be monocomponent only if the frequency of the masking signal is
chosen such that only the highest frequency component in the original signal falls
in the same octave as the masking signal [3]. A systematic method to construct
masking signals to apply on a distorted signal is given as follows:
B1. Perform FFT on the distorted signal, s(t) to estimate frequency compo-nents f1, f2, . . . , fn, where f15f25 � � �5fn. In the case of a complicatedsignal, f1, f2, . . ., fn are stationary equivalents of the possibly time-varyingfrequency components.
B2. Construct masking signals, mask2, mask3, . . ., maskn, where maskk =Mk sin(2p( fk+fk – 1)t). In the case of power quality signals [3], the valueofMk is suggested to be 5.5 times the magnitude of fk obtained in the FFTspectrum. In experiments with power system oscillation signals, this valueis found to work satisfactorily. However, it is stressed here that the value ofMk is empirical, and individual experiences about appropriate maskingsignal amplitudes may vary.
B3. Compute two signals, s(t) + maskn and s(t) – maskn Perform EMD (stepsA1–A6) on both the signals to obtain their first IMFs only, IMF+ andIMF_. Then c1(t) = (IMF+ + IMF_)/2. This is the correct IMF of thedistorted signal.
44 N. Senroy
B4. Obtain residue, r1(t) = s(t) – c1(t).B5. Perform steps B3 and B4, replacing s(t) with the residue obtained in step B4,
iteratively until n – 1 IMFs containing frequency components f2, f3, . . ., fn havebeen extracted. The final residue rn(t) will contain the remaining component f1.
As a demonstration of the masking signal-based EMD, consider the second
signal of Fig. 2.1, i.e., s(t) = sinpt + 0.7 sin 0.75pt. Using a masking signal,
5.5 sin 3.5pt, two IMFs were obtained, as shown in Fig. 2.3. The first IMF was
monocomponent in nature, as revealed by its FFT. However, there is a significant
loss of amplitude information of the 1 Hz component in the first IMF. The second
IMF contains traces of the 1 Hz component along with the 0.75 Hz component.
2.3.3 Frequency Heterodyne Technique [6]
Frequency shifting has also been proposed as an alternative to the masking
signal-based EMD [6]. As demonstrated in the previous section, the masking
0 5 10 15 20–2
0
2 sinπt + 0.7sin0.75πt
time (s)
0 5 10 15 20–2
0
2 IMFs extracted
time (s)
0 1 2 3 4 50
0.2
0.4
0.6FFT of first IMF
frequency (Hz)
III
Fig. 2.3 Masking signal based EMD on a signal, s(t) = sinpt+ 0.7 sin 0.75pt. The first IMFobtained is monocomponent as evident from its FFT. However, there is significant loss ofamplitude information in its retrieval
2 Enhanced Hilbert–Huang Technique for Power System Oscillations 45
signal-based EMD sometimes leads to loss of amplitude information while
extracting the highest frequency component. While the philosophy behind
masking signals is to include the highest frequency component in a different
octave, the basic idea behind frequency heterodyne technique is to alter the
spectral distance between two adjacent frequency components. This is achieved
by nonlinear mixing of the distorted signal with a pure tone of frequency greater
than the highest frequency present in the distorted signal. This is possible only if
the frequencies of the distorting components are known in advance. The prin-
ciple is similar to the heterodyne detection commonly used in communication
theory. The signal of interest is amplitude modulated by mixing it with an
appropriate ‘carrier signal.’ Such a signal is referred to as double sideband
modulated (DSB) with suppressed carrier as it contains two frequency shifted
copies of the original signal on either side of the carrier frequency. Further, a
single sideband modulated (SSB) signal may be obtained by removing one of
the sidebands using an appropriate filter or a Hilbert transformer. This also
effectively reduces the bandwidth of the modulated signal.Consider a signal, s(t), with its Hilbert transform sH(t). The frequency content
of s(t) is shifted around a new carrier frequency, F, bymultiplying it by the analytic
representation of the carrier signal. The DSB signal thus obtained is as follows:
fDSB tð Þ ¼ s tð Þe j2pFt (2:5)
The SSB signal is obtained by using the analytic form of s(t). The lower
sideband signal is obtained as follows
fSSB tð Þ ¼ Re s tð Þ þ jsH tð Þð Þe�j2pFt� �
(2:6)
In the context of EMD, suppose the distorted signal contains two frequencies
f1 and f2 ( f1< f2), both of which lie in an octave. The heterodyne frequency, F, is
selected as f2 < F < 2f2 – f1. The length of the original distorted signal will
determine how close F can be to f2. After heterodyning, the SSB frequencies
obtained are F– f1 and F– f2. If the value of F is chosen correctly, the two shifted
frequencies will lie in different octaves, and subsequent application of the
standard EMD results in separation of the components as individual IMFs of
frequencies F– f1 and F– f2. The IMFs are then translated back to the original
frequencies using the same heterodyne technique.The same signal of Fig. 2.3 was heterodyned with a signal of frequency
1.2 Hz. The result was the shifting of the 1 Hz component to 0.2 Hz, and the
0.75 Hz component to 0.45 Hz. Figure 2.4 shows the heterodyne technique as
applied to the signal. The new heterodyned signal contains 0.2 and 0.45 Hz,
which do not fall in the same octave. Hence, regular application of EMD is able
to separate the two components in different IMFs, which are subsequently
heterodyned back to their original frequencies. Figure 2.5 shows the IMFs
46 N. Senroy
0 0.5 1 1.5 20
0.2
0.4
0.6
0.8
1
X: 0.9995Y: 0.9984
X: 0.7496Y: 0.7015
X: 0.4498Y: 0.6991
X: 0.1999Y: 1
X=1.2Y=1
frequency (Hz)
heterodynefrequency
originalfrequencies
heterodynedfrequencies
Fig. 2.4 Frequency shifting using the heterodyne technique. The original signal contained twofrequency components at 1 and 0.75 Hz. After heterodyning with 1.2 Hz signal, thefrequencies were shifted to 0.2 and 0.45 Hz, respectively
0 5 10 15 20–2
–1
0
1
2 sinπt+0.7sin0.75πt
time (s)
0 5 10 15 20–1.5
–1
–0.5
0
0.5
1
1.5 IMFs obtained by heterodyne technique
time (s)
Fig. 2.5 IMFs obtained from applying standard EMD after heterodyning the distorted signal.Both the IMFs are monocomponent in nature with negligible loss of amplitude information
2 Enhanced Hilbert–Huang Technique for Power System Oscillations 47
obtained finally, from the distorted signal of Figure 2.3. Both the IMFs aremonocomponent and suffer negligible loss of amplitude information.
The error in extraction from the frequency heterodyne technique is signifi-cant at the edges of the window of observation and near switching instants only.This is in contrast to the masking signal-based EMD, wherein the error inextraction persists throughout the observation window. However, one of thesignificant drawbacks of the frequency heterodyne technique is the difficulty inselection of an appropriate heterodyne frequency. Accurate knowledge of thespectral content of the distorted signal is required before making a decision onthe heterodyne frequency. This knowledge could be obtained from an FFT onthe distorted signal; however, FFT-based spectral analysis of time-varyingdistorted signals yields approximate and possibly inaccurate information.
2.4 Case Studies
In this section several case studies are presented that demonstrate the applica-tion of the modifications to the Hilbert–Huang technique. All the signals to beanalyzed are derived from power system applications. However, the exactnature of the signal varies according to the particular application. The firstcase is the study of power flow oscillations in a critical transmission line in alarge interconnected power system, in response to a contingency. The objectiveis to characterize the amplitude and frequency of the undamped oscillationsusing the frequency heterodyne-based Hilbert–Huang technique. The secondcase is the application of the masking signal-based Hilbert–Huang techniquefor time–frequency–magnitude characterization of the oscillations in load tor-que and field current, in a specialized ship propulsion motor. The third case isthe application of the masking signal-based EMD to understand slow coher-ency in multimachine power systems. The fourth and final case study involvesactual wide-area measurement signals of frequency, recorded in the westernUnited States in 2005.
2.4.1 Power Flow Oscillations in Large Power Systems
This section presents the application of the Hilbert–Huang transform toanalyze the power flow oscillations in a six-area, 377-machine detailedmodel of the Mexican interconnected system [7]. Time-domain simulation ofthe outage of one of the units in the Laguna Verde nuclear power plant, wascarried out. The contingency involved disruption of 650 MW without anyfault in the southeastern part of the system. No supplementary dampingcontrollers like power system stabilizer were considered in the study.Figure 2.6 shows the resultant oscillations in the real power flow on the230 kV transmission line fromMalpaso Dos (MPD) substation toMacuspanaDos (MCD) substation, due to the tripping of the Laguna Verde unit. The
48 N. Senroy
contingency is a critical contingency as it results in undamped oscillations.
The FFT (Fig. 2.6) reveals prominent modal frequencies at 1.5 and 0.8 Hz.The 230 kV MPD–MCD transmission line is located within the southeastern
area, and the oscillations in its line flow represent the localized interaction
between machines within the area. The objective of the study is to estimate the
damping and frequency of the oscillations. Due to the close spacing of the
frequency components, the standard EMD was unable to discriminate between
the two principal modes. A masking signal-based EMD was able to accurately
extract the frequencies of the two modes; however, some amplitude information
was lost in the process. Hence, the MPD–MCD signal was heterodyned with
a frequency of 2.20 Hz. Using the procedure outline in Section 2.3.3, the
two principal modes were separated. Further application of Hilbert transform
yielded their instantaneous frequency and amplitude. Figure 2.7 shows the almost
monotonic rise in amplitude of both the frequency components (1.5 and 0.8 Hz).
This is a typical undamped oscillatory behavior, wherein the damping may be
accurately computed from the variation of the amplitude shown in Fig. 2.7.
0 5 10 15 20 25 30–50
0
50
time (s)
[MW
]
line flow
0 0.5 1 1.5 2 2.5 3 3.5 40
5
10
15
20
25 FFT of signal
frequency (Hz)
Fig. 2.6 Oscillations in thereal power flow on the230 kV MPD–MCDtransmission line in responseto the tripping of the LagunaVerde 650 MW unit. TheFFT of the oscillationsreveal prominent modalfrequencies at 1.5 and 0.8 Hz
0 5 10 15 20 25 300
20
40
60 instantaneous magnitude
time (s)
[MW
]
0 5 10 15 20 25 300
1
2
3
4 instantaneous frequency
time (s)
[Hz]
Fig. 2.7 Instantaneousfrequency and amplitudes ofthe two most prominentIMFs extracted from theoscillations in the line flowof the 230 kV MPD–MCDtransmission line. Note, themonotonic rise in amplitudeover time signifyingundamped oscillations
2 Enhanced Hilbert–Huang Technique for Power System Oscillations 49
2.4.2 Torque and Field Current Variations in HTS PropulsionMotors [8]
A slightly different kind of application of the modified Hilbert–Huang trans-
form is presented in this section. Moving away from traditional interconnected
power systems, this section devotes itself to shipboard power systems. Typi-
cally, an integrated power system on an all-electric ship implies that electricity is
used to fulfill all the energy needs of the ship including propulsion. Thus,
specialized electric machines are employed as propulsion motors that are not
only energy efficient but also contribute to the robustness of the integrated
power system.A critical aspect of machine design is the understanding of electromagnetic
stator–rotor interactions during varying load conditions. Recently, there has been
renewed interest in the application of high-temperature superconducting (HTS)
motors for the propulsions needs of future generations of warships [8]. HTS
motors are characterized by a significantly high-power density and efficiency
when compared to conventional motors. They are typically synchronousmachines
with theHTS technology for the rotor field winding only, carrying theDC current.
The stator armature containing the AC windings follow conventional technology.
While in operation at sea, a propulsion motor is subject to the application of
varying sea states, which lead to slow variations in load torque. Such low-fre-
quency oscillations in the torque result in variations in the field current. Joule
heating due to these current variations in the cryogenically cooled field windings
has significant implications on the refrigeration aspects of the machine.A transfer function between the load torque variations and the field current
oscillations is very difficult to obtain analytically. Even with finite element-
based models, computing the relationship is a challenging task, especially in the
presence of varying rotor speed. Appropriate measurements of the relevant
quantities can enable transfer function estimation, provided suitable empirical
methods are applied. Traditionally, empirical methods for transfer function
estimation are primarily FFT based, wherein the FFT of an artificially intro-
duced input signal, torque, is compared with the FFT of the output signal, field
current. Such an exercise entails systematic analysis, wherein the machine may
have to be taken offline. During sea-state conditions, the transfer function
estimation is possible only if an appropriate technique is available to study
time-varying waveform distortions. The modified Hilbert–Huang technique is
applied to accurately estimate the relationship between the motor torque and
field current variations. Figure 2.8 shows an example of the oscillations in load
torque and field current due to the application of a sea state.FFT analysis of the torque variations revealed possible modal frequencies at
2.3, 1.74, 1.26, 0.87, and 0.535Hz. Accordingly, themasking signal-based EMD
was applied, with the masking frequencies fixed at 4.04, 3.00, 2.13, 1,405, and
0.815 Hz. Similarly, the principal modes identified from an FFT analysis of the
field current variations were 1.8, 1.4, 1.07, 0.87, 0.535, and 0.28 Hz.
50 N. Senroy
Accordingly, the masking frequencies were fixed at 3.2, 2.47, 1.94, 1.405, 0.815,
and 0.38 Hz. Figure 2.9 shows the instantaneous frequencies of the principal
IMFs extracted using the masking signal-based EMD. It can observed that the
frequency tracking improves for lower frequencies. The reason for this is that
the amplitudes at these frequencies are greater. This is an inherent property of
the Hilbert–Huang technique. Another significant observation can be made
from Fig. 2.9. When the instantaneous frequencies of the modes present in the
15 20 25 30–0.1
–0.05
0
0.05
0.1
15 20 25 30–0.02
–0.01
0
0.01
0.02
time (s)
field current variationstorque variations
Fig. 2.8 Typical variations of load torque and field current of a HTS propulsion motor due tothe application of sea states. The axis for the field current variations is to the right, while thesame for the torque is to the left
15 20 25 300
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
time (s)
freq
uenc
y (H
z)
Fig. 2.9 Instantaneous frequencies extracted from torque and field current variations in anotional HTS motor due to the application of sea states. The dashed lines represent theinstantaneous frequencies of field current variations, while the solid lines represent the same forthe torque variations. Note, both the field current and torque contain identical frequency modes
2 Enhanced Hilbert–Huang Technique for Power System Oscillations 51
load torque are compared with that of the modes present in the field current,there is a clear correspondence. This indicates that all the significant modespresent in the load torque oscillations also appear in the field current variations.
Because the application of the sea states result in the excitation of a fewmodes in the load torque and field current, the relationship between the two canbe gauged for those modal frequencies only. The instantaneous amplitude gainwas calculated directly from the instantaneous amplitudes of the most domi-nant IMFs, whose frequencies are shown in Fig. 2.9. A scatter plot was used torepresent this gain, because for every sampling instant, there is a calculatedamplitude gain. In order to validate the computed amplitude gains, a traditionalFFT-based empirical transfer function estimation was also carried out for awide range of frequencies. This technique involves calculating the ratio of theamplitude of the output (oscillations in the field current) to a single frequencyinput (load torque variation). The system was found to respond linearly overthe entire frequency range, except for oscillation frequencies close to the naturalmodes of the system [8]. Figure 2.10 shows a reasonably good agreementbetween the trend of the scatter plots obtained from online sea-state analysisand the frequency response curve obtained independently using off-line FFT-based techniques. The spread evident in the scatter plots is an inherent artifactof the Hilbert–Huang technique, which defined instantaneous frequency as therate of change of the instantaneous phase of the analytic form of a real-valuedsignal. The spread at any frequency may reduce considerably if the amplitudesof the respective oscillations are higher.
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 20
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
frequency (Hz)
tran
sfer
func
tion
gain
Fig. 2.10 Transfer function gain between variations in the field current and load torque of aHTS propulsion motor. The scatter points represent results from the Hilbert–Huangtransform using measurements taken during simulated sea-state conditions. The solid linerepresents FFT-based off-line empirical transfer function estimation. Note, the increase inscatter as the frequency increases, because of which the amplitude of oscillations due to the seastates die out
52 N. Senroy
2.4.3 Analyzing Slow Coherency [9]
The aim of this section is to confirm some aspects of the linearized analysis of a
power system using generator swing curves obtained using nonlinear time-
domain simulations. Coherency of generators is used as an aggregation criter-
ion to simplify a complex power system network. It has been observed in a
multimachine system that in the aftermath of a sudden disturbance, some
machines tend to ‘swing together.’ This means that there is relatively little
difference between their swing curves. Such coherent machines may be grouped
together and represented as equivalent machines in further computations. In a
practical dynamic system, perfectly coherent machines are rarely encountered.
Hence, near-coherent machines are grouped together by examining the rows of
an eigenbasis matrix. This eigenbasis matrix is formed by combining the eigen-
vectors corresponding to some modes of interest. Coherency of these machines
is always defined with respect to a set of modes of the system. Coherent states
are insensitive to the system perturbation because only those modes are selected
that are excited by the initial perturbation and subsequently the states with the
same content of perturbed modes are identified.In a power system with n natural modes of oscillation, there are some modes
that are of high frequency and well damped. These modes do not have much
impact on the slow transients in the system. Hence, the coherency between the
machines is always defined for r modes of the system which are usually ‘slow’
and have a tangible impact on the system behavior. This kind of coherency is
referred to as slow coherency. In slow coherency theory, generators are coherent
with each other over large timescales. Coherent groups may exhibit very fast
oscillations within themselves and a disturbance will propagate within a group
very rapidly. Therefore, over very short timescales, each of these tightly bonded
groups may be studied in isolation from the rest of the system. However, over
larger timescales, the weak interactions between various tightly bonded groups
become significant. Over large timescales, these tightly bonded slow coherent
machines are represented as equivalent models that are insensitive to the level of
detail used in modeling the generators.A six-machine system was linearized around its base case operating point [9],
and the natural modes were calculated to be 1.29, 0.99, 0.72, 0.49, 0.38, and
0.0 Hz. Further eigenanalysis revealed that generator 1 and 6 were coherent for
themodes – 0.0, 0.38, and 0.49Hz. Because this coherency was established using
the linearized model of the system, the modified Hilbert–Huang method was
employed for confirmation using the nonlinear model of the system. Generator
swing curves were obtained for a line tripping without a fault. The 10 s long
swing curves are shown in Fig. 2.11. While a rough idea of the coherency
between the machines can be had from visual inspection of the swing curves,
a detailed analysis is readily provided by the masking signal-based EMD.Masking signals were used with EMD to extract monocomponent IMFs.
For both swing curves, the masking signal frequencies used were 1.6 and 1.1 Hz.
2 Enhanced Hilbert–Huang Technique for Power System Oscillations 53
Figure 2.11 shows the three important IMFs extracted from the swing curves.
The highest frequency IMFs, i.e., IMF1 for both the machines are anti-phase toeach other. The lower frequency IMFs are in phase with each other. In termsof magnitude, the third IMF was the most significant for both the generator
swing curves. Hilbert transform was applied to all the three sets of IMFs, andthe instantaneous frequency obtained is shown in Fig. 2.12. It is clear fromFig. 2.12, that the instantaneous frequency of the corresponding IMFs,
extracted from the two generators’ swing curves, show good agreement.The selection of IMFs for comparing the instantaneous phase difference must
be done on the basis of their instantaneous frequencies which must match.Hence, an instantaneous phase difference is calculated between those IMFswhose instantaneous frequencies match.
It is also interesting to relate the instantaneous frequencies obtained for the
IMFs with the natural frequencies of the system as revealed by the eigenvalueanalysis. The coherency between the machines was computed for the slow
0 2 4 6 8 10
0
5
10
15
time (s)
swing curves
0 2 4 6 8 10–0.5
0
0.5
time (s)
IMF1
0 2 4 6 8 10–1
0
1
time (s)
IMF2
0 2 4 6 8 10
–1
0
1
time(s)
IMF3
Fig. 2.11 Analysis of slow coherency between two generators using the masking signal-basedEMD. The top plot shows the original swing curves obtained for a line tripping. The lowerthree plots show the extracted IMFs from the two swing curves in decreasing order offrequency. The generators are coherent for lower frequencies, and out of phase for the highestfrequency
54 N. Senroy
modes 0.0, 0.39, and 0.49Hz. Furthermore, it was found that generators 1 and 6participate primarily in the 1.29 Hz mode (participation factors: 0.305 and0.191, respectively), and the 0.49 Hz mode (participation factors: 0.115 and0.199, respectively), with minor participation in the 0.72 Hz mode (participa-tion factors: 0.006 and 0.026, respectively). Thus, slow coherency is verified byrelating the instantaneous frequencies of the coherent IMFs and the naturalmodes of the system.
2.4.4 Wide-Area Measurement Signals [9]
This section presents the application of the modified EMD to analyze actualwide-area measurement data. An Internet-based WAMS is presented in [10]. Itconsists of frequency deviation recorders (FDRs) monitoring system frequencyat strategic locations in the USA. The FDRs measure frequency at the distribu-tion level at a sampling rate of 10 Hz, by plugging into 110 V sockets. Themeasurement data from the FDR is consequently noisy, and preprocessing isrequired before the data is ready to be analyzed. The frequency measurementsare time stamped using Global Positioning System synchronizing signals.
0 1 2 3 4 5 6 7 8 100
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
time [s]
freq
uenc
y [H
z]
Fig. 2.12 Instantaneous frequency of the three set of IMFs extracted from the swing curvesshown in Fig. 2.11. The dashed line corresponds to the IMFs extracted from the swing curvesof one machine while the solid lines correspond to the IMFs extracted from the other machineswing curve
2 Enhanced Hilbert–Huang Technique for Power System Oscillations 55
The Pacific DC Intertie is a �500 V, 3,100 MW HVDC line connecting the
Cellilo substation in Oregon to the Sylmar substation near Los Angeles, south-
ern California. This 1,354-km-long line is critical in improving the reliability of
the electricity grid located in the western United States, by supplementing the
thermal and nuclear generation in the southwest with hydropower from the
northwest. OnAugust 25, 2005 at 3:53 PM (PDT, Pacific Daylight Time) due to
the loss of a converter, the HVDC line was taken out of service, resulting in the
disruption of 1,750 MW. Figure 2.13 shows the frequency deviations resulting
out of the loss of the HVDC tie line. They were recorded by FDR14, FDR16,
and FDR21 located at Arizona State University (ASU), Tempe AZ, USA, Los
Angeles, CA, USA, and Washington State University (WSU), WA, USA,
respectively. It is clear from the oscillations that the FDR21 signal located in
the northwest is out of phase with the other two signals both located in the
southwest. However, due to the fact that all the signals aremulticomponent, it is
difficult to arrive at a reliable phase comparison without the application of
EMD.The standard EMDof [1] was applied to extract the overall trend from all the
three signals. This manifests as the last IMF extracted. Figure 2.14 shows the
extracted trends from the three WAMS signals. The system frequency was
initially at 59.8 Hz, when the HVDC line tripped. Following the tripping, the
frequency dipped to less than 59.75 Hz, after which it settles to 59.8 Hz. The
oscillations extracted from the frequency signals are shown in Fig. 2.15.
03:52:58 PM 03:53:08 PM 03:53:18 PM59.7
59.75
59.8
59.85
59.9
59.95
Pacific Daylight Time
freq
uenc
y (H
z)
FDR14
FDR16
FDR21
Fig. 2.13 Frequency oscillations recorded by FDR14 at ASU, Tempe, AZ; FDR16 at LosAngeles, CA; FDR21 at WSU, Washington, WA. The event is the tripping of the Pacific DCIntertie on August 25, 2005
56 N. Senroy
03:52:58 PM 03:53:08 PM 03:53:18 PM59.7
59.75
59.8
59.85
59.9
59.95
Pacific Daylight Time
freq
uenc
y (H
z)
FDR16
FDR14
FDR21
Fig. 2.14 Trend extracted from the WAMS signals shown in Fig. 2.13. The signals wererecorded by FDR14, FDR16, and FDR21 located in Arizona State University, Tempe, AZ,Los Angeles, CA, and Washington State University, WA, respectively
03:52:58 PM 03:53:08 PM 03:53:18 PM–0.08
–0.06
–0.04
–0.02
0
0.02
0.04
0.06
Pacific Daylight Time
freq
uenc
y (H
z)
FDR14
FDR16
FDR21
Fig. 2.15 Oscillations in frequency extracted using the standard EMD from the WAMSsignals shown in Fig. 2.13
2 Enhanced Hilbert–Huang Technique for Power System Oscillations 57
The birth of frequency oscillations are clearly observable at 03:52:58 PM,
arising out of the HVDC line tripping. These oscillations die out within 18 s,
when the frequency settles to around 59.8 Hz.The FFT spectrum of the frequency variations, reveals the multicomponent
character of the oscillations. Thus, the straightforward application of Hilbert
transform would not yield meaningful phase information. Further examination
reveals that there are possible low-amplitude frequency components in all the
three signals around 0.6–0.7 Hz. Straightforward application of the EMDwould
include these components along with the dominant 0.31 Hz component in the
first IMF.Hence, a masking signal is required to extract the weak components of
frequency ranging from 0.6 to 0.7 Hz. Amasking signal of frequency 0.88Hz and
appropriate amplitude was used for both FDR14 and FDR16 signals. Similarly,
a 0.99 Hz masking signal was used for the FDR21 signal.The dominant IMFs extracted from the three WAMS signals are shown in
Fig. 2.16(a). Hilbert transform on the IMFs revealed their instantaneous
phases. These were further used to compute the instantaneous phase differences
as shown in Fig. 2.16(b). It can be observed from Fig. 2.16(b) that the instanta-
neous phase difference between the dominant IMFs of FDR14 and FDR16
hovers around 08. Similarly, the phase difference between the dominant IMFs
of FDR21 and FDR14/FDR16 hovers around 1808. This indicates that FDR14
and FDR16 are in phase and coherent, while FDR21 is out of phase with both
FDR14 and FDR16.
03:52:58 PM 03:53:08 PM 03:53:18 PM–90
0
90
180
250
Inst
anta
neou
sph
ase
diffe
renc
e(d
eg)
(b)
FDR14 and FDR16FDR14 and FDR21FDR16 and FDR21
03:52:58 PM 03:53:08 PM 03:53:18 PM–0.05
0
0.05
Pacific Daylight Time
freq
uenc
y de
viat
ion
(Hz)
(a)FDR21
FDR14 FDR16
Fig. 2.16 Instantaneous phase difference of the dominant IMFs extracted from the WAMSsignals of Fig. 2.3: (a) dominant IMFs of FDR14, FDR16, and FDR21 and (b) instantaneousphase differences computed using Hilbert transform. Masking signal frequencies used withEMD are 0.88 Hz for FDR14 and FDR16 and 0.99 Hz for FDR21
58 N. Senroy
2.5 Discussion
In this chapter the Hilbert–Huang method has been presented along with
improvements to enhance its application to signals representing power sys-
tem dynamic behavior. All the modifications to the original Hilbert–Huang
technique presented have been aimed at improving the capability of the
EMD to discriminate between closely spaced as well as weak amplitude
frequency components in a time-varying distorted signal. Case studies of the
application of the technique to various kinds of signals have also been
presented. While the case studies clearly demonstrate the efficacy of the
suggested improvements, it is worthwhile to note the limitations of the
modified Hilbert–Huang technique.The technique of constructing masking signals depends heavily on the
approximate spectral information obtained using the FFT. When applied to
nonstationary and nonlinear waveforms, the FFT results in a stationary or
periodic interpretation. Caution must be exercised while deriving masking
signals from the FFT spectrum, as heavily damped signals result in spurious
peaks in the FFT spectrum.The frequency chosen to heterodyne the distorted signal must be such that it
shifts the frequency spectrum to a lower frequency scale, such that adjacent
frequency components lie in different octaves. While maximum value of the
heterodyne frequency is well defined, the minimum value is harder to define. A
heterodyne frequency selected very close to the highest frequency component,
results in a waveform wherein the highest frequency component is translated to
a very low frequency. Accordingly, the total length of the observed signal must
be large enough to accommodate the dynamics of this low-frequency compo-
nent. If the length of the signal is not adequate, the heterodyne frequency must
be selected as a larger value. Hence, the minimum value of the heterodyne
frequency depends on the dynamics of the highest frequency component present
in the distorted signal.A key component of the EMD technique is the interpolation between
extrema points that is achieved using spline fitting. Spline fitting is associated
with serious errors at the end points of the data, which affect lower frequency
IMFs more. In fact, spline fitting-induced errors accumulate as more and more
IMFs are extracted. Thus, the instantaneous frequency and amplitudes for
these IMFs show considerable error at the ends of the window, as can be
observed in Fig. 2.12. Additionally, one problem with the Hilbert transform is
that at very small magnitudes, it is difficult to track the instantaneous fre-
quency. One possible direction of future research is to focus on the minimum
threshold amplitude an IMF must have, to facilitate satisfactory frequency
tracking.The Hilbert–Huang transform is noncausal by nature; however, it may be
possible to implement it as an online application. Several aspects of the techni-
que must be resolved before that. One critical aspect is that the signal extrema
2 Enhanced Hilbert–Huang Technique for Power System Oscillations 59
points must be clearly and unambiguously defined for accurate sifting of the
IMFs. While this indicates that the signal must be significantly oversampled, it
also makes the algorithm sensitive to sensor-induced jitters. These jitters appear
as high-frequency components riding over the original signal, and appropriate
threshold criteria are needed to recognize such phenomena. Further, the dura-
tion of observation of the signal must be large enough to include adequate
number of extrema points to facilitate the spline algorithm. In the context of
online measurement data analysis, a sliding window of appropriate length may
be slid along the incoming data.
2.6 Conclusion
In this chapter, the time–frequency and time–magnitude localization capabil-
ities of the Hilbert–Huang transform have been significantly improved by the
suggested modifications. These modifications have been proposed keeping in
mind the specific nature of the signals to be analyzed, namely those that
represent power system dynamic behavior. The wide variety of case studies
demonstrate that it is possible to extract instantaneous signal parameters with
minimal knowledge of the system from which the signal is obtained. For
instance, the analysis of wide-area measurements implies that it is possible to
extract accurate coherency information using distributed measurements. Thus,
dynamic reduction of the system can be achieved empirically, with limited
knowledge of system parameters leading to model validation. Further coher-
ency can be monitored in real time, which will be useful in designing
special protection schemes relying on generator aggregation. The modified
Hilbert–Huang technique is also a reliable post-processing tool to analyze
sequence of events leading to a better understanding of evolving system
dynamics as well as to validate existing system models. Similarly, the adaptive
transfer function estimation technique described in this chapter may lead to
improvements in online health and condition monitoring of critical power
system equipment. Further research should focus on adapting the modified
Hilbert–Huang technique to real-time applications involving online data pro-
cessing, as well as to improve its capability for accurate instantaneous frequency
and magnitude estimation.
Acknowledgments The author acknowledges the contribution of Siddharth Suryanarayananof Colorado School of Mines, Golden, Colorado, USA in the development of the algorithmspresented in this chapter. The following other people are also acknowledged for their technicalcontributions: PauloM.Ribeiro of Calvin College,Michigan, USA;Michael ‘Mischa’ Steurerof Center for Advanced Power Systems, Florida State University, Tallahassee, Florida, USA;Stephen Woodruff of NASA Dryden Flight Research Center, California, USA; and ArturoMessina of CINVESTAV, Guadalajara, Mexico. Financial support from the Office of NavalResearch, USA, the Department of Energy, USA and the Industrial Research and Develop-ment Unit, IIT-Delhi, India, is also gratefully acknowledged.
60 N. Senroy
References
1. Huang N E, et al., ‘‘The empirical mode decomposition and the Hilbert spectrum fornonlinear and nonstationary time series analysis,’’ Proc. R. Soc. Lond. A., vol. 454, 1998,pp. 903–995
2. Gabor D., ‘‘Theory of communication,’’ IEE J. Comm. Eng., vol. 93, 1946, pp. 429–457.3. Senroy N, Suryanarayanan S, Ribeiro P F, ‘‘An improved Hilbert-Huang method for
analysis of time-varying waveforms in power quality,’’ IEEE Trans. Power Sys., vol. 22,No. 4, Nov. 2007, pp. 1843–1850.
4. RequichaAG, ‘‘The zeros of entire functions: theory and engineering applications,’’Proc.IEEE, vol. 68, no. 3, Mar. 1980, pp. 308–328.
5. Deering R, Kaiser J F, ‘‘The use of masking signal to improve empirical mode decom-position,’’ Proc. IEEE Int. Conf. Acoustics, Speech Signal Processing (ICASSP ’05), vol.454, 2005, pp. 485–488.
6. Senroy N, Suryanarayanan S, ‘‘Two techniques to enhance empirical mode decomposi-tion for power quality applications,’’ IEEE PES General Meeting, June 2007, pp. 1–6.
7. Messina A R, Vittal V, ‘‘Nonlinear, Non-stationary analysis of interarea oscillations viaHilbert spectral analysis,’’ IEEETrans. Power Sys., vol. 21, No. 3, Aug. 2006, pp. 1234–1241.
8. Senroy N, Suryanarayanan S, Steurer M, ‘‘Adaptive transfer function estimation of anotional high-temperature superconducting propulsion motor,’’ Accepted for publica-tion, IEEE Trans. Ind. Appl., Feb. 2008.
9. Senroy N, ‘‘Generator coherency using the Hilbert-Huang transform,’’ IEEE Trans.Power Sys., vol. 23, No. 4, Nov. 2008, pp. 1701–1708.
10. Wang J K, et al., ‘‘Analysis of system oscillations using wide-area measurements,’’ IEEEPES General Meeting, June 2006, pp. 1–6.
2 Enhanced Hilbert–Huang Technique for Power System Oscillations 61
Chapter 3
Variants of Hilbert–Huang Transform
with Applications to Power Systems’
Oscillatory Dynamics
Dina Shona Laila, Arturo Roman Messina, and Bikash Chandra Pal
Abstract Power system dynamic processes may exhibit highly complex spatial
and temporal dynamics and take place over a great range of timescales. When
frequency analysis requires the separation of a signal into its essential compo-
nents, resolution becomes an important issue. The Hilbert–Huang transform
(HHT) introduced by Huang is a powerful data-driven, adaptive technique for
analyzing data from nonlinear and nonstationary processes. The core to this
development is the empirical mode decomposition (EMD) that separates a
signal into a series of amplitude- and frequency-modulated signal components
from which temporal modal properties can be derived. Previous analytical
works have shown that several problems may prevent the effective use of
EMD on various types of signals especially those exhibiting closely spaced
frequency components and mode mixing. The method allows a precise char-
acterization of temporal modal frequency and damping behavior and enables a
better interpretation of nonlinear and nonstationary phenomena in physical
terms.This chapter investigates several extension to the HHT. A critical review of
existing approaches to HHT is first presented. Then, a refined masking signal
EMD method is introduced that overcomes some of the limitations of the
existing approaches to isolate and extract modal components. Techniques to
compute a local Hilbert transformation are discussed and a number of numer-
ical issues are discussed.As case studies, the applications of the various EDM algorithms in power
system’ signal analysis are presented. The focus of the case studies is to
accurately characterize composite system oscillation in a wide-area power
network.
D.S. Laila (*)Department of Electrical and Electronic Engineering, Imperial College London,Exhibition Road, London SW7 2AZ, UKe-mail: [email protected]
A.R. Messina, Inter-area Oscillations in Power Systems,Power Electronics and Power Systems, DOI 10.1007/978-0-387-89530-7_3,� Springer ScienceþBusiness Media, LLC 2009
63
3.1 Introduction
Slow electromechanical oscillations are inherent in stressed interconnected
power system operation [1]. In many occasions [2], they appear to be precursor
to severe system upset. Obviously, it is important to detect and mitigate their
influence for secure operation.The existing technology is to run series of dynamic security analysis which
essentially utilizes eigenanalysis on the linearized model of the system [3, 4].
However, the oscillatory behavior of power system is nonlinear, which causes
the eigenanalysis less accurate in estimating the damping and frequency asso-
ciated with the oscillations apart from the uncertainty in the model. Techniques
using measured data such as Prony analysis, spectral analysis, and MIMO
(multiple-input multiple-output) state–space identification were applied in [5, 6,
7, 8]. These techniques, however, rely on the linear and periodicity assumption of
the system response. In [9, 10] Fourier analysis has been used for off-line studies
of power system dynamics. However, Fourier spectral analysis requires the data
to be strictly periodic or stationary. The Fourier spectrum defines uniform
harmonic components globally and therefore needs many additional harmonic
components to simulate nonstationary data. As a result, it spreads the energy
over a wide frequency range making little physical sense. Although linear and
stationary approximation may not be as crude, the information retrieved from
this approach is often insufficient. Recently, nonlinear and nonstationary analy-
sis have been used to characterize the time–frequency attributes of system
response [11, 12], applying Hilbert–Huang transform (HHT).The HHT is an empirically based signal analysis method developed by
Huang et al. [13], which is a potentially powerful tool to analyze nonlinear and
nonstationary signals. This method has been applied to various fields such as
medical [14], geology and geodesic [15], and power systems [16]. TheHHTmethod
consists of twomain processes, the empiricalmode decomposition (EMD) and the
Hilbert transform (HT). The EMD process aims at decomposing the signal into
its frequency components called intrinsic mode functions (IMFs). The main
procedure in EMD is called sifting. The EMD process targets each IMF to be a
monofrequency signal, namely signal with single frequency component. For the
EMD to give a good analysis, resolution, namely the ability of the method to
separate each frequency component during the sifting process is very important.
It is further established that the EMD behaves as a dyadic filter bank [17]. It
decomposes white noise into IMFs whose frequency spectrum comprises an
octave. The mean frequencies of extracted IMFs show period doubling phenom-
enon. Although the original sifting procedure can perform this task quite well for
certain frequency range and composition of a signal, unfortunately, it often fails to
work for signalswith low frequency, lowmagnitude, or narrow frequency, i.e., two
consecutive frequency components within an octave. This is seen as a potential
limitation of the standard EMD when analyzing power system signals which are
dominated by some inter-area modes with frequency close to each other.
64 D.S. Laila et al.
There has been considerable recent research interest to improve the EMD
process, in particular, the sifting procedure to overcome the difficult situations
mentioned earlier. In [18], a masking technique is first introduced, and in [19,
20], the masking technique is improved, giving a constructive algorithm on how
to choose the masking signals. Other works in this directions are, for instance
[21, 22]. As the limitation of the EMD process somehow stems from the loose
definition of an IMF, another approach to improve the method is by refining
the definition of this function [23]. In this work, it is suggested to impose
additional condition to a function to be called an IMF.EMD with masking algorithm from [19] are particularly interesting.
Although it has been shown that the masking technique improves the sifting
process in the power quality analysis as shown in [19], the result is not generic.
The algorithm has worked well for signals in high-frequency range, but it has
not been able to accurately capture the close frequencies associated with inter-
area mode oscillations. Moreover, the algorithm contains some limitations as it
depends on the use of Fourier analysis.In this chapter, several extensions are to the EMDwith masking is presented,
based on the use of masking signal in the sifting procedure. A systematic way to
construct the masks for signals with various frequency ranges and compositions
is also provided, extending both the results from [18, 19]. A local HT based on
convolution is also presented that circumvents some of the limitations of dis-
crete Fourier transform (DFT)-based approaches.The rest of the chapter is as follows. In Section 3.2, we present some
preliminaries of EMD and HT. In Section 3.3, we present several algorithms
of refinement to standard EMD and HT algorithm and demonstrates its effec-
tiveness on a synthetic signal. Some applications to synthetic signals and also
real measurement signals are provided in Section 3.4 and the chapter concludes
by some remarks in Section 3.5.
3.2 Preliminaries
In an effort to make this chapter reasonably self-contained, some preliminaries
that are needed to support the techniques presented in this chapter are presented.
The standard algorithm of the HHT and its components, the EMD technique,
and HT are also briefly reviewed. Our development follows the development of
Huang [13], and we refer the readers to this source for more details.
3.2.1 Fourier Analysis
Fourier analysis is one of the most important tools in signal spectrum analysis.
Its frequencial description can be the basis of a better comprehension of the
3 Variants of Hilbert–Huang Transform with Applications 65
underlying phenomena as it complements the temporal description [24]. Given asignal xðtÞ, the Fourier transform is defined as follows:
XðfÞ ¼Z þ1
�1xðtÞe�i2pft dt : (3:1)
The computation of one frequency value XðfÞ requires knowledge of the com-plete history of the signal ranging from �1 to þ1. On the other hand, theinverse Fourier transform is defined as
xðtÞ ¼Z þ1
�1XðfÞei2pft df : (3:2)
Any value xðtÞ at one instant t can be regarded as an infinite superposition ofcomplex exponentials or everlasting and completely nonfocal waves.
However, there is the limitation of this tool. From the definition of Fouriertransform that involves the integration over the range from �1 to 1, it isobvious that the length of the signal matters. It may give quite an accuratespectrum for long or steady-state signal, but it may distort the physical reality ofshort or transient signals.
3.2.2 The Empirical Mode Decomposition Method
The EMDmethod provides an analytical basis for the decomposition of a signalxðtÞ into a set of basis functions, called IMFs. An IMF is defined as a signal thatsatisfies the following criteria.
1. Over the entire time series the number of extrema and the number of zero-crossings differ by, at most, one, i.e., an essentially oscillatory process.
2. At any point the mean value of the envelope defined by the local maxima andthe envelope defined by the local minima is zero.
The basic EMD method adopted to extract the IMFs essentially consists of athree-step procedure called sifting [12]. The goal is to subtract away the large-scalefeatures of the signal repeatedly until only the fine-scale features remain. A signalxðtÞ is thus divided into the fine-scale details cðtÞ and the residue rðtÞ, hencexðtÞ ¼ cðtÞ þ rðtÞ. The components contained in the fine-scale details are theIMFs.
The standard EMD process can be summarized as follows.
S1. Given the original signal xðtÞ; set roðtÞ ¼ xðtÞ, j ¼ 1.S2. Extract the jth IMF using the sifting procedure:
66 D.S. Laila et al.
a. Set i ¼ 1 and hi�1ðtÞ ¼ rj�1ðtÞ.b. Identify the successive local minima and the local maxima for hi�1ðtÞ.
The time spacing between successive maxima is defined to be the time-scale of these successive maxima.
c. Interpolate the local minima and the local maxima with a cubic spline toform an upper emaxi�1ðtÞ and lower emini�1ðtÞ envelope for the whole dataspan.
d. Compute the instantaneous mean of the envelopes,
mi�1ðtÞ ¼emini�1ðtÞ þ emaxi�1ðtÞ
2; (3:3)
and determine a new estimate hiðtÞ ¼ hi�1ðtÞ �mi�1ðtÞ, such thatemini�1ðtÞ � hiðtÞ � emaxi�1ðtÞ for all t. Set i ¼ iþ 1.
e. Repeat steps S2b–S2d until hiðtÞ satisfies a set of predeterminedstopping criteria (follows the criteria 1 and 2 of an IMF). Then setcjðtÞ ¼ hiðtÞ.
S3. Obtain an improved residue rjðtÞ ¼ rj�1ðtÞ � cjðtÞ. Set j ¼ jþ 1. Repeat stepS2 until the number of extrema in rjðtÞ is less than 2.
This approach allows elimination of low-amplitude riding waves in the timeseries and eliminates asymmetries with respect to the local mean, i.e., it makesthe wave profile more symmetric. At the end of this process, the EMD yields thefollowing decomposition of the signal xðtÞ,
xðtÞ ¼Xn
j¼1cjðtÞ þ rnðtÞ ¼
Xp
k¼1ckðtÞ þ
Xn
l¼pþ1clðtÞ þ rnðtÞ; (3:4)
where ckðtÞ; k ¼ 1; . . . ; p contain the physical behavior of interest and theremaining terms clðtÞ; l ¼ pþ 1; . . . ; n and rnðtÞ contain less relevant, nonsinu-soidal characteristics. We emphasize that the refined masking signal methodenables a superior analysis of dynamic behavior than Fourier-based techniquesand provides essential information that may be used to determine modalbehavior, i.e., instantaneous damping and phase as discussed below.
3.2.3 Hilbert Transform
Given a real signal xðtÞ. Its complex representation is
zðtÞ ¼ xðtÞ þ ixHðtÞ ; (3:5)
where xHðtÞ is the HT of xðtÞ, given by
xHðtÞ ¼1
pP
Z þ1
�1
xðsÞt� s
ds ; (3:6)
3 Variants of Hilbert–Huang Transform with Applications 67
with P the Cauchy principal value of the integral. Equation (3.5) can be rewrittenin an exponential form as
zðtÞ ¼ AðtÞei ðtÞ ; (3:7)
where AðtÞ ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffixðtÞ2 þ xHðtÞ2
qand ðtÞ ¼ arctan xHðtÞ=xðtÞ.
The time derivative of (3.7) is
_zðtÞ ¼ AðtÞei ðtÞði!ðtÞÞ þ ei ðtÞ _AðtÞ ; (3:8)
where !ðtÞ is the instantaneous angular frequency, which, by definition, is thetime derivative of the instantaneous angle
!ðtÞ ¼ _ ðtÞ ¼ d
dtarctan
xHðtÞxðtÞ : (3:9)
Hence, the instantaneous frequency can be defined as
fðtÞ ¼ !ðtÞ2p
; (3:10)
and using (3.7) and (3.8), it can be computed as
fðtÞ ¼ 1
2pIm
_zðtÞzðtÞ
� �¼ 1
2pxðtÞ _xHðtÞ � xHðtÞ _xðtÞ
x2ðtÞ þ x2HðtÞ: (3:11)
Remark 3.1 In HHT technique, HT is applied to each IMF to compute itsinstantaneous frequency, as well as instantaneous magnitude. As instantaneousfrequency is best defined for monofrequency signals, i.e., signals that containonly one (dominant) frequency, it makes sense to expect each IMF to bemonofrequency. However, as pointed out earlier, the IMFs may contain amixture of frequencies (frequency modulation) and are difficult to interpret interms of conventional modal analysis. This has motivated the need for demo-dulation techniques that extract from each IMF the dominant interactingfrequencies.
3.2.4 Instantaneous Damping
3.2.4.1 Computation Based on the Exponential Decay
The knowledge about the instantaneous magnitude and instantaneous fre-quency of a signal allows us to further compute the instantaneous damping ofthe signal. Damping characterization is another useful alternative to the
68 D.S. Laila et al.
analysis of local behavior of the oscillation. Consider the signal (3.7). We canrewrite the signal as [25]
zðtÞ ¼ AðtÞei ðtÞ ¼ �ðtÞe��ðtÞþi ðtÞ : (3:12)
Then the time-dependent decay function can be modeled as
�ðtÞ ¼ �Z t
0
�ðtÞ dt : (3:13)
Moreover, using (3.7) and (3.8), we obtain
_zðtÞzðtÞ ¼ ��ðtÞ þ
_�ðtÞ�ðtÞ
� �þ i!ðtÞ
� �: (3:14)
Noting that
Re_zðtÞzðtÞ
� �¼
_AðtÞAðtÞ ; (3:15)
we have the instantaneous damping coefficient � as
�ðtÞ ¼ � d�ðtÞdt¼ �
_AðtÞAðtÞ �
_�ðtÞ�ðtÞ
� �: (3:16)
Moreover, if �ðtÞ is constant, which means the signal is purely exponential,_�ðtÞ ¼ 0, and hence (3.16) is simplified into
�ðtÞ ¼ � d�ðtÞdt¼ �
_AðtÞAðtÞ : (3:17)
3.2.4.2 Computation Based on the Second-Order System Approach
Another way of computing the instantaneous damping is through the second-order system emulation approach [26, 27]. Given a signal
xðtÞ ¼ AðtÞ cosð ðtÞÞ : (3:18)
The complex representation of the signal is
zðtÞ ¼ xðtÞ þ jxHðtÞ (3:19)
with xHðtÞ the HT of xðtÞ that follows the form
3 Variants of Hilbert–Huang Transform with Applications 69
xHðtÞ ¼ AðtÞ sinð ðtÞÞ : (3:20)
A general form of an initial differential equation of motion in the analyticalsignal form can be written in a frequency-dependent (viscous) damping form as
€zþ 2�ðAÞ _zþ !2oðAÞz ¼ 0 ; (3:21)
or also in a frequency-independent (structural) damping form as
€zþ !2oðAÞ 1þ j
�ðAÞp
� �¼ 0 ; (3:22)
with z the system solution in the analytical signal form, � the instantaneousdamping coefficient, and !o the instantaneous undamped natural frequency,� ¼ 2p� the logarithmic decrement and � the damping ratio.
Using the analytical signal form (3.19) together with its two first derivatives
_z ¼ zðtÞ_AðtÞAðtÞ þ j!ðtÞ� �
€z ¼ zðtÞ€AðtÞAðtÞ � !
2ðtÞ þ j 2!ðtÞ_AðtÞAðtÞ þ _!ðtÞ
� �� � (3:23)
to solve the dynamic equation (3.21), we get the equation for free vibrationanalysis
z€A
A� !2 þ !2
o þ 2�_AðtÞAðtÞ þ j 2
_AðtÞAðtÞ!þ _!þ 2�!
� �� �¼ 0 (3:24)
where A and ! ¼ �1 are, respectively, the envelope and instantaneous fre-quency of the vibratory system solution. Solving two equations for real andimaginary parts of (3.24), we can write the expression for the instantaneousmodal parameters as functions of the first and the second derivatives of thesignal envelope and the instantaneous frequency, i.e.,
!2oðtÞ ¼ !2 �
€A
Aþ 2
_A2
A2þ
_A _!
A!(3:25)
�ðtÞ ¼ �_A
A� _!
2!; (3:26)
where !oðtÞ is the instantaneous undamped natural frequency and �ðtÞ is theinstantaneous damping coefficient of the system. The parameters ! and A arethe instantaneous frequency and the envelope (amplitude) of the oscillation,respectively. Moreover, if the signal is monofrequency, which is the case for
70 D.S. Laila et al.
IMFs when the EMD works well to decompose a composite signal, we have_!ðtÞ ¼ 0, and hence (3.26) simplifies into
�ðtÞ ¼ �_AðtÞAðtÞ : (3:27)
The instantaneous damping ratio �ðtÞ is computed following the relation
�ðtÞ ¼ �ðtÞ2!oðtÞ
: (3:28)
Remark 3.2 Observing the formulae given by Eqs. (3.16) and (3.26) it can beseen that the first approach takes into account the changing of the ‘‘unexpo-nential’’ instantaneous magnitude to compute the instantaneous damping,whereas the second approach takes into account the frequency variation ofthe signal. When the signal is purely exponential and monofrequency, the twoformulae yield an equivalent instantaneous computation as given by Eqs. (3.17)and (3.27).
3.2.5 Completeness, Orthogonality, and Orthogonality Index
The decomposition through EMD results in IMFs that possess completenessand orthogonality properties. The completeness is given by virtue of the decom-position as observed from (3.4), the sum of all decomposed elements yields theoriginal composite signal. In Section 3.4, we present some numerical examplesto demonstrate the completeness property of the decomposition (see also [13Section 6]). The orthogonality property is satisfied in all practical sense, but isnot guaranteed theoretically [13]. By virtue of the decomposition, the elementshould all be locally orthogonal to each other, because each element is obtainedfrom the difference between the composite signal and its local mean through themaximal and minimal envelopes, i.e.,
ðxðtÞ � xðtÞÞ � xðtÞ ¼ 0 : (3:29)
Ideally, if the mean is computed analytically, (3.29) is true and the IMFs areorthogonal. However, as the mean is computed through geometrical identifica-tion of the envelopes, hence it is not the exact mean. Moreover, each successiveIMF component is only part of the signal constituting xðtÞ. Because of theseapproximation, some errors, however small, cannot be avoided.
While we can assume the IMFs to be ‘‘approximately’’ orthogonal, the moreaccurate orthogonality of these EMD components should be checked a poster-iori numerically as follows. Rewrite (3.4) as
3 Variants of Hilbert–Huang Transform with Applications 71
xðtÞ ¼Xnþ1
j¼1cjðtÞ; (3:30)
where the residue is regarded as cnþ1. The orthogonality of the IMFs is thenchecked based on a parameter called the index of orthogonality. Squaring thesignal (3.30), we obtain
xðtÞ ¼Xnþ1
j¼1c2j ðtÞ þ 2
Xnþ1
j¼1
Xnþ1
k¼1cjðtÞckðtÞ: (3:31)
If the decomposition is orthogonal, the cross terms on the right-hand side of(3.31) are zero. Hence, the index of orthogonality is defined as
IO ¼XT
t¼1
Xnþ1
j¼1
Xnþ1
k¼1cjðtÞckðtÞ=x2ðtÞ
!; (3:32)
where T40 is the number of the samples. The closer the value of IO to zero, thecloser the decomposition to orthogonality.
Theoretically, the EMD guarantees orthogonality only on the strength ofEq (3.29), and orthogonality also depends on the decomposition algorithm.Moreover, orthogonality is only important for linear decomposition system,and it would not make any physical sense for nonlinear decomposition. Hence,orthogonality of EMD for nonlinear signal is not guaranteed, but neverthelessthe orthogonality index is usually quite small which means the nonlineardecomposition is still close to orthogonal in some sense. More discussion andcomputational example for the orthogonality of EMD are presented in [13].
3.3 Masking Techniques to Improve Empirical
Mode Decomposition
This section discusses extensions to conventional EMD analysis to study oscil-latory dynamics. First, a synthetic example is introduced to examine conditionsunder which the conventional masking technique may fail. Then, various algo-rithms to refine the existing HHT are proposed and tested.
3.3.1 The Standard EMD Method and Its Limitation
Consider a two-component signal, xðtÞ,
xðtÞ ¼ 8 sinð1:6ptÞ þ 20 sinðptÞ : (3:33)
The time evolution of this signal is shown in Fig. 3.1. The clear feature of signal(3.33) is that it consists of low-frequency components and the magnitude of
72 D.S. Laila et al.
the higher frequency component is significantly lower than that of the lower
frequency component.The standard EMD [28] that has been described in Section 3.2.2 is applied to
the testing signal (3.33). Figure 3.2 shows the decomposition result, with the
dashed background of the first two plots show the individual components of the
composite signal, which IMF1 and IMF2 should imitate. However it is seen that
0 5 10 15 20 25 30−30
0
30
Time (s)
Fig. 3.1 The synthetic signal (3.33) with two components
0 5 10 15 20 25 30−2
0
2
0 5 10 15 20 25 30–1
0
1
Res
idue
Time (s)
0 5 10 15 20 25 30–30
0
30
IMF 1
IMF 2
IMF 3
0 5 10 15 20 25 30–20
0
20
Fig. 3.2 IMFs of the signal (3.33) obtained using the standard EMD (components of thecomposite signal are plotted as dashed line background)
3 Variants of Hilbert–Huang Transform with Applications 73
the two IMFs do not match the background. IMF1 is not a monofrequencysignal, but instead it exhibits mode mixing, making little sense to expect usefulphysical interpretation through the application ofHilbert analysis. The error onIMF1 is transmitted to the next IMFs. This shows that the standard EMD doesnot perform the decomposition well to this signal.
This drawback of standard EMD has been recognized in many cases. In anattempt to improve the performance and effectiveness of EMD, in [18,29] amasking technique is introduced. The technique aims at solving the problem ofmode mixing due to intermittency that occurs during transient events and theproblem of ambiguity that happens when two or more frequencies are not wellseparated. This technique has helped solving parts of the problems. More in-depth discussion about the background and technicalities of this technique ispresented in [29]. Further development of EMDwithmasking is proposed in [19].
Although larger classes of composite signals can be decomposed using theexisting EMD with masking techniques, there are still signals where the techni-ques do not work. To indicate the existing problem, we also applied themaskingtechnique in [19] to signal (3.33). This masking technique also fails to workas the obtained IMFs are almost indistinguishable to those obtained using thestandard EMD. Figure 3.3 compares the spectra of the IMFs obtained usingconventional EMD with that of the approach in [19]. As discussed in
0 0.2 0.5 0.8 1.2 1.6 20
5
10
15
20(a)
0 0.2 0.5 0.8 1.2 1.6 20
5
10
15
20
Frequency (Hz)
(b)
Fig. 3.3 Fourier spectrum of the first IMFs of signal (3.33) obtained using (a) the standardEMD and (b) the EMD with masking [19]
74 D.S. Laila et al.
Section 3.1, the EMD results in mode mixing in which the 0.5 and 0.8Hzmodes are seen to interact and create a lower frequency mode at 0.2Hz whichin fact is not the component of the original signal. Techniques to identify andisolate the individual frequency components are discussed in the followingsubsection.
3.3.2 EMD Method with Fourier-Based Masking Technique
Simulation results in Section 3.3.1 show that the ability of EMD to deal withmode mixing is affected by three main factors:
� The signal consists of low frequency components;1
� The magnitude of the highest frequency component is much lower thanothers, particularly the second component, which is directly next to it inthe Fourier spectrum;
� The frequency components are high enough, but they are relatively close toeach other.
Based on the above considerations, the use of masking signal to improveempirical mode decomposition was investigated that extends the previouswork in [19]. The proposed approach permits the analysis of more generalsignals that exhibit mode mixing and results in improved characterization offrequency behavior. The algorithm called the Refined EMD (R-EMD), can besummarized as follows:
R1. Perform FFT on the original signal xðtÞ to estimate the frequency compo-nents f1; f2; . . . ; fn, with f14f24 � � �4fn. These captured frequencies arethe stationary equivalence of the possibly time varying frequency compo-nents of the signal xðtÞ.
R2. Construct the masking signals mask1;mask2; . . . ;maskn�1 using the fol-lowing sinusoidal signal:
maskkðtÞ ¼ Mk sinð2pðfk þ fkþ1ÞtÞ : (3:34)
The value of Mk is empirical and following [19] it is chosen to beMk ¼ 5:5 �Mk, with Mk40 the magnitude of the spectrum of the kthfrequency component.
R3. Identify two cases depending on the physical values of the highest fre-quency components f1 and f2, and their associated amplitudesM1 andM2:Case 1: If one of the following conditions hold:
(a) f1 � 1 and M15R21M2,(b) f141 and f1 � R1f2,
1 We consider the frequency 1Hz as the boundary between the low- and high-frequencysignals, and this is seen from the highest frequency component of a signal.
3 Variants of Hilbert–Huang Transform with Applications 75
(c) f141 and R1f25f15R2f2 and M15R22M2,(d) f141 and f1 � R2f2 and M15R23M2,
where R21 ¼ 1:1, R1 ¼ 1:5, R2 ¼ 2, R22 ¼ 2, and R23 ¼ 0:5, then
1.1. Use only the first masking signal
mask1ðtÞ ¼ M1 sinð2pðf1 þ f2ÞtÞ (3:35)
for the whole process.1.2. Construct two signalsxþðtÞ ¼ xðtÞ þmask1ðtÞ andx�ðtÞ ¼ xðtÞ �mask1ðtÞ.
PerformEMDon each signal following steps S1 to S3 from the standardEMDto obtain all IMFs from each of them, i.e., cþi ðtÞ and c�i ðtÞ; i ¼ 1; 2; . . . ; n andalso the residue rþn ðtÞ and r�n ðtÞ.
1.3. The IMFs and the residue of the signal xðtÞ are
ciðtÞ ¼ðcþi ðtÞ þ c�i ðtÞÞ
2; i ¼ 1; 2; . . . ; n; (3:36)
rnðtÞ ¼ðrþn ðtÞ þ r�n ðtÞÞ
2: (3:37)
1.4. The total reconstructed signal ~xðtÞ is
~xðtÞ ¼Xn
i¼1ciðtÞ þ rnðtÞ : (3:38)
Case 2: If other than conditions (a) – (d) hold, then
2.1. Use all the constructed masking signals (3.34).2.2. Construct two signalsxþðtÞ ¼ xðtÞ þmask1ðtÞ andx�ðtÞ ¼ xðtÞ �mask1ðtÞ.
Perform EMD to each signal to obtain the first IMF only from each one, i.e.,cþ1 ðtÞ and c�1 ðtÞ. The first IMF of xðtÞ is
c1ðtÞ ¼ðcþ1 ðtÞ þ c�1 ðtÞÞ
2: (3:39)
2.3. Obtain the residue r1ðtÞ ¼ xðtÞ � c1ðtÞ.2.4. Use the next masking signal, perform steps 2.2 and 2.3 iteratively using
each masking signal while replacing xðtÞ with the residue obtained ateach iteration, until n� 1 IMFs containing the frequency componentsf2; f3; . . . ; fn are extracted. The final residue rnðtÞ will contain theremainder.
2.5. Compute the final residue, rnðtÞ ¼ xðtÞ � cnðtÞ.2.6. If the residue rnðtÞ is above the threshold value of error tolerance, then
repeat step S2 of the sifting process presented in Section 3.2.2 on rnðtÞ toobtain the next IMF and new residue.
76 D.S. Laila et al.
2.7. The total reconstructed signal ~xðtÞ is
~xðtÞ ¼Xn
i¼1ciðtÞ þ rnðtÞ : (3:40)
Remark 3.3 In the complete R-EMD algorithm, we combine the proposedmasking algorithm, referred to as Case 1, and the masking algorithm from
[19], referred to as Case 2. As can be seen from the required conditions stated in
the algorithm, Case 1 is active when f1 � 1, it is clear that this algorithm takes
care of low-frequency signals and on the other hand Case 1 takes care of the
decomposition of high-frequency signals, so that the whole process can handle
the decomposition for a large sets of signals both with high-and low-frequency
components.Moreover, the values of the parameters R1, R2, R21, R22, and R23 in Case 1
are chosen based on the relation between the frequency as well as the amplitude
of the first two highest frequency components of the composite signals. In this
paper, the values are chosen to suit the application for signals that contain inter-
area oscillation. The choice helps classifying signals that satisfies the three
reasons given at the beginning of this section. Although they are not optimal,
the chosen combination yields effective decomposition for a large set of signals.
In general, seeing the EMD algorithm as a filtering process, we can think of the
parameters as filter gains that are possible to tune if necessary.The R-EMD algorithm gives different procedures for dealing with high-
and low-frequency signals. The main difference is in the way the masking
signals are utilized. For Case 2, we use as many masking signals as the number
of frequencies (or ideally the number of frequencies minus one) we want to
extract from the signal, and we subtract the effect of each masking signal at
every sifting stage, after each IMF is obtained. On the other hand, for Case 1,
we use only the first masking signal, constructed from the first two highest
frequency components peaking on the Fourier spectrum and let the masking
signal stay until the end of the decomposition process. The effect of this
masking signal is then automatically removed from the signal through the
use of formula (3.36).
Remark 3.4 The use of only one masking signal constructed using the twohighest frequency components of the spectrum in Case 1 is justified, as it
satisfies the condition of a masking frequency to be higher than the frequency
to be masked. The significant advantage of this algorithm is that it preserves
well the magnitude of the signal components, which is not the case for other
algorithms as the decomposition often fails. Hence, not only that the instanta-
neous frequency of the IMFs obtained using the R-EMD algorithm is more
meaningful but also we can obtain a quite good estimation of the instantaneous
magnitude of the IMFs.
3 Variants of Hilbert–Huang Transform with Applications 77
3.3.3 EMD Method with Energy-Based Masking Technique
In the previous subsection we use FFT to construct the masking signals, which
implies that to some extent we rely on FFT to separate the frequency compo-
nents of the composite signals. Due to the limitation of FFT, particularly in
dealing with nonlinear and nonstationary signals, this step may deteriorate the
decomposing power of the proposed algorithm as FFT may give wrong infor-
mation in selecting the frequency of the masking signals.Drawing on Case 1 in Section 3.3.2 and the notion of instantaneous mean
frequency in [18], an alternative approach to determining an appropriate mask-
ing signal is suggested, relaxing the dependence to Fourier spectrum for detect-
ing the frequency components of the signal. The algorithm, called A-EMD, is
summarized as follows:
A1. Perform the standard EMD algorithm on the original signal xðtÞ to obtainthe IMFs. Use only the first IMF, c1ðtÞ, which is expected to contain thehighest frequency component of the signal, fmax, but may also containmode mixing with other lower frequency components. Perform HT onc1ðtÞ to obtain its instantaneous frequency f1ðtÞ and instantaneous magni-tude A1ðtÞ.
A2. In the spirit of Hilbert analysis, compute the energy weighted mean of f1ðtÞover L samples, i.e.,
�f ¼PL
i¼1 A1ðiÞf 21ðiÞPLi¼1 A1ðiÞf1ðiÞ
: (3:41)
A3. Observe Case 1 from R3, then replace step 1.1 with the following.
1.1. Construct the masking signal
mask1ðtÞ ¼ M1 sinð2pðm �f ÞtÞ ; (3:42)
whereM1 ¼ maxi¼1;...;L A1ðiÞ and m41.
The rest follow the steps given in the R-EMD algorithm.
Remark 3.5 Note that if the maximum frequency of the composite signal, fmax,is lower than 1Hz, it is common to choose m ¼ 2 as choosing a higher value of
mmay cause the masking signal to be ineffective as its frequency, m�f, would be
much higher than fmax. Comparing with [18], where the masking signal is
computed as mask1ðtÞ ¼ a0 sinð2pð�f=fsÞtÞ, the parameter m replaces the para-
meter fs, the sampling rate. Moreover, we introduced M1 ¼ maxi¼1;...;L A1ðiÞfor analytical choice of a0 in [18]. To complete the formulation of the method,
an efficient algorithm to extract instantaneous attributes is now explored based
on the use of a local HT.
78 D.S. Laila et al.
3.3.4 Local Hilbert Transform
Existing approaches to the calculation of the complex trace (3.5) are based onthe computation of the analytic signal through the Fourier transform. Thistransform, however, has a global character and suffers from problems such asend effects and leakage. In this section, an alternative approach based on filterbanks is proposed that circumvents some of these effects.
Given a signal
xðtÞ ¼X
!
að!Þ cosð!tÞ þ bð!Þ sinð!tÞ ; (3:43)
where a and b are the Fourier coefficients
að!Þ ¼ 1
T
Z T
0
xðtÞ cosð!tÞ dt ;
bð!Þ ¼ 1
T
Z T
0
xðtÞ sinð!tÞ dt :(3:44)
The transformation to a complex time series is
zðtÞ ¼X
!
að!Þ cosð!tÞ þ bð!Þ sinð!tÞ
þ i bð!Þ cosð!tÞ � að!Þ sinð!tÞ½ �
¼ xðtÞ þ ixðtÞ ;
(3:45)
where xðtÞ ¼ xHðtÞ is the quadrature function, or the HT in (3.5).The HT used in this construction is obtained directly by operating the real
component with a convolution filter
xðtÞ ¼ xHðtÞ ¼XM
l¼�Mxðt� lÞhðlÞ ; (3:46)
where hð�Þ is the convolution filter with unit amplitude response and 90 phaseshift. A simple filter that provides an adequate amplitude response and p=2phase response is given by [30] as
hðlÞ ¼2lp sin
2ðpl=2Þ; l 6¼ 0;
0; l ¼ 0;
((3:47)
where �M515M. As M!1 the filter (3.47) yields an exact HT. For Mfinite, the filter introduces ripple effects. To limit these effects, a local HT has
3 Variants of Hilbert–Huang Transform with Applications 79
been developed based on filter banks. As suggested in [31,32], the filter banksare developed such that the flatness of the frequency response is maximal for thelength of the filter. Defining z ¼ ej!, a maxflat filter can be defined by
hðzÞ ¼ 1þ z�1
2
� �2p
Q2p�2ðzÞ (3:48)
where p is the number that determine the zeros at ! ¼ p, and Q is chosen suchthat hðzÞ is halfband. The filter hðzÞ is shifted in frequency by p=2.
3.4 Applications
To test the accuracy of the method, we consider both synthetic data and datafrom transient stability simulations. For comparison, the system response isanalyzed using various algorithms described in previous subsections.
3.4.1 Application to a Synthetic Signal
To verify the accuracy of the method we examine again the synthetic signal(3.33) that we have used in Section 3.3.1. It has been noted earlier, conventionalanalysis fails to identify the individual modal components. In this subsectionwe perform two different major tests of our proposed algorithms. First, wetest the decomposing capability and second, we test the reliability to deal withnonlinear/nonstationary signals.
3.4.1.1 Decomposing Capability Test
Figure 3.4 shows the first three IMFs obtained using the R-EMD algorithm,whilst Fig. 3.5 shows the spectra of the first and the second IMFs. For erroranalysis, IMF1 and IMF2 are also compared with the corresponding compo-nents of the composite signal (3.33), which are also plotted in the backgroundwith dashed lines. It can be seen that the two IMFs match quite well thecorresponding components of frequency 0.8 and 0.5Hz, respectively. More-over, Fig. 3.6 shows the correctness of the whole decomposition results.
Figure 3.7 (a) and (b) shows the instantaneous frequency of IMF1 and IMF2,respectively, which show the frequency components of the composite signal.This figure also compares the instantaneous frequency obtained utilizingthe command hilbert in Matlab, with the convolution approach proposed inSection 3.3.4, where the latter is seen to reduce end effects.
The overall observations have shown two things. First, that R-EMDachieves a higher temporal resolution than the standard methods. Second, the
80 D.S. Laila et al.
0 5 10 15 20 25 30−3
0
3
0 5 10 15 20 25 30
0
0.2
0.4
Time (s)
Res
idue
0 5 10 15 20 25 30−10
0
10
IMF 1
IMF 2
IMF 3
0 5 10 15 20 25 30−20
0
20
Fig. 3.4 IMFs of signal (3.33) obtained using the R-EMD (components of the compositesignal are plotted as dashed line background)
0 0.4 0.8 1.2 1.6 20
2
4
6(a)
0 0.2 0.5 0.8 1.2 1.6 20
5
10
15
20
Frequency (Hz)
(b)
IMF2
IMF1
Fig. 3.5 Fourier spectra of the first and second IMFs of signal (3.33) with R-EMD
3 Variants of Hilbert–Huang Transform with Applications 81
convolution-based Hilbert transformer provides smoother transformation of
the signal by reducing end effects. These yield a more accurate physical char-
acterization of temporal behavior of the signal.We have also tested the energy-based A-EMD algorithm on signal (3.33).
However, we do not include the simulation plots in this paper as they are very
similar to the results from the R-EMD algorithm. We will show the application
of the A-EMD in the next example.
0 5 10 15 20 25 30−30
0
30
Time (s)
Reconstructed
Original
Fig. 3.6 Reconstruction of signal (3.33) from IMFs obtained using R-EMD
0 5 10 15 20 25 300
0.8
1.6
2.4(a) Frequency IMF1
0 5 10 15 20 25 300
0.5
1
1.5(b) Frequency IMF2
0 5 10 15 20 25 300
0.2
0.4
0.6(c) Frequency IMF3
Time (s)
StandardConvolution
Fig. 3.7 Instantaneous frequency of the IMFs of the synthetic signal (3.33).
82 D.S. Laila et al.
3.4.1.2 Reliability to Handle Nonlinear/Nonstationary Signals
The reliability of R-EMD to cope with nonlinearity/nonstationarity is also
tested using clipped signal, where we distort signal (3.33) by clipping its com-
ponents at some ranges. The decomposition performance can be observed in
Fig. 3.8 where we can see that IMF1 and IMF2 preserve the frequency of each
corresponding component, while at the same time trying to capture the shape of
the distorted signals.Although we only show two IMFs in Fig. 3.8, this decomposition actually
yields another three insignificant IMFs plus a residue (as shown partly in
Fig. 3.4). However, feeding the distorted signal using a standard Prony analysis
tool (we have used the BPA/PNNL Ringdown Analysis Tool) for comparison,
we obtain more elements of the signals. Moreover,applying the instantaneous
damping computation formula (3.16), we obtain the comparison between HHT
with R-EMD and Prony analysis as provided in Table 3.1.From the quantities shown in Table 3.1, we can observe that HHT with
R-EMD is more reliable than Prony in dealing with nonlinearities, in this case
clipped signals. While the higher order IMFs and the residue from HHT with
R-EMD are insignificant in terms of magnitude with respect to IMF1 and IMF2,
which represent the components of the signal, some of the extra components
yielded by Prony have higher relative energy than the 0.8Hz component, that
0 5 10 15 20 25 30
0
5
10
IMF 1
IMF 2
0 5 10 15 20 25 30–20
–10
–10
–5
0
10
20
Time (s)
Fig. 3.8 The first two IMFs of the distorted signal (3.33) obtained using the R-EMD(distorted components are plotted as dashed line background)
3 Variants of Hilbert–Huang Transform with Applications 83
may cause misleading in the analysis. Moreover, the damping computationshows that for the 0.8Hz component HHT with R-EMD is twice more accu-rate, while for the 0.5Hz component it is eight times more accurate than Prony.While in HHT with R-EMD the positive damping comes more due to the signaldistorsion and IMF1 and IMF2 do not show any increment in their magnitude,in Prony the 0.8006 and 0.4994Hz components really show the increment intheir amplitude, whereas they should not.
3.4.2 Instantaneous Damping Computation
3.4.2.1 Test I
Given a monofrequency sinusoidal signal
yðtÞ ¼ sinð2ptÞ : (3:49)
as plotted in Fig. 3.9. It is obvious that the signal contains a single frequencyf ¼ p Hz with unit amplitude. Using the information obtained from applyingthe Hilbert transformation to signal (3.49), the instantaneous damping compu-tation gives the result as depicted in Fig. 3.10.
Table 3.1 Comparison of R-EMD results and Prony analysis results
Distorted signal HHT (mean values) Prony
Frequency(Hz) z
Frequency(Hz) z
Frequency(Hz) z Relative Energy
0.8 0.0000 0.8016 0.0013 0.8006 0.0030 0.1496
0.5 0.0000 0.5083 0.0001 0.4994 0.0048 1.0000
0.2151 –0.0238 0.3390 0.0409 0.0375
0.1369 0.0106 0.1866 0.1980 0.0301
0.0833 0.0055 0.0921 0.9536 0.3269
6.3298 0.0615 0.3180
0 5 10 15 20 25 30−1
0.5
0
0.5
1
Time (s)
Fig. 3.9 Test signal (3.49)
84 D.S. Laila et al.
Note that as the signal is monofrequency the decomposition process is notrequired. It can be observed clearly that the instantaneous magnitude of thesignal AðtÞ ¼ 1, the instantaneous frequency fðtÞ ¼ 1 and the instantaneousdamping �ðtÞ ¼ 0.
From this example we have been able to see that the instantaneous dampingcomputation is valid for a monofrequency signal. In the next example, we willtest the case when the signal contains several frequency components.
3.4.2.2 Test II
Given the following sinusoidal signal:
yðtÞ ¼ 6 sinð1:6ptÞ þ 2t sinðptÞ : (3:50)
It is obvious that the signal contains two frequency components 0.8 and 0.5Hz,respectively, with the first component has constant amplitude and the secondone with increasing amplitude. As the signal is a composite signal, we will firstneed to run an EMD process to decompose the components of the signal, andcompute the instantaneous damping of each component.
0 5 10 15 20 25 300.5
1
1.5
0 5 10 15 20 25 300.5
1
1.5
0 5 10 15 20 25 30−1
0
1
Time (s)
(a) Magnitude
(b) Frequency
(c) Damping
Average 2−28 s = −0.000000378≈0
Fig. 3.10 Instantaneous attributes of signal (3.49)
3 Variants of Hilbert–Huang Transform with Applications 85
Up to this point, we have verified that our proposed algorithms provide a better
alternative implementation of HHT in certain applications, and it has shown
to be particularly more powerful when dealing with nonlinear/nonstationary
signals.
0 5 10 15 20 25 30
−60
−30
0
30
60
Time (s)
Fig. 3.11 Test signal (3.50)
0 5 10 15 20 25 300
5
10
Mag
nitu
de
0 5 10 15 20 25 300
0.8
1.6
Freq
uenc
y
0 5 10 15 20 25 30–0.2
0
0.2
Time (s)
Dam
ping
0 5 10 15 20 25 30–6–3036
IMF 1
FilteredOriginal
FilteredOriginal
Average = −0.0048519
Fig. 3.12 Instantaneous attributes of IMF1 from signal (3.50)
86 D.S. Laila et al.
3.4.3 Application to Simulated Data
To verify the proposed method further, we consider simulation data from
transient stability simulations of a complex system. Figure 3.14 depicts a
simplified diagram of the test system showing the study area and major inter-
faces selected for study [11].Several simulation studies have been conducted to assess the applicability of
the proposed technique to analyze composite oscillations resulting from major
system disturbances. In these studies, the southeastern–central interface was
chosen for analysis because this corridor has a dominant participation in three
major inter-area modes. Figure 3.15 shows the power flow response of a key
transmission line interconnection, to the loss of Laguna Verde (LGV) unit #1 at
the Southeastern system. This particular contingency results in undamped
oscillations involving three major inter-area modes at 0.25, 0.50, and 0.78Hz.Using the R-EMDmethod, we decompose the signal into four nonstationary
temporal IMFs and a trend. Figure 3.16 shows the decomposition over the
entire simulation window.It is interesting to compare our result with the one presented in [11, Fig. 5]
which shows the IMFs of the same signal obtained from the standard EMD
algorithm. While in [11, Fig. 5] (the first three IMFs are re-plotted in Fig. 3.17)
0 5 10 15 20 25 300
30
60
Mag
nitu
de
0 5 10 15 20 25 300
0.5
1
Freq
uenc
y
0 5 10 15 20 25 30−1
−0.5
0
Time (s)
Dam
ping
0 5 10 15 20 25 30
−50−25
02550
IMF 2
Filtered
Original
Filtered
Original
Average = −0.12643
Fig. 3.13 Instantaneous attributes of IMF2 from signal (3.50)
3 Variants of Hilbert–Huang Transform with Applications 87
all frequency components appear in IMF1, or in other words the shifting
process does not decompose the signal properly, in Fig. 3.16 we can see that
particularly the first three IMFs are pretty monofrequency. Moreover, it can be
seen fromFig. 3.18 that the proposed R-EMD algorithm accurately extracts the
Fig. 3.14 Geographical scheme of the Mexican interconnected power system
0 5 10 15 20 25 30
600
650
700
750
Time (s)
Rea
l pow
er f
low
(M
W)
Fig. 3.15 Tie-line oscillations following the loss of Laguna Verde unit #1
88 D.S. Laila et al.
0 5 10 15 20 25 30–20
020
IMF 1
0 5 10 15 20 25 30–80
080
IMF 2
0 5 10 15 20 25 30–20
020
IMF 3
0 5 10 15 20 25 30–1
01
IMF 4
0 5 10 15 20 25 30–2
–1
0
Time (s)
Res
idue
Fig. 3.16 The IMFs obtained using R-EMD algorithm
0 5 10 15 20 25 30–80
0
80
IMF 1
0 5 10 15 20 25 30–25
0
25
IMF 2
0 5 10 15 20 25 30–7
0
7
Time (s)
IMF 3
Fig. 3.17 The first three IMFs obtained using standard EMD algorithm
3 Variants of Hilbert–Huang Transform with Applications 89
three dominant frequencies as we can see the instantaneous frequency value of
each IMF is quite constant throughout the time. This has shown that the
decomposition works well.The frequency component of the inter-area modes obtained from the power
signal in this study are 0.7625, 0.4888, and 0.2542Hz; these modes coincide very
well with detailed eigenvalue analysis of the system [11].Another advantage of this approach over other existing methods is that
modal damping can be determined more accurately as the individual (modal)
components are isolated and extracted. This issue will be addressed in our
future research.In order to demonstrate that Hilbert analysis correctly identifies system beha-
vior, we also show that the damping ratio listed in [11, Table III] for the frequency
components 0.7625, 0.4888, and 0.2247Hz, which are respectively 0.0173,
–0.0209, and –0.0351, matches the trend of magnitude of each frequency compo-
nent. As we can observe from Figs. 3.16 and 3.19, the 0.7625Hz component is
decreasing, the 0.4888Hz is increasing, and the 0.2542Hz is also increasing.Figures 3.20 and 3.21 are the corresponding IMFs and instantaneous frequency
computed using the A-EMD method. Comparisons between Figs. 3.20 with 3.16
and Figs. 3.21 with 3.18 show that the two methods give results that show good
agreement. In both cases, the local HT is found to reduce the end effects.
0 5 10 15 20 25 30
0.4
0.8
1.2
(a) Frequency IMF1
0 5 10 15 20 25 30
0
0.5
1
(b) Frequency IMF2
0 5 10 15 20 25 30
0
0.3
0.6
(c) Frequency IMF3
Time (s)
StandardConvolution
Fig. 3.18 Instantaneous frequency of the IMFs showing the frequency of the inter areaoscillation
90 D.S. Laila et al.
0 5 10 15 20 25 300
10
20
(a) Magnitude IMF1
0 5 10 15 20 25 300
40
80
(b) Magnitude IMF2
0 5 10 15 20 25 30
10
20
(c) Magnitude IMF3
Time (s)
StandardConvolution
Fig. 3.19 Instantaneous amplitude of the IMFs showing the growth of each component
0 5 10 15 20 25 30–20
020
IMF 1
0 5 10 15 20 25 30–80
080
IMF 2
0 5 10 15 20 25 30–20
020
IMF 3
0 5 10 15 20 25 30–2
02
IMF 4
0 5 10 15 20 25 30
–2–1
0
Time (s)
Res
idue
Fig. 3.20 The IMFs obtained using A-EMD algorithm
3 Variants of Hilbert–Huang Transform with Applications 91
The use of a modified masking signal in our numerical formulation deservessome comment. It may be tempting to question why we are using (3.42) with theterm m�f instead of using the maximum value of the instantaneous frequencyf1ðtÞ of the first IMF that is logically the maximum frequency component of thesignal and replacem�fwithmf1;max where 15m52. Figure 3.22 shows the reason
0 5 10 15 20 25 30
0.4
0.8
1.2
(a) Frequency IMF1
0 5 10 15 20 25 30
0
0.5
1
(b) Frequency IMF2
0 5 10 15 20 25 30
0
0.3
0.6
(c) Frequency IMF3
Time (s)
StandardConvolution
Fig. 3.21 Instantaneous frequency of the IMFs obtained using A-EMD algorithm
0 5 10 15 20 25 300
2
4
6
Time (s)
Fig. 3.22 Instantaneous frequency of the first IMF obtained using the standard EMD algorithm
92 D.S. Laila et al.
for this. The odd peak appears on the plot will give a wrong information of thevalue of the maximum frequency component that lead to the frequency of theconstructed masking signal too high. Clearly, the use of �f in (3.42) helps infiltering fictitious variations which in turn results in improved systemcharacterization.
To complete our study, we also make a comparison between HHT withA-EMD and Prony analysis tool. The result is shown in Table 3.2. It can beobserved that the results obtained using Prony involve some ambiguities as canbe seen for the components 0.4915Hz and 0.5276Hz as well as the components0.2494 and 0.2758 as they are coming as pairs. Although the relative energy ofthe pairing components are significantly different, it tells us that the dampinginformation does not show the real damping ratio of the true component 0.5and 0.25Hz, respectively. If the components of the monitored signal are notknown, this creates confusion in interpreting the results. On the contrary, HHTwith A-EMD gives more reliable and consistent results for the decompositionand the damping computation.
These findings are very useful for monitoring and analysis of the inter-areaoscillation for power system. It has simplified the analysis, as in this way theinstantaneous frequency and instantaneous damping of the inter-area oscilla-tion can be seen clearly and directly from visual observation, which is veryuseful when engineers have to make quick decision to take action in urgentsituations.
3.4.4 Application to Measured Data
To further test the ability of themethod to treat complex data, we analyze in thissection data from time-synchronized phasor measurements and the results arecompared to those of previous investigations based on standard EMD analysis.The data set chosen for investigation is the real power flow between two majorsubstations obtained from a recording of an oscillatory event in the Mexicansystem [12] which is depicted in Fig. 3.14. The two substations are TTH, which
Table 3.2 Comparison between HHT with A-EMD and Prony for inter-area modes analysis
Modes HHT (mean values) Prony
Frequency Frequencyz
Frequencyz Relative Energy(Hz) (Hz) (Hz)
0.78 0.7625 0.010 0.7678 0.0119 0.1032
0.5 0.4888 �0.010 0.4915 �0.0271 1.0000
– – 0.5276 0.0250 0.0489
0.25 0.2542 �0.007 0.2494 �0.0257 0.2542
– – 0.2758 0.0257 0.0622
0.0978 �0.030 1.1983 0.1548
0.8635 0.0420
3 Variants of Hilbert–Huang Transform with Applications 93
is located at the Central system, and TTE, which is located at the Southeastern
system. The measurements were obtained over a 4 s period at a sampling rate of
2ms.Figure 3.23 shows the time evolution of the selected signal and Fig. 3.24
shows its corresponding spectrum. The power spectrum discloses the presence
of three major modes at 0.48, 0.44, 0.39, and 0.25Hz. Detailed examination of
the record suggests the presence of nonstationary characteristics. Also, the
signal exhibits a strong nonlinear trend associated with the frequency recovery
characteristics of the system.Based on the recorded data, the refined EMD method was used to decom-
pose the signal into several components, each with single frequency. The
objective is to assess the ability of the method to capture transient behavior in
real-world observational data.Figure 3.25 shows the three dominant components decomposed from the
data by the A-EMD. The other components have small amplitudes compared
with the first two IMFs and are not studied here. Because each component
0 100 200 300 400350
360
370
380
390
400
Time (s)
Rea
l pow
er f
low
(M
W)
Fig. 3.23 Measured real power flow between TTH and TTE of the Mexican system
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
Frequency (Hz)
X:0.4472Y:0.9899
X:0.2566Y:0.4291
X:0.3935Y:0.4718
X:0.4863Y:0.7534
Fig. 3.24 Fourier spectrum of the measured signal
94 D.S. Laila et al.
captures a particular timescale, the EMD technique effectively decomposes theoriginal signal into a set of IMF components, which can be linear or nonlinearand time-varying. Visual inspection of the IMFs in Fig. 3.25 shows that therefined method produces nearly sinusoidal oscillations from which meaningfulinstantaneous parameters can be extracted.
A key advantage of the method is that the exact time at which each modebecomes dominant can be determined. Useful insight into the physical inter-pretation of the IMFs can be obtained from the spectra of the first two IMFs inFig. 3.26. These results are consistent with the spectra of the original signalshown in Fig. 3.24, hence confirm the accuracy of themethod. The refinedHHTmethod, however, captures, additionally, the evolution of the frequencies pre-sent at every time instant of the records.
0 50 100 150 200 250 300 350 400–50
0
50IM
F 1
0 50 100 150 200 250 300 350 400–10
0
10
IMF 2
0 50 100 150 200 250 300 350 400–5
0
5
Time (s)
IMF 3
Fig. 3.25 Extracted IMFs from the R-EMD algorithm
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
Frequency (Hz)
(a)
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
Frequency (Hz)
(b)
x:0.4863y:0.6995
x:0.4448y:0.928
x:0.391y:0.3572
x:0.2542y:0.4002
Fig. 3.26 Spectra of the IMF1 (a) and IMF2 (b)
3 Variants of Hilbert–Huang Transform with Applications 95
Having determined the individual IMFs, efforts were directed toward asses-
sing the ability of the method to capture abrupt changes in modal behavior.
Figures 3.27 and 3.28 show themodal attributes for modes 1 and 2, respectively,
obtained using the theory in Section 3.3. In these plots, the upper panel and
middle panel show the instantaneous amplitude and frequency computed using
the conventional approach in [18]. The lower panel shows the instantaneous
damping obtained from (3.16).As it may observed from these plots, in all cases, the modal frequency is very
accurately by the refined method. In addition, careful inspection of the middle
panel in Fig. 3.4 shows that the A-EMDmethod suppresses anomalous peaks in
the instantaneous frequency computations. This, in turn, results in improved
characterization of system damping as discussed below.As noted above, each IMF emphasizes a different oscillation mode with
different amplitude and frequency content. The first IMF captures the highest
frequency mode at 0.44 and 0.48Hz, while IMF 2 captures the time evolution of
the 0.25Hz mode.Also of interest, the analysis of modal damping in the lower panels in Figs. 3.27
and 3.28 show that the method is able to characterize local temporal behavior.
This information might be used to trigger control corrective or preventive control
actions. This is, however, not discussed here.
125 130 135 140 145 150 155 160 165 170 1750
10
20
Mag
nitu
de
125 130 135 140 145 150 155 160 165 170 1750
0.5
1
Freq
uenc
y
125 130 135 140 145 150 155 160 165 170 175–0.4
–0.2
0
Time (s)
Dam
ping
FilteredUnfiltered
Average = –0.00124
0.48 Hz0.44 Hz
IMF1
Fig. 3.27 Instantaneous attributes of IMF1
96 D.S. Laila et al.
To verify the accuracy of the proposed method, detailed studies were con-
ducted using Prony analysis. Prior to Prony analysis, the slow trend was
removed by subtracting the nonlinear trend from the original signal; this
makes the analysis more accurate and enable closer comparison between both
approaches. Table 3.3 shows Prony analysis results for the recorded bus fre-
quency signal in Fig. 3.23.Overall, the results are found to be in good agreement although it should also
be stressed that Prony cannot capture the modes simultaneously as HHT does.
Consistent Prony results are found only when the time intervals match those of
Hilbert analysis. While several techniques can be used to improve damping
estimates this is not the main focus of this chapter.As discussed in Section 3.2.5, the IMFs form a set of complete basis for
modal decomposition of the oscillatory processes, which is derived from the
data themselves. Also, for linear signals, the resulted IMFs are numerically
100 110 120 130 140 150 160 170 180 190 2000
5
10M
agni
tude
100 110 120 130 140 150 160 170 180 190 2000
0.25
0.5
Freq
uenc
y
100 110 120 130 140 150 160 170 180 190 200–0.5
0
0.5
Dam
ping
Time (s)
Average = –0.00169
0.25 Hz
IMF2
Fig. 3.28 Instantaneous attributes of IMF2
Table 3.3 Prony analysis results
Time intervals Dominant modes Frequency Damping Energy
125–175 2 0.440 0.0025 1.0000
125–140 2 0.414 0.0192 1.0000
3 0.229 0.0436 0.7930
140–160 1 0.487 0.0228 0.0000
3 Variants of Hilbert–Huang Transform with Applications 97
orthogonal. Following [13], in Section 3.2.5 the orthogonality index is alsointroduced for the case of nonlinear signals.
To verify the completeness of the decomposition, the full data was recon-structed using the decomposed IMFs. Figure 3.29 compares the original timeseries with the reconstructed signal obtained using the EMD. As shown in thisplot, the results are practically indistinguishable over the entire length of thesignal, thus giving confidence to the results.
Moreover, Table 3.4 compares the IO for the conventional approach and therefined EMD method. As may be seen from this table, the IO is very smallshowing that the proposed method produces physically and numerically realis-tic IMFs. It is also interesting to note that the refined algorithm reduces slightlythe orthogonality condition of the decomposition.
3.5 Conclusion
In this chapter various refinement of EMD algorithms and a local Hilberttransformation have been presented. These algorithms are well suited forextracting and characterizing temporal behavior and can be applied to typicalsignals found in power oscillatory processes. Applications to synthetic signalsas well as measurement signals show that the proposed algorithms provideimproved visualization and characterization of complex oscillations involvingmulti-timescale behavior. Several extensions to the proposed analytical toolsare possible including improvements in the masking signal method, the compu-tation of local Hilbert transformations and the estimation of modal properties.These issues warrant further investigation.
0 100 200 300 400–30
–20
–10
0
10
20
Time (s)
Reconstructed signalOriginal signal
Fig. 3.29 Reconstruction of signal from IMFs obtained using A-EMD
Table 3.4 Index of orthogonality of the measurement signal
Conventional approach A-EMD algorithm
IO 7.128e� 06 6.977e� 06
98 D.S. Laila et al.
Acknowledgments The authors are thankful to Jegatheeswaran Thambirajah and NinaThornhill from the Chemical Engineering Department, Imperial College London, for usefuldiscussion and good teamwork in pursuing the research in this topic that make it possible forthe authors to contribute this chapter. The authors also thank ABB, Switzerland, andNational Grid, UK, for the research collaboration done within this research project.
References
1. G.Rogers. Power System Oscillations. Kluwer Acad., Boston, 2000.2. J. Paserba. Analysis and control of power system oscillation. CIGRE Special Publication
38.01.07, Technical Brochure, Technical Brochure 111, 1996.3. P. Kundur. Power System Stability and Control. McGraw Hill, 1994.4. P. W. Sauer and M. A. Pai. Power System Dynamics and Stability. Prentice Hall, 1998.5. J. F. Hauer, C. J. Demeure, and L. L. Scharf. Initial results in Prony analysis of power
system response signals. IEEE Trans. Power Syst., 5(1):80–89, 1990.6. J. F. Hauer. Application of Prony analysis to the determination of modal content and
equivalent models for measured power system response. IEEE Trans. Power Syst.,6:1062–1068, 1991.
7. D. J. Trudnowski, M. K. Donnelly, and J. F. Hauer. A procedure for oscillatory para-meter identification. IEEE Trans. Power Syst., 9(4):2049–2055, 1994.
8. I. Kamwa and L. Gerin-Lajoie. State-space system identification toward MIMO modelsfor modal analysis and optimization of bulk power systems. IEEE Trans. Power Syst.,15(1):326–335, 2000.
9. D. R. Ostojic and G. T. Heydt. Transient stability assessment by pattern recognition inthe frequency domain. IEEE Trans. Power Syst., 6(1):231–237, 1991.
10. D. R. Ostojic. Spectral monitoring of power system dynamic performances. IEEE Trans.Power Syst., 8(2):445–451, 1993.
11. A. R. Messina and V. Vittal. Nonlinear, non-stationary analysis of interarea oscillationsvia Hilbert spectral analysis. IEEE Trans. Power Syst., 21(3):1234–1241, 2006.
12. A. R. Messina, V. Vittal, D. Ruiz-Vega, and G. Enr’iquez-Harper. Interpretation andvisualization of wide-area PMU measurements using Hilbert analysis. IEEE T. PowerSyst., 21(4):1763–1771, 2006.
13. N. E. Huang, Z. Shen, S. R. Long, M. L. Wu, H. H. Shih, Q. Zheng, N. C. Yen C. C.Tung, and H. H. Liu. The empirical mode decomposition and the Hilbert spectrum fornonlinear and nonstationary time series analysis. Proc. Royal Soc. London, 454:903–995,1998.
14. J. C. Echeverria, J. A. Crowe, M. S. Woolfson, and B. R. Hayes-Gill. Application ofempirical mode decomposition to heart rate variability analysis.Med. Biol. Eng. Comput.,39(4):471–479, 2001.
15. B. M. Battista, C. Knapp, T. McGee, and V. Goebel. Application of the empirical modedecomposition and Hilbert-Huang transform to seismic reflection data. Geophysics,72(2):H29–H37, 2007.
16. M. A. Andrade, A. R. Messina, C. A. Rivera, and D. Olguin. Identification of instanta-neous attributes of torsional shaft signals using the Hilbert transform. IEEE T. PowerSyst., 19(3):1422–1429, 2004.
17. Z. Wu and N. E. Huang. A study of the characteristics of the white noise using theempirical mode decomposition method. Proc. Royal Soc. London A, 460:1597–1611,2004.
18. R. Deering and J. F. Kaiser. The use of a masking signal to improve empirical modedecomposition. In Proc. IEEE Int. Conf. on Acoustic, Speech and Signal Proc. (ICASSP’05), 4:485–488, 2005.
3 Variants of Hilbert–Huang Transform with Applications 99
19. N. Senroy and S. Suryanarayanan. Two techniques to enhance empirical mode decom-position for power quality applications. In Proc. IEEE Power Eng. Soc. Gen. Meet.,Tampa, Florida, 1–6, 2007.
20. N. Senroy, S. Suryanarayanan, and P. F. Ribeiro. An improved Hilbert-Huang methodfor analysis of time-varying waveforms in power quality. IEEE Trans. on Power Syst.,22(4):1843–1850, 2007.
21. E. Del’echelle, J. Lemoine, and O. Niang. Empirical Mode Decomposition: An analyticalapproach for sifting process. IEEE Signal Process. Lett., 12:764–767, 2005.
22. R. Srinivasan, R. Rengaswamy, and R. Miller. A modified empirical mode decomposi-tion (EMD) process for oscillation characterization in control loops. Control Eng. Pract.,15:1135–1148, 2007.
23. R. C. Sharpley andV. Vatchev. Analysis on the intrinsic mode functions.Constr. Approx.,24:17–47, 2006.
24. P. Flandrin. Time-Frequency Time-Scale Analysis. Acad. Press, 1999.25. R. R. Zhang, L. vanDemark, J. Liang, and Y. Hu. On estimate site damping with soil
non-linearity from earthqueake recordings. 2004.26. M. Feldman. Non-linear system vibration analysis using Hilbert transform � Free
vibration analysis method ’FREEVIB’.Mech. Syst. Signal Process., 8:119–127, 1994.27. M. Feldman. Non-linear free vibration identification via the Hilbert transform. J. Sound
Vib., 208:475–489, 1997.28. G. Rilling. Empirical mode decomposition, Matlab codes with examples. http://perso.
ens-lyon.fr/patrick.flandrin/emd.html, retrieved January 2008.29. R. Deering. Final-Scale Analysis of Speech using Empirical Mode Decomposition: Insight
and Applications, PhD Thesis. Duke University, 2006.30. W. B. White and S. E. Pazan. Hindcast/forcast of ENSO events based upon the redis-
tribution of observed and model heat content in the western tropical Pacific, 1964–86.Control Phys. Oceanogr., 17:264–280, 1987.
31. R. L. C. Spaendonck, F. C. A. Fernandes, R. G. Baraniuk, and J. T. Fokkema. LocalHilbert transformation for seismic attributes. InProc. EAEG 64th Conference &TechnicalExhibition, 2002.
32. G. Strang and T. Nguyen.Wavelets and Filter Banks. Wellesley - Cambridge Press, 1997.
100 D.S. Laila et al.
Chapter 4
Practical Application of Hilbert Transform
Techniques in Identifying Inter-area Oscillations
T.J. Browne, V. Vittal, G.T. Heydt, and Arturo Roman Messina
Abstract Disturbances in large power systems can exhibit nonlinear, time-
varying behavior. Traditional modal identification from field data is via tech-
niques, such as Prony analysis, which assume data stationarity. The Hilbert
transform and analytic function can be used to analyze inter-area oscillatory
behavior of power systems with the stationarity assumption relaxed. However,
reducing the data to simple numerical results can be achieved more effectively
when stationarity is assumed. The application process is not straightforward
and subtle changes can yield considerable variation in the results observed. An
example is the effect of discrete time calculations of the Hilbert transform over a
window of finite length. Application of the newer modal identification techni-
que, Hilbert analysis, is examined relative to the more established Prony ana-
lysis, with particular reference to the considerable structural differences which
exist between the two methods. Prony analysis yields modes which are directly
expressed as exponentially modulated sinusoids, whereas the Hilbert method
provides amore general solution. Synthetic andmeasured signals are used in the
comparison. Some general conclusions are drawn from the analysis of several
signals, including sets of measured field data.
4.1 Inter-area Oscillations in Power Systems
Inter-area oscillations in power systems [1, 2] occur when geographically iso-
lated generator groups swing against each other. When inter-area oscillations
are unstable, the interconnected network can be broken up, potentially leading
to islanding. Real-time analysis of power system measurements is intended to
assist in assessing the existence and stability of inter-area oscillations.
T.J. Browne (*)Ira A. Fulton School of Engineering, Department of Electrical Engineering, ArizonaState University, Tempe, AZ, USAe-mail: [email protected]
A.R. Messina, Inter-area Oscillations in Power Systems,Power Electronics and Power Systems, DOI 10.1007/978-0-387-89530-7_4,� Springer ScienceþBusiness Media, LLC 2009
101
Present analysis techniques are yet to be applied to real-time power systemmeasurements for consistent stability identification. One modal parameteridentification method, Prony analysis [3], provides estimates of dampingwhen the power system can be satisfactorily modeled with linear approxima-tions. Prony analysis can be applied successfully to some sections of the mea-sured bus frequency signal given in Fig. 4.1; however, application of the Pronymethod to the entire signal does not yield meaningful results.
The Hilbert transform H of a signal x(t) is obtained by convolution with aninverse time function:
H½xðtÞ� ¼ 1
pt� xðtÞ: (4:1)
Unlike the Fourier transform (FT), the Hilbert transform of a time-domainsignal is itself a time-domain signal. When applied to a single mode, the Hilberttransform offers some useful properties in modal parameter identification. Arelatively new technique, empirical mode decomposition (EMD) [5], isemployed to isolate individual modes within the signal prior to application ofthe Hilbert transform.
This chapter will demonstrate practical considerations associated with theapplication of the Hilbert transform and EMD in identifying modal parametersunder inter-area oscillatory conditions [4], which are typically in the range0.1–2 Hz [6].
4.2 Present Identification Techniques
4.2.1 Prony Analysis
Prony analysis [3, 7, 8] is a curve-fitting technique, which decomposes a signalinto sinusoids modulated by exponential growth or decay – that is, into modesof the form
Fig. 4.1 Measured bus frequency signal during a disturbance in the Mexican InterconnectedSystem. Reproduced, with permission, from [4]. # 2008 IEEE
102 T.J. Browne et al.
xðtÞ ¼Me�t sinð!dtþ ’Þ: (4:2)
Although the underlying mathematics, which is based on linear predic-tion, is eighteenth-century technology, numerical ill-conditioning makes thetechnique computationally intensive [3], delaying widespread use untilaround the 1980s.
Being an eigenvalue-based technique, Prony analysis delivers the real andimaginary parts of the mode, corresponding respectively to the attenuationfactor � and the modal damped frequency !d, fromwhich the natural frequency!n and damping ratio � can be calculated using well-known linear systemtheory. In addition, Prony analysis is able to identify the two additionalmodal parameters, namely amplitude and phase shift, which allow a mode tobe described uniquely in the time domain, making reconstruction of the signalfrom its Prony-identified components feasible.
The key assumption involved in Prony analysis is that the signal is station-ary: where the assumption is not met, the linear system theory underpinning theProny algorithm is inapplicable and so the decomposition is not necessarilymeaningful. Factors such as window length and sampling rate can affect theaccuracy with which a nonstationary signal can be approximated by a sum ofstationary modes.
Prony software [9] has been used to assist in verifying results of Hilbertanalysis and to provide a point of reference. Adjustments to options in thesoftware can have significant impact on the results obtained.
4.2.2 Fourier Methods
Certainly, the oldest and perhaps the most widely used signal analysismethods are based on Fourier technology. This is the resolution of a signalx(t) into its component sine waves, that is, a frequency spectrum. Themethod is based on the Fourier series for x(t). If x(t) is periodic with periodT, then 2p=T rad/s is the fundamental frequency, !o; and one finds that theFourier series is
xðtÞ ¼ ao2þX1
n¼1an þ a�n� �
cosðn!otÞ þ j an � a�n� �
sinðn!otÞ (4:3)
where ai are the Fourier coefficients
ai ¼2
T
Zt1þT
t1
xðtÞe�ji!ot dt: (4:4)
4 Practical Application of Hilbert Transform Techniques 103
In engineering, the series is often rewritten as
xðtÞ ¼ ao2þX1
n¼1bn cosðn!otÞ þ cn sinðn!otÞ (4:5)
where
bn ¼2
T
Zt1þT
t1
xðtÞ cosðn!otÞ dt (4:6)
cn ¼2
T
Zt1þT
t1
xðtÞ sinðn!otÞ dt: (4:7)
Note again that x(t) must be periodic, and also note that there are no ‘‘inter-frequency components’’ (i.e., no ‘‘fractional harmonics’’).
In the present application, it is convenient to use a numerical analysismethod, the discrete Fourier transform (DFT), which is
FDðxðk�TÞÞ ¼ Xðm��Þ ¼ 1ffiffiffiffiNp
XN
m¼1xði�TÞe�jim���T (4:8)
where FD(.) denotes the DFT, �T is the sampling time,N is the number of pointsin the sample, and N must span an integer number of periods of the signalx(t).The frequency domain resolution of the DFT is �� which is 2p/N�T r/s.In the 1960s, several researchers found a clever way to find the DFT of a signalusing certain symmetry properties of the complex exponential. The method iscalled the fast Fourier transform (FFT), and it is an exact method of evaluatingthe DFT of a signal. The advantage of the FFT is that not every term in the DFTneed be explicitly evaluated, and therefore the computational burden is reduced.It is well known that the computational burden in the FFT becomes N log(N)rather than N2 as implied by the definition of the DFT.
As the sampling interval becomes smaller and smaller, and as N goes toinfinity, it is a simple matter to show that the DFT becomes the FT,
FffðtÞg ¼ Fðj!Þ ¼Zþ1
�1
fðtÞe�j!t dt: (4:9)
The inverse FT is
F�1fFðj!Þg ¼ fðtÞ ¼ 1
2p
Zþ1
�1
Fðj!Þeþj!t d!: (4:10)
104 T.J. Browne et al.
Note that:
� ‘‘Almost all’’ periodic functions x(t) possess a Fourier series.� A Fourier series is a representation of x(t) as a sum of sines and cosines that
have frequencies that are integer multiples of 1/T Hz (T = the period of thesignal f(t)).
� Each term in the series is called a harmonic.� The Fourier series, the DFT, and the FFT are the minimum least-squares fit
of a sine–cosine series to a signal x(t).That is, there is no better calculation ofthe series coefficients that is possible. However, if the series or the discretetransforms are calculated using faulty assumptions, for example, x(t) isassumed periodic but it is not actually periodic, then one may not obtain asuitable representation of x(t) in the frequency domain.
As an historical note, Fourier methods are credited to Baron Jean BaptisteJoseph Fourier (1768–1830) who worked at the Ecole Normale in Paris. Theliterature of Fourier methods is huge. A sampling of useful references for thesubjects touched upon is [10–14].
Proceeding to examine the use of Fourier technology to find signal damping,use the DFT of the sampled time signal x(k�T),
Xðm��Þ ¼ 1ffiffiffiffiNp
XN�1
i¼0xði�TÞe�jim�T�� (4:11)
where the signal x(k�T) is assumed to be periodic with periodicity N and�� =2p/N�T. Of course, all DFTs have a discrete spectrum. Note that theassumption of periodicity of x(k�T) results in discrete spectrum X(m��) withno ‘‘leakage.’’ Leakage refers to omission of terms in theDFT sum due to failureto span an integer number of periods of x(t). If the DFT (and hence the FFT) dospan an integer number of periods of x(t), and, indeed, if x(t) is truly periodic,X(m��) is ‘‘exact.’’
Using X(m��), the parameter � is calculated,
� ¼ 2pBW
(4:12)
where BW is the (3 dB) bandwidth of the spectrum |X(m��)|. This parameter �is the damping parameter in the case of a single-mode signal.
Because actual measurements give x(k�T) that is not generally periodic, therequired time window for the calculation of X(m��) is infinite and finitesampling is undersampling. Undersampling results in low-pass filtering ofX(m��). This problem is not important if the signal x(t) is nearly periodic,but a damped sine wave, the expected signal, is clearly aperiodic.
Most engineers are familiar with minimum phase systems, and therefore maybe unfamiliar with the trivial property that the pole locations in the right (i.e.,
4 Practical Application of Hilbert Transform Techniques 105
nonminimum phase systems) or left half-planes may be readily determined fromthe phase spectrum of the DFT (and FFT) near resonance. All methods basedon the assumption of signal behavior as a second-order system suffer from thedifficulty that if the signal is not a simple second-order system, erroneous resultsmay occur – and the DFT approach shown here has that property. Actualpower signal are results of nonlinear, time-varying, high-order systems, andtherefore all methods of identifying modes may give inaccurate results whendissimilar modes are nearly collocated or time variation of the signal (and itssource system) come into play. The short-time Fourier transform (STFT) mayaddress this situation. Parameter identification techniques have been devel-oped, which exploit sliding windows [15] and spectral decay [16].
4.3 The Hilbert Transform and Analytic Function
4.3.1 Hilbert Transform Properties
The Hilbert transform XH(t) of a signal x(t) is defined [17] as the convolution
XHðtÞ ¼ H½xðtÞ� ¼ 1
pt� xðtÞ: (4:13)
An alternative definition [18] reverses the sign. In contrast to the FT, theHilbert transform does not introduce a frequency variable. Rather, the trans-form of a time-domain signal remains in the time domain.
Properties of the convolution mean that the inverse Hilbert transformH–1 ofa signal YH(t) is given by
yðtÞ ¼ H�1 YHðtÞ½ � ¼ �1pt� yðtÞ ¼ �H YHðtÞ½ � (4:14)
and so a double application of the Hilbert transform to a signal returns thenegative of the original signal.
Because the signals of principal concern are damped sinusoids, Hilberttransforms of trigonometric and exponential functions are of interest. It canbe shown that the Hilbert transform of a cosine function is a sine function.Similarly, (4.14) implies that the Hilbert transform of a sine function is anegative cosine function. Hahn [17] gives the Hilbert transform of a rotatingphasor as
H½ej!t� ¼ �sgnð!Þej!t (4:15)
where sgn is the signum (sign) function.Many other Hilbert transform pairs areprovided in [17]. Damped sinusoids will be addressed by the analytic function inSection 4.3.2.
106 T.J. Browne et al.
The linearity of the convolution operation confers useful properties upon theHilbert transform. The Hilbert transform of the sum of two functions is the sumof the Hilbert transforms of the individual functions. Similarly, if k is a constantthen for any function x(t),
H½kxðtÞ� ¼ kH½xðtÞ�: (4:16)
Products of functions are more complicated. Bedrosian [19] showed that theproduct of two analytic functions x1(t) and x2(t) is given by
H½x1ðtÞx2ðtÞ� ¼ x1ðtÞH½x2ðtÞ� (4:17)
when x1(t) and x2(t) are spectrally disjoint. However, this spectral condition istoo restrictive to be useful in modal identification: exponential functions arewideband signals, and so an exponential decay is unsuitable for use as eitherx1(t) or x2(t). Cain [20] developed an unrestricted expression for a Hilberttransform product of two complex (or, by implication, real) functions,
H½x1ðtÞx2ðtÞ� ¼ x1ðtÞH½x2ðtÞ� þ x2ðtÞH½x1ðtÞ� þH½H½x1ðtÞ�H½x2ðtÞ��: (4:18)
Because the final termmeans that evaluation of this expression still requires aHilbert transform of a product, (4.18) does not appear to have found wide-spread use, at least in the power engineering literature. More heuristic techni-ques tend to be utilized when Hilbert transform products are required.
4.3.2 Modal Parameters in Terms of the Analytic Function
The analytic function [17, 21] is a complex signal derived from both the originalsignal and its Hilbert transform as
XAðtÞ ¼ xðtÞ þ jH½xðtÞ� (4:19)
For a single-mode signal of the form of (4.2) for any real t, the analytic signalcan be expressed as
XAðtÞ ¼Me�t cosð!dtþ ’Þ þ jMe�t sinð!dtþ ’Þ
¼Me�tejð!dtþ’Þ:(4:20)
Therefore, the magnitude of the analytic function is related to the attenua-tion factor � of the mode via
ln XAðtÞj j ¼ �tþ ln M (4:21)
4 Practical Application of Hilbert Transform Techniques 107
while the phase angle of the analytic function can be related to the modaldamped frequency !d as
ffXAðtÞ ¼ !dtþ ’ (4:22)
For a signal of the form (4.2), (4.21), and (4.22) imply that the attenuationfactor can be found from the slope of a plot of ln |XA(t)| versus time, whereas themodal damped frequency can be established from the slope of a plot of ffXA(t)versus time. By implication, the two parameters could be found instead using aleast-squares fit for online applications.
As an example of the technique, consider the signal
fðtÞ ¼ e�t=10 sinð2p0:27tÞ; �15t51: (4:23)
The Hilbert transform of f(t) is
FHðtÞ ¼ �e�t=10 cosð2p0:27tÞ; �15t51: (4:24)
giving the analytic signal
FAðtÞ ¼ �je�t=10 ej2p0:27t; �15t51: (4:25)
It can be seen that ln |FA(t)| ¼ –0.1t and so the attenuation factor � ¼ –0.1.Similarly, ffFA(t) ¼ 0.27t, yielding !dt ¼ 0.27 rad/s.
The problem remains to decompose a measured signal xm(t) into individualcomponents of the form (4.2). Two techniques which have been applied to thisend are Prony analysis [3, 7, 8], discussed in Section 4.2, and EMD [5, 22]. AsProny analysis delivers the modal properties without further calculation, onlythe EMD is applied in conjunction with the Hilbert transform.
It should be noted here that it is not the entire set of modal parameters whichis critical to the intended application. The first priority, stability assessment,implies that the sign of the attenuation factor � is the most important parameterto be identified for each mode: a negative value indicates a stable mode whereasa positive value denotes instability. Errors in the other parameters can betolerated acceptably should stability assessment be of sole interest. However,the following implementation of the technique for stability assessment and forthe purposes of control implementation become of increasing interest. If themodal frequency is assessed incorrectly then any control action will be misdir-ected and may be, at best, ineffective.
4.3.3 Hilbert Transform Implementation
The theory presented in Section 4.3.2 was concerned purely with the calculationof the Hilbert transform in continuous time. For the modal identification
108 T.J. Browne et al.
problem, utilizing sampled power systemmeasurements, continuous time beha-vior can only be approximated. Some work has been carried out on a specificdiscrete Hilbert transform [17].
It can be shown [23] that the Fourier spectra of the original and analytic signalsare identical at positive frequencies, whereas at negative frequencies, the FT of theanalytic signal is zero. Software [24] employed for calculation of the Hilberttransform exploits this property, using the discrete approximation to the FT togenerate the analytic signal directly, without first determining a Hilbert transform.The Hilbert transform is extracted as the imaginary part of the analytic signal.
A key advantage of calculation via the DFT is the speed and simplicity of theFFT, compared with the operations required for the direct convolution. How-ever, disadvantages exist also, primarily aliasing problems inherent in applica-tion of the DFT to a signal which is unlikely to be band-limited. A highsampling frequency can mitigate some, but not all, of these aliasing problems.
4.3.4 Instantaneous Frequency
The concept of ‘‘instantaneous frequency,’’ arguably meaningless for a single-frequency sinusoid, has been examined for nonlinear signals and is used [5] as ajustification for the decomposition to be introduced in Section 4.5. Only oscilla-tions which lie within the frequency range of 0.1–2.0 Hz characteristic of inter-area modes [25] are of interest. It can be argued that when the sampling rate issufficiently high, a meaningful instantaneous frequency of a signal can becalculated using only two or three samples in the vicinity of every samplinginstant under consideration. Boashash [26] describes several possible techni-ques, and notes that the instantaneous frequency is not necessarily reconcilablewith frequencies obtained from amore global Fourier analysis. The logic of thisstatement can be verified by considering an infinite sine wave: the instantaneousfrequency around the zero crossings is necessarily greater than that at the peaksand troughs, and so a range of frequencies, which would not be identified bydirect application of Fourier techniques, can be conceived.
Instantaneous damping is less obviously definable, despite an assertion [27]in a civil engineering context that the concept makes physical sense. Thedamping in an oscillatory signal has no obvious meaning if only a singleoscillatory cycle is to be examined.
4.4 Application to Single-Mode Signal
The general expression
H cos t ¼ sin t (4:26)
holds over the domain –1< t <1.
4 Practical Application of Hilbert Transform Techniques 109
However, when applying the theory to data gathered from field measure-ments, only a finite signal window is available. Some examination of theproperties of time-limited signals and their Hilbert transforms is, therefore,warranted.
Consider the signal given in (4.23). Over an infinite window the Hilberttransform is well-defined and offers a closed-form expression, specifically(4.24). Because an infinite window is impractical to measure, suppose that a50 s interval of this signal is recorded, as given in Fig. 4.2.
The Hilbert transform calculated for the signal in Fig. 4.2 will dependheavily on the boundaries over which the transform is to be calculated. Ifonly the 0–50 s window is used, the result in Fig. 4.3 develops. The oscillationseen in the Hilbert transform appears well damped until around 45 s into therecord, when significant change occurs. The effect on the damping calculationis evidenced in Fig. 4.3. The mathematics developed earlier suggests thatFig. 4.3 should be a straight line of slope –1 pu/s. Whilst this is a reasonablerepresentation in the middle part of the record, distortion toward the endcould yield differing interpretations.
Fig. 4.2 Signal given in (4.23), truncated to the 0–50 s interval. Reproduced, with permission,from [4]. # 2008 IEEE
Fig. 4.3 Hilbert transform, calculated over 0–50 s window only, of signal in Fig. 4.2
110 T.J. Browne et al.
To alleviate the situation, the original signal can be zero padded prior to
finding the Hilbert transform and the analytic signal. Figures 4.5 and 4.6
demonstrate the improvement made to Figs. 4.3 and 4.4, respectively when
this action is taken.Note that Figs. 4.5 and 4.6 show significant variation before t=0, when the
input to the transform is zero. That is, the Hilbert transform operator is not
causal. This is apparent from the definition via the convolution, and is con-
firmed here by the calculation.
Fig. 4.4 Logarithm of analytic signal magnitude, calculated fromHilbert transform in Fig. 4.3
Fig. 4.5 Hilbert transform of zero-padded record
Fig. 4.6 Logarithm of analytic signal of zero-padded record
4 Practical Application of Hilbert Transform Techniques 111
The reason for the discrepancy can be traced to the Hilbert transform of arectangular pulse. A time-limited signal xm(t), as recovered frommeasurements,can be constructed as the product of an infinite signal xn(t) and a rectangularpulse ((t – a)/T) having duration T and midpoint t = a.
An alternative perspective can be obtained through consideration of thecalculation method. It has been established that the Hilbert transform can becalculated from the inverse FT of the positive frequency spectrum of the signal.Fourier methods assume periodicity, which is not a property held by the signalunder examination. Aliasing is therefore unavoidable. However, by zero pad-ding the signal on either side, the extent of the aliasing is reduced by making thefalse repetition less significant relative to the actual signal.
The domain on which the Hilbert transform of a signal is calculated can havea strong impact on the result. Distortion around the endpoints of the consideredwindow is most severe when only a short time is used for the calculation. Toreduce the impact of distortion around the endpoints of the signal section underexamination, the first 10% and last 10% of the time interval are discarded whenassessing modal parameters using least squares.
4.5 Multiple Mode Signals: Empirical Mode Decomposition
EMD [5], is a relatively new method of extracting components of a signal.Unlike linear transform techniques, EMD does not assume a constant fre-quency for each component. Instead, components exhibiting ‘‘fast’’ (in somesense) variations are isolated from components varying more slowly. Being atime–frequency analysis, the technique bears closer resemblance to wavelettransforms than to Fourier analysis. However, whereas wavelets offer closed-form expressions, EMD is defined by an algorithm and not a formal mathema-tical expression.
Traditional linear analysis methods, such as Fourier and Prony decomposi-tion, find constant parameters to fit a signal to a specific expression; in the caseof modal analysis, the expression takes the form
xðtÞ ¼X
i
Mie�it sinð!ditþ ’iÞ: (4:27)
By contrast, the components identified by EMD do not take a form which hasbeen previously assigned. EMD is described as a data-driven technique [5]:whereas Prony analysis assumes that a signal can be satisfactorily decomposedinto damped sinusoids, EMD makes no such assumption.
Huang [5] denoted the signal components identified by EMD as ‘‘intrinsicmode functions,’’ or IMFs, and proposed two defining properties of an IMF.The first requires that the difference between the number of zero crossings andthe number of maxima andminima must be –1, 0, or 1. The second requires that‘‘at any point, the mean value of the envelope defined by the local maxima and
112 T.J. Browne et al.
the envelope defined by the local minima is zero’’ [5]. An oscillatory mode of theform (4.2) meets – and exceeds – the definition of an IMF. In practice, a signalwith any significant DC component cannot be an IMF.
Huang [5] provides comprehensive instructions on application of the EMD.Many shortcomings in the implementation are examined. Only an overview isgiven here; for further details [5] is recommended. Available EMD software [28]provides additional mechanics.
EMD ‘‘sifts’’ out fast local variations from the signal content. To find thefirst, and fastest, IMF, the mean of the maximum and minimum envelopes isfound at each point and subtracted from the signal. Envelopes are presentlyfitted to the data points using a cubic spline algorithm; problems with the cubicspline fitting have been identified as a key deficiency of EMD [5]. The process offinding envelopes and subtracting their mean from the signal continues until theresulting signal meets the definition of an IMF. This IMF is subtracted from theoriginal signal and the process begins again, continuing until a residue whichnot possessing an envelope is found.
As an example of EMD application, Fig. 4.7 gives a bimodal test signaldefined as
gðtÞ ¼ e�t=10 sinð2p0:27tÞ þ e�t=8 sinð2p0:60tÞ; �15t51 (4:28)
and shows the decomposition into the three most significant IMFs. A secondmode has been added to (4.23) to derive g(t). As expected, the IMFs showing the
Fig. 4.7 Empirical mode decomposition of bimodal test signal given in (4.28). Residue issuppressed
4 Practical Application of Hilbert Transform Techniques 113
0.60 Hz (IMF1) and 0.27 Hz (IMF2) detail account for the bulk of the energyassociated with the signal. Several spurious IMFs are falsely identified, themostsignificant being IMF3 (displayed) which still has negligible energy relative toIMF1 and IMF2. The most likely cause of these spurious components is endeffects associated with the decomposition. The residue from the EMD is negli-gible and so not shown.
In this instance the additional mode does not present significant cause forconcern as the oscillations are well damped and of negligible energy. However,with other signals the potential exists for spurious unstable components toappear. In such a situation, control logic could incorrectly trigger separation ofthe network into islands. Alternate methods are therefore required to confirm theoutput from the EMD when used in a potential supervisory control scheme.
EMD suffers from a paradox: it analyzes time and frequency behavior of asignal simultaneously, but can only resolve one of these two characteristics bymaking assumptions on the other. In practical terms, EMD can only be used toassess the modal content of a signal numerically by assuming that the signal is infact composed of linear modes; that is, that (4.27) holds. If modal analysis is tobe undertaken, any online modal assessment must therefore operate on a slidingwindow.
Much of the discrepancy arises from a lack of clarity in the meaning of theword ‘‘mode.’’ In linear signal analysis, ‘‘mode’’ implies one of the components of(4.27). By contrast, in the EMD sense ‘‘mode’’ refers to a single componentisolated by the heuristic EMD process, which may bear no resemblance to amode in the linear sense [22, 28]. Neither interpretation of ‘‘mode’’ is necessarilywrong; rather, the intended meaning depends entirely on the context of thediscussion. Because the intended application is online numerical assessment ofmodal content, with a view toward determining the stability or otherwise of theunderlying system, the assumption of a linear system is implicit in the analysis.
In the field of power quality, EMD encounters difficulty in isolating sinu-soids closely spaced in frequency [29]. A solution which was employed involvesthe derivation of masking signals from the FFT of the signal. This solution,whilst having merit for sinusoids, is not appropriate for damped sinusoids:derivation of a masking signal is arguably more complicated than identificationof the mode itself and therefore provides no benefit.
4.6 Factors Affecting Performance of the Technique
4.6.1 Modal Separation
Section 4.5 has demonstrated the application of EMD in order to prepare foridentification of modal parameters in amultimodal signal. The test signal whichwas used in Section 4.5 contained two modes, separated in damped frequencyby a factor of more than two (0.27 and 0.60 Hz) and sharing the same
114 T.J. Browne et al.
attenuation factor of 0.1. It was noted that the EMD algorithm was able tosuccessfully separate the two components.
By contrast, in a signal containing twomodes closely spaced in frequency butwidely separated in damping, identification of the two components is notreadily achieved. As an example, the upper trace of Fig. 4.8 gives a signalderived from (4.23) with the addition of a second mode [4]:
hðtÞ ¼ e�t=10 sinð2p0:27tÞ þ e�t=1 sinð2p0:29tÞ (4:29)
The remaining traces in Fig. 4.8 give the IMFs and residue identified fromthe EMD. Application of the Hilbert analysis yields only the 0.27 Hz compo-nent from IMF1; the 0.29 Hz mode is not apparent. Altering the analysis tosuccessive windows of 10 s duration, or to the 0–50 s window examined earlier,provides no improvement. Application of Prony analysis to the same signal,however, successfully identifies both components of the signal, despite themuchlarger energy associated with the 0.27 Hz mode than the 0.29 Hz mode.
These results provide reason for caution in implementation of the EMD andHilbert analysis for any future real-time control application. In this instance, themissedmode is stable and is damped out quickly. However, it is quite conceivablethat the method could fail to identify a slightly unstable – but initially small –mode until after catastrophic system consequences have occurred.
For a further examination of EMD and Hilbert analysis applied to multi-modal signals, Fig. 4.9 gives a tie-line power flow signal from the classic two-areasystem [25, 30]. Over the full 50 s window, both Prony and Hilbert analysis
Fig. 4.8 Empirical mode decomposition of signal described by (4.29) over the 0–10 s interval
4 Practical Application of Hilbert Transform Techniques 115
identify a 0.21 Hz mode of damping ratio 0.044 and 0.043, respectively [4],comparing well with a damping ratio of 0.04 identified previously [30]. BothProny analysis and Sanchez-Gasca et al. [30] have identified a second harmonic ofthe 0.21 Hz as being present during the first half of the window in Fig. 4.9. TheHilbert analysis does not identify this mode the same way. Rather, a nonlinearityis identified.
4.6.2 Noise Tolerance
In order for the Hilbert analysis to be applicable for the intended stabilityassessment or correction applications, it must be capable of identifying modalparameters even in the presence of noise. In order to examine this capability,sample test signals are corrupted with noise to test the capabilities of theapproach. Figure 4.10 gives three signals corrupted by white Gaussian noise.
Fig. 4.9 Tie-line flow signal from two-area system [25, 30]. Reproduced, with permission,from [4]. # 2008 IEEE
Fig. 4.10 Three noise-corrupted test signals. The first two signals are based on the single-mode signal (0.27 Hz) of (4.29); the third is based on bimodal signal (0.27 and 0.60 Hz)described by (4.28). Reproduced in part, with permission, from [4]. # 2008 IEEE
116 T.J. Browne et al.
The upper andmiddle [4] traces add noise, with signal-to-noise ratios (SNRs) of10 and 2, respectively, to themonomodal signal of Fig. 4.2. The lower trace addsnoise, in an SNR of 5, to the bimodal signal of Fig. 4.7.
Considering the monomodal, SNR¼ 10, signal first, the Hilbert analysis is ableto identify the 0.27Hzmode, albeit as 0.269Hz, over both thewindow shown inFig.4.10 and the first 10 s intervalwithin thatwindow.However, the decay time constantidentified is 6.4 s for the 50 s interval and 9.8 s for the 10 s interval, comparedwith theexpected 10.0 s. Prony analysis provides superior estimates of 10.0 and 10.1 s,respectively [4]. The poor time constant estimated by the Hilbert technique overthe longer window is not overly surprising: toward the end of the window themodehas decayed substantially, enabling the noise to mask the signal more effectively.
The center trace in Fig. 4.10 is essentially a substantially noisier version of theupper trace. As such, it is to be expected that modal analysis is more difficultand results are less likely to be satisfactory. Such an expectation is verified byresults [4]: the Hilbert technique cannot identify the 0.27Hz mode in either thefull 50 s interval or in shorter 10 s intervals. Instead, a spurious 0.10Hz compo-nent dominates, and is coupled with different additional modes which varyaccording to the analysis window. Over the first 10 s interval, the dominantadditional mode is at 0.17Hz, whereas over the full 50 s interval a 0.38Hzcomponent is seen. Whilst the sum and the difference respectively of the twokey frequencies in the windows investigated approximate the actual 0.27Hzmode in the original signal, there is no apparent reason to assume that the0.27Hz signal can be derived from the two identified components ex post. Bycontrast, Prony analysis identifies the 0.27Hz mode directly [4], albeit with apoor time constant estimate (7.2 s) from the 0 s through 10 s interval and animproved assessment (9.3 s) over the 0–50 s interval. Components identified byboth techniques to represent the noise are not of interest for stability assessmentor control purposes; results are therefore not provided for these components.
The bimodal corrupted signal in the lower trace of Fig. 4.10 is repeated forthe 0–10 s interval in Fig. 4.11, which includes the three slowest IMFs identifiedby the EMD applied to this signal. Because the EMD isolates the fastestcomponents first, and noise can be thought of as a high-frequency process, itis logical that the modes of interest are the final IMFs developed. It should benoted that the scales on the vertical axes are not consistent throughout Fig. 4.11;therefore, IMFs 6 and 7 are of more interest than IMF 5, which appears toaccount only for the initial primary oscillation in the signal. Additionally, thecontribution of IMF to a signal reconstruction appears to distort IMF 6 fromits expected shape as a regular decaying sinusoid. By inspection, IMFs 6 and 7correspond, respectively, to the 0.60 and 0.27Hz signal components.
The distorting effect of IMF5 on IMF6 is made more apparent when modalparameter identification is undertaken. Figure 4.12 gives the natural logarithmof the analytic signal magnitudes for IMFs 6 and 7. Based on the theory given inSection 4.3.2, it would be expected that these two plots would approximatestraight lines. Whilst this is a fair description of the plot for IMF7, the same isnot true of the plot for IMF6. The residual from the least-squares estimate of
4 Practical Application of Hilbert Transform Techniques 117
attenuation factor for IMF6 will be significant; the potential exists for high
residuals to be used to indicate to any control algorithm the considerable
uncertainty in such an estimate.
Fig. 4.12 Corrupted test signal and logarithms of analytic signal magnitudes for two keyIMFs identified. Reproduced in part, with permission, from [4]. # 2008 IEEE
Fig. 4.11 Corrupted bimodal test signal and three most significant IMFs identified.Reproduced in part, with permission, from [4]. # 2008 IEEE
118 T.J. Browne et al.
Furthermore, although the Hilbert technique is able to correctly identify themodal frequencies from the IMF6 and IMF7 analytic function phase angles, theattenuation factors identified from Fig. 4.12 are reversed: a time constant of 8 s,and not 10 s, is attributed to the 0.27Hz mode, whereas a time constant of 10 s,and not 8 s, is attributed to the 0.60Hz mode. No explanation is readilyapparent for this phenomenon. However, the Hilbert technique is in thisinstance superior to the Prony analysis. Over the 10 s interval Prony analysiscannot identify the 0.27Hz mode and identifies the 0.60Hz mode of 8 s timeconstant as a 0.57Hz, 2.4 s component. Prony is able to identify both frequen-cies over the longer 50 s interval, but without a realistic damping assessment.
The investigation of Hilbert analysis as applied to the three corrupted testsignals of Fig. 4.10 has produced some interesting anomalies. The shorterwindows appear to favor the Hilbert technique over Prony analysis; bothmethods appear to suffer as the level of distortion increases. It could reasonablybe argued that some form of noise filtering of a signal, prior to application ofeither assessment method, would result in more accurate modal parameterestimation, and ultimately improved stability assessment and correction.
4.6.3 Changes in Underlying System Dynamics
A key property of a transmission network affecting modal identification is thecontinual change in network conditions, as evidenced by variations in loading,generation, and switching conditions. To obtain an indication that a mode changehas occurred, two options present themselves: direct use of themodal identificationtechnique, or use of external signals such as breaker statuses. The latter option is tobe preferred, as the external signals can be expected to be available.
The initial part of the test signal [4] given in Fig. 4.13 is constructed from thesingle-mode signal developed in Section 4.4. At t =16 s, however, the singlemode changes from 0.27Hz to 0.40Hz. The intent is to represent a change in thepower system as might occur in a switching event. Figure 4.14 isolates the
Fig. 4.13 Test signal with mode change at t¼ 16 s. Reproduced in part, with permission, from[4]. # 2008 IEEE
4 Practical Application of Hilbert Transform Techniques 119
interval 10–20 s, which includes the mode change, and includes an EMD and
Hilbert analysis over this interval. As the first IMF, IMF1, dominates and
accounts for the bulk of the energy in the signal, Hilbert analysis is confined
to this IMF.In particular, the slope of the analytic signal angle plot changes at the t =16 s
breakpoint. Because this slope represents the modal frequency, this effect is
expected. As the Hilbert analysis technique requires a least-squares estimate of
the slope of the analytic signal angle plot, the mode change will not become
apparent until the next 10 s intervalwhen the plot can be expected to track a single
straight line.Whilst techniques for identifying the change in slope are feasible, the
alternative – namely, relying on external signals to indicate a system change and
reset the modal analysis window – is a simpler and more elegant solution.
4.7 Application to Physical Signals
The foregoing discussion has applied EMD to synthetic signals, demonstrating
behavior that might be expected of power systemmeasurements. In this section,
recorded power system measurements are employed, in order to provide real-
world examples of how EMD and Hilbert analysis can be applied.The two signals in Fig. 4.15were recorded during a disturbance in theMexican
Interconnected System [31]. The upper graph gives a bus frequencymeasurement,
Fig. 4.14 Test signal of Fig. 4.13 (upper plot), along with empirical mode decomposition intotwo IMFs (center plot; residue suppressed) and the angle of the analytic signal for IMF1 (lowerplot). Reproduced in part, with permission, from [4]. # 2008 IEEE
120 T.J. Browne et al.
the lower a tie-line power flow measurement. Both measurements are represen-tative of the signal types which might be expected to accommodate Hilbertanalysis under disturbance conditions. The disturbance is evident from a cursoryinspection of the two graphs.
Immediate application of the Hilbert analysis to either of these signals,without consideration of windowing effects noted throughout this chapter, isunlikely to yield meaningful estimates of, for example, modal damping para-meters. Again it is stressed that the least-squares estimates in the present Hilbertanalysis are most reliable when the signal under test can be decomposed intoIMFs closely resembling damped sinusoids. The substantial nonlinear effectsapparent in the signals of Fig. 4.15 implies that such a condition is unlikely.
Figure 4.16 gives the bus frequency signal over three successively shorterintervals, making sinusoidal behavior more apparent. Hilbert analysis of theshortest interval, from 126 to 136 s, yields only one mode: 0.23Hz of unstabledamping ratio –0.003. Although Prony analysis identifies a stable (dampingratio 0.001) mode at the same frequency, the discrepancy is minor: the Pronyversion is sufficiently close to instability to warrant correction. Prony alsoidentifies three further modes in the frequency band corresponding to inter-area oscillations, but their amplitudes are sufficiently small as to be likelyartifacts of the decomposition rather than actual modes excited by the physicalsystem.
Figure 4.17 applies a similar analysis to the tie-line power flow signal. Again,neither Hilbert nor Prony analysis on the longest window given is logical: thereis no obvious way for the signal to be decomposed into damped sinusoids. For
Fig. 4.15 Bus frequency and tie-line flow signals recorded during a disturbance in theMexicantransmission network. Reproduced in part, with permission, from [4]. # 2008 IEEE
4 Practical Application of Hilbert Transform Techniques 121
Fig. 4.16 Recorded bus frequency signal over three successively shorter time intervals.Reproduced, with permission, from [4]. # 2008 IEEE
Fig. 4.17 Recorded tie-line power flow signal over three successively shorter time intervals.Reproduced in part, with permission, from [4]. # 2008 IEEE
122 T.J. Browne et al.
illustration purposes, however, such an analysis is attempted on the second
interval. Figure 4.18 gives the three component IMFs for this second
interval.The tie-line flow signal decomposition given in Fig. 4.18 highlights well the
difficulties faced by a future online modal identification technique. Of the three
IMFs given, none bears realistic resemblance to a traditional linear mode. The
theoretical basis for the Hilbert and Prony analysis techniques cannot apply to
this signal. Any attempt to identify modal parameters is not readily meaningful.
Instead, analysis in terms of amplitude-modulated sinusoids might be more
appropriate.However, for the same signal over the shorter interval of 126–136 s, as given
in the lower trace of Fig. 4.17, modal analysis appears to warrant application.
The Hilbert technique yields – in addition to a large pseudo-DC component – a
dominant mode at 0.201Hz with a (stable) time constant of 3.62 s and damping
ratio of 2.14. Two slightly unstable contributions, of sufficiently low relative
magnitude to be suspect, are also identified, at 0.16Hz and at 0.37Hz. Prony
analysis provides very different results: a dominant lightly damped (time con-
stant 67 s) 0.23Hz component, a lightly damped (time constant 3.8 s) 0.63Hz
component, and an unstable 0.41Hz component.For this shorter 10 s long interval, neither set of results is necessarily wrong;
rather, each reflects a different objective in fitting the recorded data to a
theoretical ideal. Further, the substantial changes which occur in the behavior
Fig. 4.18 Decomposition of tie-line flow test signal into (dimensionless) three most significantintrinsic mode functions over the 120–180 s interval. Reproduced, with permission, from [4].# 2008 IEEE
4 Practical Application of Hilbert Transform Techniques 123
of the signal soon after this short interval render the analysis immaterial:
changes in the system mean that modes excited during the interval between
126 and 136 s are not identical to the modes exhibited a short time later.
4.8 Conclusions
For assessment of modal parameters, the Hilbert analysis offers an alternative
approach to the more standard Prony techniques. Practical considerations such
as window selection and zero padding have been shown to influence the accu-
racy of results significantly.Considerable differences exist in the principles underpinning the twomethods.
Prony analysis is fundamentally directed toward extracting modal parameters
from stationary signals recorded in linear systems. The Hilbert method, in
principle, is applicable to more general conditions and forms a time–frequency
analysis technique. However, seeking numerical values for modal parameters
requires that stationarity is assumed to be an implicit signal property, removing
the flexibility introduced by the Hilbert method.Scope exists for improvement in both the EMD algorithm and the Hilbert
transform calculation. Nevertheless, the Hilbert analysis has demonstrated the
ability to identify modal parameters in many of the test signals examined,
especially where noise conditions are not excessive and modal frequencies are
well separated.
References
1. I. Kamwa, R. Grondin, Y. Hebert, ‘‘Wide-area measurement based stabilizing control oflarge power systems – a decentralized/hierarchical approach,’’ IEEE Trans. Power Syst.,vol. 16, no. 1, pp. 136–153, Feb. 2001.
2. I. Kamwa, L. Gerin-Lajoie, G. Trudel, ‘‘Multi-loop power system stabilizers using wide-area synchronous phasor measurements,’’ Proc. American Control Conference, vol. 5, pp.2963–2967, June 1998.
3. J. F. Hauer, ‘‘Application of Prony analysis to the determination of modal content andequivalent models for measured power system response,’’ IEEE Trans. Power Syst., vol. 6,no. 3, pp. 1062–1068, Aug. 1991.
4. T. J. Browne, V. Vittal, G. T. Heydt, A. R. Messina, ‘‘A comparative assessment of twotechniques for modal identification from power system measurements,’’ IEEE Trans.Power Syst., vol. 23, no. 3, pp. 1408–1415, Aug. 2008.
5. N. E. Huang, Z. She, S. R. Long, M. C.Wu, S. S. Shih, Q. Zheng, N.-C. Yen, C. C. Tung, H.H. Liu, ‘‘The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis,’’ Proc. Royal Society of London, vol. 454, pp. 903–995, 1998.
6. G. Rogers, Power System Oscillations. Boston: Kluwer, 2000.7. J. Hauer, D. Trudnowski, G. Rogers, B. Mittelstadt, W. Litzenberger, J. Johnson, ‘‘Keep-
ing an eye on power system dynamics,’’ IEEE Comput. Appl. Power, vol. 10, no. 4, pp.50–54, Oct. 1997.
124 T.J. Browne et al.
8. C. E. Grund, J. J. Paserba, J. F. Hauer, S. Nilsson, ‘‘Comparison of Prony and eigen-analysis for power system control design,’’ IEEE Trans. Power Syst., vol. 8, no. 3, pp.964–971, Aug. 1991.
9. BPA/PNNL, Dynamic Systems Identification (DSI) Toolbox, [Online], available: ftp://ftp.bpa.gov/pub/WAMS_Information/.
10. D. Sundararajan, The Discrete Fourier Transform: Theory, Algorithms and Applications.New York: World Scientific Publishing Company, 2001.
11. E. Chu, Discrete and Continuous Fourier Transforms: Analysis, Applications, and FastAlgorithms, Boca Raton, FL: CRC Press, Taylor and Francis Publishing Co., 2008.
12. C. D.McGillem, G. R. Cooper,Continuous and Discrete Signal and System Analysis, NewYork: Prentice Hall, 1995.
13. G. R. Cooper, C. D. McGillem, Methods of Signal and System Analysis, New York:Prentice Hall, 1997.
14. A. V. Oppenheim, A. S. Willsky, S. Hamid, Signals and Systems, New York: Prentice Hall,1996.
15. P. O’Shea, ‘‘The use of sliding spectral windows for parameters estimation in powersystem disturbance monitoring,’’ IEEE Trans. Power Syst., vol. 15, no. 4, pp.1261–1267, Nov. 2000.
16. K. P. Poon, K. C. Lee, ‘‘Analysis of transient stability swings in large interconnectedpower systems by Fourier transformation,’’ IEEE Trans. Power Syst., vol. 3, no. 4, pp.1573–1581, Nov. 2007.
17. S. L. Hahn, Hilbert Transforms in Signal Processing. Boston: Artech House, 1996.18. R.N. Bracewell,The Fourier transform and its applications. 3rd ed. Boston:McGraw-Hill,
2000.19. E. Bedrosian, ‘‘A product theorem for Hilbert transforms,’’ Proc. IEEE, pp. 868–869,
May 1963.20. G. D. Cain, ‘‘Hilbert transform relations for products,’’ Proc. IEEE, pp. 673–674, May
1973.21. J. W. Brown, R. V. Churchill, Complex Variables and Applications. 6th ed. New York:
McGraw-Hill, 1996.22. G. Rilling, P. Flandrin, P. Goncales, ‘‘On empirical mode decomposition and its algo-
rithms,’’ IEEE-EURASIP Workshop on Nonlinear Signal and Image Processing (NSIP-03), Grado (Italy), Jun. 8–11, 2003.
23. L. Cohen, Time-Frequency Analysis. Englewood Cliffs, NJ: Prentice-Hall, 1995.24. MathWorks, Matlab, http://www.mathworks.com/products/matlab25. P. Kundur, Power System Stability and Control. Palo Alto: EPRI, 1994.26. B. Boashash, ‘‘Estimating and interpreting the instantaneous frequency of a signal – Part
1: Fundamentals,’’ Proc. IEEE, vol. 80, no. 4, pp. 520–538, Apr. 1992.27. R. Ceravolo, ‘‘Use of instantaneous estimators for the evaluation of structural damping,’’
J. Sound and Vibration, vol. 274, no. 1–2, pp. 385–401, Jul. 6, 2004.28. P. Flandrin, Empirical Mode Decomposition, [Online] available: http://perso.ens-lyon.fr/
patrick.flandrin/emd.html, Mar. 2007.29. N. Senroy, S. Suryanarayanan, P. F. Ribeiro, ‘‘An improved Hilbert-Huang method for
analysis of time-varying waveforms in power quality,’’ IEEE Trans. Power Syst., vol. 22,no. 4, pp. 1843–1850, Nov. 2007.
30. J. J. Sanchez-Gasca, V. Vittal, M. J. Gibbard, A. R. Messina, D. J. Vowles, S. Liu,U. D. Annakkage, ‘‘Inclusion of higher order terms for small-signal (modal) analysis:Committee report – Task force on assessing the need to include higher order termsfor small-signal (modal) analysis,’’ IEEE Trans. Power Syst., vol. 20, no. 4, pp.1886–1901, Nov. 2005.
31. A. R. Messina, V. Vittal, D. Ruiz-Vega, G. Enrıquez-Harper, ‘‘Interpretation and visua-lization of wide-area PMU measurements using Hilbert analysis,’’ IEEE Trans. PowerSyst., vol. 21, no. 4, pp. 1763–1771, Nov. 2006.
4 Practical Application of Hilbert Transform Techniques 125
Chapter 5
A Real-Time Wide-Area Controller
for Mitigating Small-Signal Instability
Jaime Quintero and Vaithianathan (Mani) Venkatasubramanian
Abstract This work proposes a real-time centralized controller for addressingsmall-signal instability-related events in large electric power systems. The pro-posed system ismeant to be a safety net type control strategy that will detect andmitigate small-signal stability phenomena as they emerge in the system. Speci-fically, it will use wide-area monitoring schemes to identify the emergence ofgrowing or undamped oscillations related to inter-area and/or local modes.
The damping levels of the associated inter-area and local oscillatory modeswill be estimated by analyzing predefined sets of signals using multi-Pronymethod. Rules are developed for increasingmulti-Pronymethod’s observabilityand dependability. These rules are applied to simulated signals, but also to realnoisy measurements.
Rules for operating the SVC (static VAR compensator) controls in thedamping enhancement mode and for the application of the multi-Prony algo-rithm on detecting the onset of the oscillations are proposed and tested in a two-area power system and in large-scale simulation example. The controller isshown to be effective on a validated western American large-scale power systemmodel of the August 10, 1996 blackout event.
5.1 Introduction
In modern day power systems, the power flows across distant portions of thetransmission network have been growing steadily to accommodate growingconsumer demands. Moreover, owing to deregulation, the power transfershave also become somewhat unpredictable as dictated by market price fluctua-tions. As a result, the system operation can find itself close to or outside thesecure operating limits under severe contingencies. Recent occurrences of large-scale blackouts all over the world reinforce the significance of developing safety
J. Quintero (*)Faculty of Engineering, Universidad Autonoma de Occidente, Cali-Valle, Colombiae-mail: [email protected]
A.R. Messina, Inter-area Oscillations in Power Systems,Power Electronics and Power Systems, DOI 10.1007/978-0-387-89530-7_5,� Springer ScienceþBusiness Media, LLC 2009
127
net type control mechanisms, which are specifically designed to handle suchunforeseen operating conditions [1, 2, 3].
Many efforts have beenmade in order to increase the damping of rotor modesunder a wide range of operating conditions, allowing larger steady-state powertransmissions. Classical lead–lag and bang–bang compensators have beendesigned as auxiliary controls on static VAR compensators (SVCs) [4, 5, 6].Also, robust tuning algorithms has been used like the linear quadratic Gaussianfor SVCs in [7] and for thyristor-controlled series capacitors in [8], or H1 and m-synthesis designs in [9, 10], respectively. Because the interactions between unco-ordinated damping controllers can be destabilizing, robust MIMO (multiple-input multiple-output) algorithms have also been developed for coordination ofmultiple power system stabilizers (PSSs) [11, 12, 13], or multiple flexible ACtransmission systems (FACTS) [14], or various PSS units and FACTS together asin [15, 16]. With the recent implementation of synchronized phasor measure-ments, wide-area measurement-based coordinated control techniques have alsobeen used [17]. Specific comparative studies may be found in [18, 19].
Although, all of these techniques achieve very good results on improving themode shape of the system and many achieve robust stability and robust per-formance over a wide range of operating conditions, none of them is specificallydesigned for the emergency condition when the systemmay become small signalunstable. Therefore, the main purpose of this work is to design a reliableemergency control strategy when the system despite all these well-designedcontrols gets close to or goes into small-signal instability.
Themain difference in our approach from the previous research works is thatthe controller as proposed here will initiate targeted damping control actionsonly when it detects the emergence of small-signal instability in the powersystem. Therefore, this controller is aimed at being a safety net type automaticsupervisory mechanism which is aimed toward handling growing or sustainedoscillations from poor damping of inter-area modes under unforeseen highlystressed operating conditions. It is meant to buy the system operators some timeto react, by damping out the oscillations in the interim, while the operators canpossibly initiate other corrective actions such as reducing the tie-line flows and/or initiate load shedding to relieve the system stress.
Our controller is motivated toward preventing oscillatory instability events suchas the August 10, 1996 western blackout that was caused by the negative dampingof the 0.25Hzwestern inter-areamode (Fig. 5.1). Such oscillatory instability eventstypically take a minimum of four to five swings of growing oscillations before theoscillations become large enough to lead to system separation. Therefore, for suchsmall-signal stability-related instability events, there does exist sufficient time for anautomatic controller such as the one postulated inFig. 5.2 to detect the event and totake corrective actions before the oscillations become critical.
On the other hand, Prony analysis [20], has been successfully applied foroff-line signal and model identification in power systems [21, 22, 23], and awide variety of fields [24, 25, 26]. It has been shown that comparing withFourier transform-based techniques, Prony’s method is more powerful when
128 J. Quintero and V. (Mani) Venkatasubramanian
it works [27]. Issues such as system’s nonlinearities [28], power system’s high-
order model [29, 30], noise, low signal-to-noise ratios (SNRs), and defining the
time window or data length [31, 32] have limited its use in real-time applications.
However, the problem suggested by [29] was already solved by one upgrade in
[33], and the high-frequency modes [30] are not an issue in small-signal stability
studies, as the low-frequency nature of inter-area and local modes. Reference
[32] analyzes the issues of sample size selection, data length selection, and noise
in Prony method. However, as seen later in this chapter with simple rules and a
SVC SVC
SVC LOAD
LOAD
LOAD
LOAD
CENTRAL CONTROL
P
PP
Fig. 5.2 Real-time small-signal stability centralized control strategy
Fig. 5.1 Growing oscillations on the California–Oregon tie-lines during the August 10, 1996blackout
5 A Real-Time Wide-Area Controller for Small-Signal Instability 129
smoothing filter, these issues can be strongly reduced in power system’s small-signal stability studies.
In [28], the authors apply the Hilbert spectral analysis obtaining the instan-taneous frequency and damping. In [34], the authors use matrix pencil methodand Hankel total least-squares method to determine also the frequency anddamping of power system modes. They results are comparable to Pronymethod. Moreover, in [34] the authors use the same principles presented inthis work to increase the reliability of the oscillation monitoring system (OMS).
The controller framework is summarized in Fig. 5.2. Rules for applying theright power swing phase compensation by strategically located SVCs are devel-oped. Also, they are demonstrated in the Kundur’s two-area power system [35]and briefly in the Western Systems Coordinating Council (WSCC) large-scalepower system. Practical rules for applying multi-Prony analysis successfully onreal-time basis for detection of low damped oscillations are developed. Off-linerules for selection and use of the appropriate signals for improving the accu-racy, the observability, and the reliability of multi-Prony analysis are obtainedthrough simulations in the two-area power system. Later, these rules are shownwith real noisy data and final online rules are proposed and proved with theWSCC power system for a validated case of the August 10, 1996 blackout [36].
The proposed approach is based on previous works [37, 33, 38] and tested withreal data and a validated simulation case [36]. Reference [37] introduces theapplicability and usefulness of Prony in power system, while [38] improves sub-stantially the accuracy of Prony’s estimations. Also, the idea of grouping signals bydominant modes to increase the observability of certain modes was first suggestedin [33]. We have taken these previous results and developed some more rules toimprove the accuracy, the reliability, and the observability of the multi-Pronyanalysis for the purpose of a real-time controller. The framework has beenextended into an OMS at Washington State University [34] and a prototypeversion ofOMShas been recently implemented at Tennessee ValleyAuthority [39].
Finally, the main objective of this work is on designing a practical real-timewide-area controller, independent of system configuration, to work as an auto-matic emergency controller to improve dynamic security in the small-signalsense. The main contributions are as follows:
� Targets small-signal instability.� Improves Prony’s method accuracy, observability, and reliability.� Analyzes the effects of full SVC PSDC (power swing damping control)
location and compensation on inter-area modes.� Makes use of available resources.
5.2 The Controller
The centralized control strategy introduced in this work (Fig. 5.2) will con-stantly monitor the inter-area modes as well as the local modes presented in thepower system. Data from predetermined signals with relevant content of local
130 J. Quintero and V. (Mani) Venkatasubramanian
and inter-area modes are sent from phasor measurement units (PMUs) atdifferent control areas and intertie lines to the central control unit via suitablecommunication channels. The central controller analyzes these signals in theform of different sets by using multi-Prony’s method grouped by dominantmodes, in order to estimate the power system’s mode shape. If there is one ormore underdamped inter-area modes, the central unit will issue triggers topreviously identified SVCs located at the affected areas and or along the intertielines. The SVCs will switch from the voltage regulation control mode to the fullPSDCmode until safe damping ratio values �i and safe operating conditions arereached again. Constant monitoring will assure that other local nor inter-areamodes are not weakened during the control action.
The two-area power system presented by Kundur [35] will be used in thiswork (Fig. 5.3). Although this is a small test system, the system has beenproven to be of great value for the study of inter-area oscillations in real powersystems.
5.3 The Central Control Unit
Selected signal measurements, received from dedicated PMUs, grouped bydominant modes are processed by the central control unit using multi-Pronyanalysis [38].
The data suitable for Prony analysis are the ones collected after a small- ormedium-scale disturbance. The initial period of fast and very nonlinear oscilla-tions just after the disturbance will give bad estimations, as Prony’s method is alinear approach.We present techniques for distinguishing between useful versusunreliable Prony estimations of modal data, from real-time measurement data.Also, we show that two or three swings of data are sufficient to obtain a fairapproximation of the mode shape.
P
G1
G2
G3
G4
1
2
3
4
5 6 7 8 9 10 11
Area 1 Area 2
SVC7SVC9C7
C9
SVC8
C8
Fig. 5.3 Two-area power system
5 A Real-Time Wide-Area Controller for Small-Signal Instability 131
It has been demonstrated in [38] that multi-Prony analysis of various signals ismore robust than single Prony analysis. Also, it was noted in [33] that the Pronyaccuracy is improved by having different signals with the same dominant modeanalyzed simultaneously. Moreover, our tests have shown that the observability onthe estimation of inter-area and localmodes is increased by analyzing different sets ofsignals grouped by dominant modes. In extension, for the common case of havingtwoormore inter-area or local areamodeswith almost the same frequency, it is likelythat the Prony’s methodwill have difficulties on differentiating these modes if we tryto extract the information froma single signal or a single set of signals, as shownnext.We have addressed this problem by analyzing the signals in different sets, previouslycharacterized by dominantmodes. That is, the signals with the same dominantmodeare grouped together, and we form a few such different sets to ensure redundancy.
Moreover, it is not unusual in Prony’s analysis to make bad estimations froma signal set at any time. Crosschecking results from different sets of data willmake sure that our estimations are reliable and will provide a certain redun-dancy level. In case of bad estimations, the analysis will be repeated over thosesets using another least-squares solution. If the results are still not consistent,the total estimations would be disregarded, and a complete new estimationprocess should be applied over the next time window.
Based on the estimated mode (eigenvalue) parameters, if the central con-troller determines that one or more inter-area modes are under damped ornegative damped, it will send triggers to corresponding SVCs.
Prony’s method, or multi-Prony’s method, estimates parameters of modesassociated with a signal, or certain set of signals. These are mode frequency fi,amplitude Ai, phase �i, and damping ratio �i. The Prony’s method which wasoriginated in a very early century [20, 37], calculates signal estimations byapproximating in the least-squares sense to a certain set of equally sampleddiscrete data, a linear exponential function of the form
yðtÞ ¼Xm
i¼1Ai expð�itÞ cosð2pfitþ �iÞ for t � 0: (5:1)
This real equation may be written in a complex exponential form as
yðtÞ ¼Xm
i¼1Bi expðlitÞ þ B�i exp l�i t
� �for t � 0 (5:2)
where m is the number of modes, Bi ¼ Ai=2ð Þ expðj�iÞ, li ¼ �i þ j!i, and*
indicates complex conjugate. Moreover, in a simplified manner (5.2) could beexpressed as
yðtÞ ¼Xp
i¼1Bi expðlitÞ for t � 0 (5:3)
132 J. Quintero and V. (Mani) Venkatasubramanian
where p is the number of the estimated eigenvalues. Without loss of general-ization, consider that for Eqs. (5.2) or (5.3) the Bi’s and the li’s are distinct. LetyðtÞ be N samples evenly spaced by �t such that
yðtkÞ ¼ yðkÞ for k ¼ 0; 1; . . . ;N� 1:
The Prony’s estimated output signal at time tk ¼ k will be
yðkÞ ¼Xp
i¼1Bi expðlik�tÞ for k ¼ 0; 1; . . . ;N� 1:
For convenience, define zi ¼ expðli�tÞ, then
yðkÞ ¼Xp
i¼1Biz
ki for k ¼ 0; 1; . . . ;N� 1: (5:4)
As can be seen, the objective with Prony’s method is to find the values of Bi
and Zi that produce
yðkÞ ¼ yðkÞ for all k: (5:5)
Multi-Prony analysis [38] is a vector–matrix extension of Prony analysis,considering multiple outputs at the same time.
5.3.1 Setting Up the Central Unit – Off-line Rules
Here, our control strategy is applied to the two-area power system as shown inFig. 5.3. A series of empirical rules for selecting the power system variables to bemonitored by the central control unit are developed and are explained throughnumerical simulations.
5.3.1.1 Defining the Time Window
Because Prony analysis is a linear approximation, there is a risk of not gettinggood estimations during the first strong nonlinear oscillations just after thedisturbance has been cleared. When this happens, the estimations obtainedfrom the different signals sets will not agree and the estimation will be disre-garded. The multy-Prony method will then be applied to the next time window.
For the two-area system, Table 5.1 shows relative energy Eri values of inter-area and local modes for two intertie line signals at different time windows. Itcan be seen that the most dominant mode, the inter-area mode, increasesits energy about three times with respect to the second most dominant modewhen the time windowmoves from 4.1–8.1 to 5.0–9.0 s. Also, simulations like inFig. 5.4 show how the SNR of these two most dominant modes is considerably
5 A Real-Time Wide-Area Controller for Small-Signal Instability 133
increased for the same time window change. As a consequence of these results,
in this case a time window after the fifth second will allow more accurate
estimations than a time window just after the disturbance.On the other hand, good parameter estimations were achieved by multi-
Prony analysis with only two complete swings of data. Table 5.1 and Fig. 5.4
show no significant advantages on using a longer time frame in the two-area
system. However, in a real system, due to more complex dynamics, three to four
cycles are a safer option.
5.3.1.2 Grouping Signals by Dominant Modes
Prony or multi-Prony analysis of signals with dominant modes that have very
close frequencies will likely estimates only one equivalent mode. This could be
seen in Table 5.1 on every single Prony estimation for the 5.1Hz local area 1 and
area 2 modes in the two-area power system. Inter-area active power flow P78
and voltage magnitude at Bus 8 V8 are signals that have inter-area and local
Table 5.1 Prony’s Ringdown analysis
Symbol Signal Mode fi �i Eri
Time window: 4.1–8.1 sec
P78 Interarea activepower flow
InterareaLocal Area 1 or 2
0.5091.088
0.00000.0835
1.00000.0189
V8 Bus 8 voltagemagnitude
InterareaLocal Area 1 or 2
0.5081.110
�0.00480.0685
1.00000.0069
Time window: 5–9 s
P78 Interarea activepower flow
InterareaLocal Area 1 or 2
0.5081.088
0.00290.0822
1.00000.0066
V8 Bus 8 voltagemagnitude
InterareaLocal Area 1 or 2
0.5051.106
0.00190.0677
1.00000.0028
Time window: 5–13 s
P78 Interarea activepower flow
InterareaLocal Area 1 or 2
0.5081.087
0.00020.0827
1.00000.0035
V8 Bus 8 voltagemagnitude
InterareaLocal Area 1 or 2
0.5091.107
0.00150.0697
1.00000.0015
The two–area power system configuration is as shown in Fig. 5.3, with an SVC at Bus 8operating in voltage regulation form, and with P78 ¼ 330 MW.Dynamic simulation was performed using Extended Transient Midterm Stability Program(ETMSP) from the Electric Power Research Institute (EPRI).One line from Bus 8 to Bus 9 is removed at 4.00 s and reconnected at 4.08 s. Samplingfrequency is 62.5 Hz.Prony signal estimation was calculated using the Ringdown GUI program from BPA/PNNLDynamic System Identification (DSI) Toolbox. Signal mean values were removed.Mode parameters calculated using the Multi–Area Small-Signal Stability Program (MASS)from EPRI are as follows interarea mode, fi ¼ 0.506 and �i ¼ �0.0007; local area 1 Mode,fi ¼ 1.078, and �i ¼ 0.0638; and local area 2 mode, fi ¼ 1.111, and �i ¼ 0.0574.
134 J. Quintero and V. (Mani) Venkatasubramanian
area mode content of the two local modes. However, Prony analysis is not able
to distinguish between the two local modes.Table 5.2 presents a multi-Prony analysis of two different groups of signals.
Both groups include signals that have different dominant modes. In the first
group, the generator 2 rotor angle �2 has local area 1 mode and inter-area mode
as dominant modes, while the generator 4 rotor angle �4 has local area 2 mode
and inter-areamode as dominantmodes. In the second group, line 6 to 7 current
magnitude I67 has local area 1 mode and inter-area mode as dominant modes,
and, line 9 to 10 current magnitude I910 has local area 2 mode and inter-area
mode as dominant modes. However, multi-Prony analysis at the first two rows
fails on the estimation of the two local modes and gives an equivalent estimation
of only one mode.The estimation problem of modes with very close frequencies may be
addressed by analyzing signals that are predominantly affected by only one
of these modes. For this case, we may see in Table 5.3 how analyzing
signals grouped by local area modes and by inter-area modes give a clear
estimation of each mode. Therefore, observability is improved by classify-
ing signals by dominant modes in groups and analyzing them
independently.
4.1 4.6 5.1 5.6 6.1 6.6 7.1 7.6–4
–3
–2
–1
0
1
2Bus 8 Voltage Magnitude Swings (pu)
Time (sec)
Measured DataProny's Approximationsnr = 2.0
5 5.5 6 6.5 7 7.5 8 8.5–1.5
–1
–0.5
0
0.5
1
1.5
2× 10
–3× 10
–3
× 10–3
Bus 8 Voltage Magnitude Swings (pu)
Time (sec)
Measured DataProny's Approximationsnr = 8.3
5 6 7 8 9 10 11 12–1.5
–1
–0.5
0
0.5
1
1.5
2Bus 8 Voltage Magnitude Swings (pu)
Time (sec)
Measured DataProny's Approximationsnr = 10.3
Fig. 5.4 Prony’s Bus 8 voltage magnitude signal approximation
5 A Real-Time Wide-Area Controller for Small-Signal Instability 135
Table 5.2 Multi–Prony analysis – Case 1 – Signals with different dominant modes combined
Signals Least-square solution Mode fi �i
�8, �2, and �4 Singular value Interarea 0.521 0.0164
decomposition Local area 1 or 2 1.104 0.0710
I78, I67, and I910 Singular value Interarea 0.564 �0.1920decomposition Local area 1 or 2 1.079 �0.0560
I78, I67, and I910 Total least squares Interarea 0.525 0.0143
Local area 1 1.066 0.0332
Local area 2 1.302 0.0948
The two–area power system configuration is as shown in Fig. 5.3, without SVCs, with onlytwo lines from Bus 8 to Bus 9, and with P78 ¼ 400 MW.Dynamic simulation was performed using Extended Transient Midterm Stability Program(ETMSP) from the Electric Power Research Institute (EPRI).One line from Bus 8 to Bus 9 is removed at 4.00 s and reconnected at 4.08 s. Sample data aretaken from 5.00 s to 9.00 s. Sampling frequency is 62.5 Hz.Prony andmulti–Prony signal estimation were calculated using theRingdownGUI program fromBPA/PNNL Dynamic System Identification (DSI) Toolbox. Signal mean values were removed.
Table 5.3 Prony mode parameter estimation – Case 1 – signals grouped by dominant modes
Symbol Signal Ai fi �i Ari
Interarea modea 0.523 0.0111
I78 Interarea current magnitude 0.224 0.524 0.143 1.000
P78 Interarea active power flow 0.156 0.522 0.0177 1.000
Q78 Interarea reactive power flow 0.061 0.520 0.0183 1.000
V6 Bus 6 voltage magnitude 0.015 0.524 0.0176 1.000
V8 Bus 8 voltage magnitude 0.018 0.521 0.0167 1.000
V10 Bus 10 voltage magnitude 0.009 0.519 0.0177 1.000
�10 Bus 10 voltage angle 2.461 0.522 0.0178 0.761
�8 Bus 8 voltage angle 1.522 0.520 0.0181 0.652
�6 Bus 6 voltage angle 0.076 0.559 �0.0143 0.068
Local area 1 modea 1.090 0.0694
�1 Generator 1 rotor angle 0.163 1.084 0.1113 0.145
I67 Line 6 to 7 current magnitude 0.009 1.081 0.0558 0.033
�6 Bus 6 voltage angle 0.033 1.074 0.0582 0.030
�2 Generator 2 rotor angle 0.170 1.083 0.0753 0.031
V1 Bus 1 voltage magnitude 2e–4 1.041 0.0529 0.027
V2 Bus 2 voltage magnitude 1e–4 1.100 0.0062 0.008
V6 Bus 6 voltage magnitude 1e–4 1.075 �0.0032 0.007
Local area 2 modea 1.118 0.0704
I910 Line 9 to 10 current magnitude 0.011 1.060 0.0356 0.071
�4 Generator 4 rotor angle 0.153 1.120 0.0718 0.056
V3 Bus 3 voltage magnitude 2e–4 1.207 0.0847 0.029
�3 Generator 3 rotor angle 0.080 1.142 0.0672 0.028
V10 Bus 10 voltage magnitude 2e–4 1.249 0.1213 0.024
�10 Bus 10 voltage angle 0.029 1.143 0.0748 0.009
V4 Bus 4 voltage magnitude 3e–4 1.226 0.1551 0.004aValues calculated using theMulti–Area Small-Signal Stability Program (MASS) from EPRI.Same signal measurements and programs as Table 5.2
136 J. Quintero and V. (Mani) Venkatasubramanian
5.3.1.3 Using Mode Content: Group Using Ai and Ari
It is clear from Table 5.3 that signals with the greatest dominant mode contentare the more likely to give the best estimations. However, even with thesesignals, it is still possible to have bad estimations when using Prony analysis.It has been demonstrated in [38] that using multi-Prony will increase theaccuracy of the estimations, but still due to the nonlinear nature of powersystem signals we may have bad estimations. Our goal with this rule is makingas many subgroups as we want in order to increase the redundancy level in ourestimations divided by dominant modes.
When using Prony’s method, amplitude Ai is a good approximation of modei content in the processed signal. However, signals have very different valueranges even in per unit basis. Therefore, it is also important considering therelative amplitude Ari when judging the real mode content on a signal. In Table5.3, signals like voltage magnitudes V1,V2,V4, andV6, and bus voltage angle �6are between the signals with the lowest Ai and Ari values at each group and arethe ones with the largest estimation differences with respect to the dampingratio calculated with the Multi-Area Small-Signal Stability (MASS) Program.
Combining signals that have high Ai and Ari values with fewer signals thathave lowAi orAri values may improve the general estimation, as it is the case forthe group formed by �10, �8, and �6, and the group formed by �2, I67, and V6. Inmost cases, relative amplitude Ari is increased when using multi-Prony analysisof signals with similar dominant modes, as seen in Table 5.4. On the other hand,multi-Prony’s method for analyzing signal sets where all signals have very lowAi and Ari values does not improve the estimation, as it is the case for the groupof voltage magnitudes V1, V2, and V6 in Table 5.4.
5.3.1.4 Validating the Groups
With this rule we want to select the final subgroups that are going to be used formonitoring the mode shape of our power system in real time. While we want tohave a good redundancy level on the estimation of each mode, we also want tomake sure that the chosen subgroups has the ability of providing good estima-tions of the excited modes when a small or medium disturbance occurs in thepower system.
Because Prony or multi-Prony analysis is a linear approach applied to non-linear signals, we should expect certain margin of error in estimations fromdifferent signals during the same time window. Therefore, an acceptable errormargin � should be defined in order to admit or not the estimations. In Table5.4, for example, an accuracy level � of�0.01 with respect to the mode dampingratio values found with the MASS Programmay be selected. Consequently, theestimation from the group of voltage magnitudes V1, V2, and V6 will beconsidered as a bad estimation and will be eliminated as a subgroup.
Cross-checking damping ratio values obtained from subgroups selectedaccording with rules 1, 2, and 3, is another way for determining the finalsubgroups if no actual mode value is known. According to our results in the
5 A Real-Time Wide-Area Controller for Small-Signal Instability 137
two-area and the WSCC power systems, a difference of �2% damping ratio
value between the minimum and the maximum estimation for one damping ratio
may be acceptable. Bad groups, or groups with signals with weak contain of
certain mode will be discarded as they are detected with the cross-checking
process.Nevertheless, it is important to keep in mind that good data that fulfill the
best conditions of the above rules may give bad results. This can be seen in
Table 5.2 for the second group of signals I78, I67, and I910, where in spite of being
part of the signals with higher amplitudes and relative amplitudes at each group
in Table 5.3, we do get completely unacceptable results at the first attempt of
solving the least-squares problem. The results have shown that changing the
least-squares solution method originally used for the bad estimation will
Table 5.4 Multi–Prony mode parameter estimation – Case 1 – signals grouped by dominantmodes
Symbol Signal Ai fi �i Ari
Interarea modea 0.523 0.0111
V6, 0.015 1.000
V8, 0.018 0.522 0.0199 1.000
and V10 0.010 1.000
�10, 2.488 0.202
�8, 1.534 0.522 0.0191 0.124
and p78 0.152 1.000
�10, 2.454 1.000
�8, 1.490 0.525 0.0189 0.815
and �6 0.175 0.095
Local area 1 mode a 1.090 0.0694
V1, 9e–5 0.009
V2 7e–5 1.058 �0.0107 0.005
and V6 9e–5 0.006
�2, 0.173 0.117
I 67, 0.011 1.082 0.0765 0.040
and �6 0.040 0.028
�2, 0.173 0.145
I67, 0.011 1.083 0.0765 0.041
and V6 2e–4 0.012
Local area 2 mode a 1.118 0.0704
�4, 0.153 1.120 0.0718 0.056
and V3 3e–4 0.046
I910, 0.012 0.085
�3, 0.177 1.077 0.1048 0.057
and V10 7e–4 0.074
I910, 0.018 0.122
�3, 0.086 1.140 0.0705 0.030
and �10 0.021 0.009
Same signal measurements and programs as Table 5.2.
138 J. Quintero and V. (Mani) Venkatasubramanian
contribute to solve the problem for these cases as shown in Table 5.2, and later
in the WSCC application example.
5.3.2 Monitoring and Control – Online Rules
Here, measured data from the WSCC power system will be used in order to
illustrate the online rules. Signals from BCH-Boundary, BCH-Custer, Malin-
Round Mountain, Grand Coulee, and Tacoma were selected from a set of
measurements provided from Bonneville Power Administration [33]. These
measurements were taken during the events previous to the August 10, 1996
blackout (Fig. 5.1).According with the above off-line rules, two subgroups were created in order
to monitor the California–Oregon inter-area mode. The online rules will be
described, for the time window shown in Fig. 5.5 after the Keeler–Allston line
opens.
5.3.2.1 Activation Deactivation Criteria
Multi-Prony monitoring will be triggered only after a small or medium distur-
bance has been detected. If not, triggers are issued and if no any other distur-
bance has occurred, the monitoring will be cancelled after 7–14 swings. Also, if
triggers have been issued and based on the monitoring damping ratios are going
wrong, the controller will be deactivated.
–340.3 –330.3 –320.3 –310.3 –300.31320
1330
1340
1350
1360
1370
1380
1390
1400
Time in Seconds
Filtered Malin-Round Mountain #1 MW Data
Fig. 5.5 First oscillatory stage. Case 1
5 A Real-Time Wide-Area Controller for Small-Signal Instability 139
5.3.2.2 Validating Criteria
Cross-checking results from the same time window and for consecutive timeintervals will increase reliability in the online estimations. Therefore, two cri-teria are developed in order to validate the Prony estimations:
1. A Prony damping estimate is considered valid only if the net variation of thedamping estimates from the different subgroups is less than a threshold, say�1. Then, the average damping estimate for the current time window isdefined as the average of the estimates from the subgroups.
2. When we consider the estimates from a consecutive set of moving timewindows, the damping estimates are used for control purposes: (a) if eachof the damping estimates in the different time windows is valid (i.e., they allobey the �1 rule) and (b) if the net variation of the average dampingestimate from the different time windows is less than a predefined thresholdsay �2.
These rules are explained through Table 5.5. Here �1 has been set to 2%. Asshown, only window five is out of the threshold and is considered not valid. Thiscan be understood because the Keeler–Allston line opening is a small distur-bance and Prony’s method is very accurate when there are no strong nonlinea-rities presented in the signals. On the other hand, for control purposes, considerthe estimates of two consecutive time windows and �2 set at 1.5%. It may beseen in Table 5.5 that the difference between the average damping ratio esti-mates of window 1 (5.3%) and window 2 (4.95%) are within the acceptedmargin difference of less than 1.5% (5.3 – 4.95% ¼ 0.35%). However, thisdifference is out of the threshold for windows 2 and 3 (4.95 – 1.35% ¼ 3.6%)and windows 3 and 4. Window 5 is not a valid estimate and windows 6 and 7agree within the threshold.
Finally, if a cutoff value of 2% damping ratio may issue triggers, theywould be issued only after the third and fourth windows agree. If the cutoff
Table 5.5 Multi–Prony COI mode damping–Case 1a – Estimation in percent
Time windowbPmrl, Pbound,
and ftacoma
Pcuster, Vmalin,and Pcoulee Average Validity
�330.3 to �318.3 5.0 5.6 5.3 Yes
�328.3 to �316.3 4.5 5.4 4.95 Yes
�326.3 to �314.3 0.4 2.3 1.35 Yes
�324.3 to �312.3 �0.4 0.0 �0.2 Yes
�322.3 to �310.3 0.3 5.0 2.65 No
�320.3 to �308.3 1.6 1.0 1.3 Yes
�318.3 to �306.3 3.0 2.3 2.65 YesaMeasured data taken just after Keeler-Allston line trips at �332.3 s.bTime with respect to Ross–Lexington line trip.Sampling frequency is 20 samples/s. Prony estimation was calculated using the RingdownGUI program from BPA/PNNLDynamic System Identification (DSI) Toolbox. Signal meanvalues were removed. A smoothing filter with 1 Hz cutoff frequency was used.
140 J. Quintero and V. (Mani) Venkatasubramanian
value is 3%, triggers may be issued after the fourth window or after theseventh window.
5.3.2.3 Selecting the SVC
An SVC priority list corresponding to each inter-area mode may be built usingthe criteria demonstrated next in Section 5.4.4.1 for the SVC local units. Thecentral unit will make sure that the selected SVC will be available and workingunder the expected conditions by checking the actual mode content in thesignals coming from the SVC location.
5.3.2.4 Determining the Phase Compensation
Additional to the trigger, the central unit will determine the type of phasecompensation to be applied for the SVC with respect to the direction of thecorresponding major inter-area power flow. This rule is demonstrated next inSection 5.4.4.2.
5.4 The SVC Local Unit
This control strategy assumes that some of the SVC units have been earmarkedfor potential operation under the proposed scheme, and that these SVCs arenormally operating in the voltage regulation mode. Specifically, when the SVCreceives the external trigger from the central control unit, the SVCwill switch tofull PSDC form.
The SVC mode switch occurs only when the oscillations are already so largeas to threaten the integrity of the large system. Under such highly stressedconditions, we assume that the SVCs can be switched to the full dampingcontrol mode to quickly damp out the otherwise growing (or sustained) oscilla-tions in the system. The damping control at the SVCs will be designed by usingthe active power flow from a specific intertie line as the control input signal.Details will be discussed next.
As we will show in Section 5.4.3, the damping control actions of an SVCworking on full PSDC located at the sending or at the receiving sides of anintertie line are more effective than those of an SVC located close to the middleof the intertie line. Additionally, as seen in our results, an SVC located at theactive power sending area offers the most effective location for full PSDC whenthe respective intertie active power flow is used as the control input for the SVCdamping control.
In this work, we are interested on making an analytic study of the effects ofthe SVC location with respect to an intertie line, on the associated inter-areamode. In order to keep a low-order model for our analysis purposes, generatorsare modeled for this part using a classical representation or swing equation
5 A Real-Time Wide-Area Controller for Small-Signal Instability 141
(5.6), which includes only two state variables per generator, rotor angle �, androtor speed deviation �!. Because inter-area modes are mainly associatedwith these two state generator variables, inter-area modes are still presentedand their characteristics are preserved when using this simple classical model foranalytical study. Later, we show that the design rules that are derived from theclassical model appear to be valid in simulations of detailedmodels as well as fora large power system model.
5.4.1 The Classical Power System Model
The swing equation for a generator may be written as
_� ¼ �!
� _! ¼ !s
2HPm � Pe � KD
�!
!s
� �� � (5:6)
where �! ¼ !� !s and !s is the synchronous speed, Pm is the shaft mechanicalpower in pu, Pe is the air-gap power in pu, and KD is a damping factor thatincludes mechanical and electrical damping components.
The classical representation can be derived from detailed power systemmodels bymaking the assumptions stated at [40].With generators characterizedby the transient reactance machine model and loads by constant impedances,the interconnected transmission network represented by
I ¼ YV (5:7)
may be portioned as
IG
0
� �¼
YGG YGR
YRG YRR
� �VG
VR
� �(5:8)
where subscript G relates to all internal generators nodes and subscript Rcomprises the remaining nodes with zero injection current. Reducing Eq. (5.8)by eliminating VR, we obtain
IG ¼ YGG � YGRY�1RRYRG
� �VG: (5:9)
Defining YGEN as
YGEN ¼ YGG � YGRY�1RRYRG (5:10)
142 J. Quintero and V. (Mani) Venkatasubramanian
we may write
IG ¼ YGENVG (5:11)
or
~IGi ¼Xng
j¼1
~YGENij~VGj for i ¼ 1; . . . ; ng (5:12)
where ng is the number of generators.From (5.12) and considering that air-gap power Pe is equal to generator
active output power P, neglecting stator resistance we may write that
Pi ¼Xng
j¼1EGiEGjYGENij cos �i � �j � ’GENij
� �for i ¼ 1; . . . ; ng: (5:13)
Initial values may be found from a pre-transient power flow load study.
5.4.2 The Linearized State–Space Classical Modelfor a Reduced Two-Area Power System
Here we reduce the original power system of four generators presented byKundur, as in Fig. 5.3, to only two generators. The compensator model usedfor the SVC in full PSDC operating mode, in this modeling case, is given onlyfor a low-pass filter followed by a phase compensation filter. Then, the PSDCmodel equations are
_eLP ¼1
T2LPP79REF � P79 � eLPð Þ
_BSVC ¼1
T2K eLP þ T1 _eLPð Þ � BSVC½ �
(5:14)
where P79 is the inter-area active power flow
P79 ¼V7V9
X79sinð�7 � �9Þ (5:15)
~V7; ~V9 bus voltages may be obtained approximately by
~V7 ffi E1ff�1 1� ~YGEN11 þ ~YGEN12
� jX0d1 þ jXT1
þ jX56 þ jX67
� �h i(5:16)
5 A Real-Time Wide-Area Controller for Small-Signal Instability 143
~V9 ffi E2ff�2 1� ~YGEN22 þ ~YGEN21
� jX0d2 þ jXT2
þ jX11�10 þ jX10�9� �h i
:
The nonlinear state–space model for the reduced two-area power system isdefined by the above equations, and the state–space variables are
x ¼ �1 �!1 �2 �!2 eLP BSVC½ �T:
An equilibrium point for the generated nonlinear state–space model can becalculated by
�10 ¼ cos�1Pm1 � YGEN110E
21 cosð’GEN110Þ
YGEN120E1E2
� �þ �20 þ ’GEN120
�!10 ¼ 0
�20 ¼ cos�1Pm2 � YGEN220E
22 cos ’GEN220ð Þ
YGEN210E2E1
� �þ �10 þ ’GEN210
�!20 ¼ 0
eLP0¼ P79REF �
V70V90
X79sin �70 � �90ð Þ
� �
BSVC0¼ KeLP0
:
(5:17)
Assuming that the equilibrium point stays away from the singularity of thenetwork equations, the linearized state–space model may be written as
� _x ¼ J�x: (5:18)
Therefore, the Jacobian matrix J around the equilibrium point (5.17) for thistwo-area power system, may be written as
J ¼
0 1 0 0 0 0
J21 � KD1
2H1J23 0 0 a
0 0 0 1 0 0
J41 0 J43 � KD2
2H20 b
J51 0 J53 0 � 1T2LP
c
J61 0 J63 0 d e
26666666664
37777777775
(5:19)
where the following J parameters depend mostly on the initial values and thepower system configuration. If we neglect variations in voltage magnitudes,these parameters are constant for the different PSDC phase compensatorstested
144 J. Quintero and V. (Mani) Venkatasubramanian
J21 ¼!sYGEN120E1E2
2H1sin �10 � �20 � ’GEN120ð Þ
J23 ¼ �J21
J41 ¼ �!sYGEN210E1E2
2H2sin �20 � �10 � ’GEN210ð Þ
J43 ¼ �J41:
(5:20)
The next J values are affected by the low-pass filter design included in thePSDC function, but independent of the other PSDC variables:
J51 ¼ �1
T2LPX79
@V7
@�1V90 sin �70 � �90ð Þ þ @�7
@�1V70V90 cos �70 � �90ð Þ
� �
J53 ¼ �1
T2LPX79
@V9
@�2V70 sin �70 � �90ð Þ � @�9
@�2V70V90 cos �70 � �90ð Þ
� �:
(5:21)
For convenience in analysis, we set the gain K as K ¼ T2=T1 so that thecompensator transfer function is of the pole-zero formGcðsÞ ¼ ðsþ 1=T1Þ=ðsþ 1=T2Þ. We study the impact of the phase-lead andphase-lag provided by the compensator on the damping of the inter-areamode, as they were the dominant contributors to the overall effectivenessof the compensator. The compensator gain has a lesser impact on thedamping as compared with the phase, as we have been able to test throughour simulations. With the gain K ¼ T2=T1, the parameters satisfy thefollowing property:
J61 ¼KT1
T2J51 ¼ J51
J63 ¼KT1
T2J53 ¼ J53:
(5:22)
On the other hand, the following parameters are going to change accordingto the SVC location:
a ¼ !s
2H1
� @YGEN11
@BSVCE21 cos ’GEN110ð Þ
þ @’GEN11
@BSVCYGEN110E
21 sin ’GEN110ð Þ
� @YGEN12
@BSVCE1E2 cos �10 � �20 � ’GEN120ð Þ
� @’GEN12
@BSVCYGEN120E1E2 sin �10 � �20 � ’GEN120ð Þ
2
666666666664
3
777777777775
(5:23)
5 A Real-Time Wide-Area Controller for Small-Signal Instability 145
b ¼ !s
2H2
� @YGEN22
@BSVCE22 cos ’GEN220ð Þ
þ @’GEN22
@BSVCYGEN220E
22 sin ’GEN220ð Þ
� @YGEN21
@BSVCE2E1 cos �20 � �10 � ’GEN210ð Þ
� @’GEN21
@BSVCYGEN210E2E1 sin �20 � �10 � ’GEN210ð Þ
2
666666666664
3
777777777775
(5:24)
c ¼ � 1
T2LPX79
@V7
@BSVCV90 sin �70 � �90ð Þ
þ @V9
@BSVCV70 sin �70 � �90ð Þ
þ @�7@BSVC
V70V90 cos �70 � �90ð Þ
� @�9@BSVC
V70V90 cos �70 � �90ð Þ
2666666666664
3777777777775
: (5:25)
Finally, it is clear that the last two parameters d and e are affected mostly bythe values of the PSDC design
d ¼ K
T21� T1
T2LP
� �(5:26)
e ¼ KT1
T2c� 1
T2:
Next, we will show how the parameters J21, J23, J41, J43, J51, J53, J61, and J63remain constant while the parameters a, b, and c vary when the SVC locationchanges. We will also see that the d and e values are modified mostly by thePSDC phase compensation type. By analyzing the parameters a, b, and c, wecan then derive rules on the placement of the SVC. Similarly, by studying theproperties of the parameters d and e, we can derive rules on the lead versus lagdesign principle of the PSDC compensator.
5.4.3 Numerical Results
Figure 5.6 shows the Jacobianmatrices JSVCB7_LP, JSVCB8_LP, and JSVCB9_LP of thereduced two-area power system above with an SVC located at Bus 7, Bus 8, andBus 9, respectively. The SVCs are in the PSDCmode and the controller includes alow-pass filter anda zero phase compensator (5.14).Note that as described in (5.19)and (5.20), the submatrix related with the original system without SVCs does notchange with the addition of the SVCs. This is because the change in the admittanceYGEN matrix is too small when an SVC is added. Therefore, values included into
146 J. Quintero and V. (Mani) Venkatasubramanian
this submatrix do not have any significant effect in the mode shape variation that
takes place when an SVC working in full PSDC is included into the system.On the other hand, the values of parameters a, b, and c change with respect to
the SVC location at any of the two areas or along the intertie line. Table 5.6 shows
these parameter value variations when an SVC is located in Bus 7, Bus 8, and Bus
9. From (5.23), (5.24), and (5.25) and the approximations we havemade, changes
in a, b, and c parameter values, when moving from one side to the other, depend
mostly on the changes suffered by admittance sensitivities. Between them, the
parameter c value is the one that experienced the smaller changes and, therefore,
has little effect on the observed differences in the inter-area mode damping.Admittance sensitivities for an SVCatBus 7, Bus 8, andBus 9were found for a
positive step in its BSVC bus output value (5.14). These sensitivities are shown in
Fig. 5.7, as well as the variation suffered by each of the four right terms in a and b
formulations, (5.23) and (5.24). If we compare corresponding sensitivities in a
and b terms, sensitivities @’GEN11=@BSVC and @’GEN22=@BSVC are the
only sensitivity factors with opposite slopes when moving the SVC from the
sending area to the receiving area, through the intertie. Therefore, they are the
main cause of the value difference between second right terms of a and b expres-
sions, called here a2 and b2, and ultimately for the value difference between terms
a and b. We may say here, that the difference between the a and b values is a
deterministic characteristic of the SVC location in the PSDC mode.
Fig. 5.6 The J matrix with SVC compensation at Bus 7, Bus 8, and Bus 9. The SVC controlincludes a low-pass filter and a 08 phase compensator. KD=0.5. Matlab model
5 A Real-Time Wide-Area Controller for Small-Signal Instability 147
Table5.6
Two-areapower
system
a. .(M
atlabresults)
SVClocation
PSDCcompensation(lowpass
filter
+)
ab
cd
el 1
,2f
�
NoSVC
––
––
––
�0.020�7.214i
1.148
0.003
Bus7
LPfilter
only
�6.847�3.418�0.233
0.033
�6.516�0.378�6.914i
1.100
0.055
LPfilter
+158lag
2.083
�4.995�0.580�6.995i
1.113
0.083
LPfilter
+308lag
5.514
�3.567�0.817�7.230i
1.151
0.112
LPfilter
+458lag
10.416
�2.614�0.989�7.559i
1.203
0.130
LPfilter
+158lead
�1.488
�8.567�0.214�6.913i
1.100
0.031
LPfilter
+308lead
�2.916�11.998�0.084�6.965i
1.108
0.012
LPfilter
+458lead
�3.869�16.900�0.020�7.029i
1.119
0.003
Bus8
LPfilter
only
�5.406�5.967�0.411
0.033
�6.694
0.031�7.262i
1.156�0.004
LPfilter
+158lag
2.083
�5.173
0.058�7.257i
1.155�0.008
LPfilter
+308lag
5.514
�3.745
0.100�7.242i
1.152�0.014
LPfilter
+458lag
10.416
�2.792
0.152�7.217i
1.149�0.021
LPfilter
+158lead
�1.488
�8.745
0.009+
7.261i
1.156�0.001
LPfilter
+308lead
�2.916�12.176�0.009�7.254i
1.154
0.001
LPfilter
+458lead
�3.869�17.078�0.019�7.244i
1.153
0.003
Bus9
LPfilter
only
�3.263�7.429�0.580
0.033
�6.863
0.322�7.554i
1.202�0.043
LPfilter
+158lag
2.083
�5.342
0.489�7.54li
1.200�0.065
LPfilter
+308lag
5.514
�3.913
0.743�7.503i
1.194�0.098
LPfilter
+458lag
10.416
�2.961
1.059�7.474i
1.189�0.140
LPfilter
+158lead
�1.488
�8.913
0.185�7.542i
1.200�0.024
LPfilter
+308lead
�2.916�12.345
0.059�7.495i
1.193�0.008
LPfilter
+458lead
�3.869�17.247�0.014�7.431i
1.183
0.002
aThisisareducedtw
o-areapower
system
,withonly
twogenerators,theSVC
consistsonly
ofalow-pass
filter
andaleadorlagfilter,KD=
0.5,
p78=
90MW.
Resultswerecalculatedusingadesigned
Matlabmodel.
148 J. Quintero and V. (Mani) Venkatasubramanian
In this particular Matlab model, the value difference term, b – a, is positivewhen the SVC is at the sending side, almost zero if the SVC is located close tothe center of the intertie line, and negative when the SVC is at the receiving side.Also, we note that the greater the absolute value difference between a and b, thegreater the stabilizing or destabilizing capacity of the SVC in the full PSDCmode. This is based on several simulations and the eigenvalue analysis. In Table5.6, this phenomenon may be observed when comparing the stabilizing anddestabilizing effects of the SVCs in Bus 7 and Bus 9 with respect to an SVC inBus 8. Simulations using the MASS software package presented in [41] confirmthis SVC location relation with values a and b.
Additionally, we have parameters d and e (5.26) that depend on the SVCscontroller specifications and on the parameter c, which does not changemuch inthis study. Therefore, as seen in Table 5.6, parameters d and e have similarvalues at Bus 7, Bus 8, and Bus 9, as they vary mostly with the type of PSDCcompensation used and not with the SVC location. Note that in order to obtaincomparative results and to avoid the gain K effect in d and e values, we setK ¼ T2=T1. As may be seen in (5.26), withK ¼ T2=T1, the parameter d dependsmostly onT1 while the parameter e relates mostly toT2; the magnitudes of d ande vary inversely proportional to T1 and T2, respectively.
5.4.4 SVC Rules
Based on the above analysis and numerical results, we state that parameters aand b values have the biggest impact on damping inter-area modes with an SVCin the full PSDC operation. That is, the location of the SVC with respect to theintertie active power flow direction is the most important factor to consider forimplementing this type of control. Then, based on the SVC location, the inter-area active power phase compensation should be decided, for suitably settingthe d and e values.
7 8 9–0.05
0
0.05
0.1
0.15
0.2
BUS
Sensitivities for Parameters a and b
7 8 9–0.18
–0.14
–0.1
–0.06
–0.02
0.02Term Components for a and b
BUS
svc
GEN
B∂∂ϕ 11
svc
GEN
B∂∂ϕ 22
srasrb
Fig. 5.7 Sensitivities and term component variations in a and b
5 A Real-Time Wide-Area Controller for Small-Signal Instability 149
5.4.4.1 Selecting the SVC Location
The SVC priority list corresponding to each inter-area mode may be built usingthe criteria shown in Section 5.4.3 for the SVC local units. These criteria may bewritten as follows:
1. An SVC at the sending inter-area active power flow area has more steeringpower than one located at the receiving area.
2. Between SVCs at the same areas, the greater the corresponding SVCs b –aabsolute value difference the greater the steering power.
5.4.4.2 Determining the Phase Compensation
From Table 5.6, it is clear that an intensification of the intertie active power-based phase-lag compensation, or the d magnitude increment, is particularlybeneficial on increasing the inter-area mode damping for an SVC located at thesending area, as Bus 7, while it is disadvantageous when the SVC is located atthe receiving area at Bus 9. However, it is important to note that increasing dtoo much in Bus 7, could lead to another pole close to the imaginary axis tomove to the right-hand side attracted by a zero on this side. Alternatively, theintensification of the phase-lead compensation has a good effect on dampingthe inter-area mode, when the SVC is located at the receiving area. An oppositeeffect results, when it is located on the sending side. For an SVC close to themiddle of the intertie line at Bus 8, the effects are much lower.
This rule may be expressed as follows:
1. An SVC in full PSDC is operated as a phase-lag compensator if the SVC is atthe sending side for the active power flow of the major intertie line that isbeing used as the control input.
2. An SVC in full PSDC is operated as a phase-lead compensator if the SVC isat the receiving side for the active power flow of the major intertie line that isbeing used as the control input.
5.5 WSCC Power System Example
Measured data taken at BPA Dittmer Control Center during the events pre-vious to the August 10, 1996 WSCC Power System breakup [33] are analyzedhere in order to illustrated our approach. Also, a validated model [36] of thisphenomenon is used to simulate our control strategy.
FromFig. 5.1 and [36], two different oscillatory stagesmay be identified. Thefirst oscillatory stage initiates when the Keeler–Allston line trips. Ten secondslater, the multi-Prony analysis on recorded signals shows a center of inertia(COI) mode with a damping ratio around 1%. After this disturbance, measuredata are analyzed using the multi-Prony approach presented in this work. Thewhole analyzed time window is presented in Fig. 5.5.
150 J. Quintero and V. (Mani) Venkatasubramanian
The improved accuracy on the estimations of using multi-Prony by groups
over the single Prony, may be seen in Tables 5.7 and 5.8 with real data.
Estimations in Table 5.7 give a wide range of possible values for the COImode damping ratio for that time window. As seen, the estimated damping
ratio values have differences of more than 4%. Table 5.8 shows how groupingmeasurements according to the rules developed in Section 5.3.1 makes the
estimated damping ratio differences less than 2% for the same time window.Note that, according with the analysis developed in Section 5.4.4, as an SVC
located at Maple Valley at that time would be at the active power sending side,
the COI active power phase compensation should be lagging and this SVC is thefirst choice in the priority list. On the other hand, an SVC located in Adelanto
would be at the receiving active power side, the COI active power phasecompensation should be leading at that bus and it is our second choice.
Table 5.7 Prony mode parameter estimation – WSCC power systems – (Measured signalswith COI dominant mode)
Symbol Measured signalsa Ai fi �i Ari
COI mode
Pmr1 Malin-Round Mountain #1 21.57 0.259 0.0452 1.000
Pcuster BCH-Custer 23.08 0.263 0.0125 1.000
Pbound BCH-Boundary 8.372 0.259 0.0103 1.000
Pcoulee Grand Coulee Generation 4.041 0.244 �0.0152 1.000
VMalin Malin bus voltage magnitude 1.785 0.259 0.0019 1.000
fTacoma Tacoma bus frequency 0.006 0.233 �0.0028 0.885aTime window is from �320.3 to �308.3 s, Fig. 5.5. Ten seconds, after Keeler–Allston linetrips. Sampling frequency is 20 samples/s.Prony estimation was calculated using the Ringdown GUI program from BPA/PNNLDynamic System Identification (DSI) Toolbox. Signalmean values were removed.A smoothingfilter with 1 Hz cutoff frequency was used.
Table 5.8 Multi-Prony mode parameter estimation – WSCC power system – (Groupedmeasured signals)
Symbol Ai fi �i Ari
COI mode
Pmr1,
Pbound,and fTacoma
16.62
8.7080.006
0.254 0.0158 1.000
0.6010.830
Pcuster,
Vmalin,and PCoulee
(SVD) (No Good Estimations)
Pcuster,
Vmalin,and PCoulee
(QR Factorization) (No Good Estimations)
Pcuster,Vmalin,and PCoulee
(TLS) 24.751.8894.548
0.256 0.0097 1.0001.0000.533
Same signal measurements and programs as Table 5.7.
5 A Real-Time Wide-Area Controller for Small-Signal Instability 151
If due to historical or technical reasons, the full PSDC compensation is not
activated during the first oscillatory stage, the second stage which starts at
time zero in our plots after the Ross–Lexington line trips (Fig. 5.1) presents a
worse oscillatory scenario. At that time, a sequential tripping of generators
and governor actions began, introducing a stronger oscillatory behavior
and also stronger nonlinearities in the signals. A closer look is presented in
Fig. 5.8.Using the same measurements from BPA, we tried to estimate the damping
of these oscillations applying our multi-Prony approach by groups through
the time frame defined by the dashed lines in Fig. 5.8. Results are presented in
Table 5.9.
–10 0 10 20 30 401240
1260
1280
1300
1320
1340
1360
Time in Seconds
Filtered Malin-Round Mountain #1 MW Data
Fig. 5.8 Second oscilatory stage. Case 2
Table 5.9 Multi-Prony COI mode damping estimation in precent–Case 2a
Timewindow (0)
Pmrl, PBound
and fTacoma
PCuster, VMalin
and PCoulee Average Validity
2.0–14.0 �2.7 3.6 0.45 No
4.0–16.0 �1.3 �3.6 �2.45 No
6.0–18.0 1.0 1.1 1.0 Yes
8.0–20.0 �0.9 0.2 �0.35 Yes
10.0–22.0 0.0 0.8 0.4 Yes
12.0–24.0 1.2 1.6 1.4 Yes
14.0–26.0 3.5 1.9 2.7 Yes
16.0–28.0 1.1 0.8 0.95 Yes
18.0–30.0 �1.3 2.2 0.45 No
20.0–32.0 �2.1 �3.4 �2.75 Yes
22.0–34.0 �3.4 �3.1 �3.25 YesaMeasured data taken just after Ross–Lexington line trips at 0.0s.bTime with respect to Ross–Lexington line trip.Sampling frequency is 20 samples/s. Prony estimation was calculated using theRingdown GUI program from BPA/PNNL Dynamic System Identification(DSI) Toolbox. Signalmean values were removed.A smoothing filter with 1Hzcutoff frequency was used.
152 J. Quintero and V. (Mani) Venkatasubramanian
Due to the strong nonlinearities and with an �1 equal to 2%, the first
two time windows estimations are inaccurate and they are neglected. Later,
our results show that the COI mode is going lightly damped less than 1%
damping ratio. If for control purposes we consider the estimates of two
consecutive moving time windows, �2 equal to 1.5%, and a cutoff value of
1% damping ratio, the central controller will issue triggers to the assigned
SVC after 20 s. If we use three consecutive windows and �2 equal to 1.5%,
then the triggers would be issued at 22 s for the same cutoff damping
of 1%.However, if we want to be very conservative and wait for the mode going
really unstable we may set the cutoff value at say –2% damping ratio. Then, in
Table 5.9 considering only two consecutive time windows the PSDC is triggered
at 35 s. Figure 5.9 represents the simulated behavior for this case, provide the
right phase compensation as described in Section 5.4.4 for an SVC located at
Maple Valley, the active power sending area, and for an SVC located at
Adelanto, the active power-receiving area.Finally, Table 5.10 presents the frequency and damping ratios estimated
from a time window from 85 to 97 s for triggers issued at –15 s and at 35 s,
when the right compensation control actions are taken and when the wrong
compensation control actions are applied. Note that, as we showed in Section
5.4.3, an SVC at the sending side has more steering power than one at the
receiving side, so it could add more damping but also could be more destabiliz-
ing when right phase compensation is not applied.
0 10 20 30 40 50 60 70 80 90 1003800
4000
4200
4400
4600
4800
5000
0 10 20 30 40 50 60 70 80 90 1003800
4000
4200
4400
4600
4800
5000
Time in seconds
Maple Valley full PSDC Lag Compensation at 35 seconds
Adelanto full PSDC Lead Compensation at 35 seconds
Fig. 5.9 Simulated COI behavior for an SVC triggered at 35 s to full PSDC. At MapleValley with COI phase-lag compensation or at Adelanto with COI phase-leadcompensation
5 A Real-Time Wide-Area Controller for Small-Signal Instability 153
5.6 Conclusions
This work presents a powerful and practical real-time wide-area controller which is
specifically aimed at improving the dynamic security in the small-signal sense. This
real-time control strategy showed how, with increased observability and reliability, a
signal analysis tool likemulti-Prony (ormatrix pencil, orHankel TLS, or a combina-
tion of them as in [34]) can be successfully applied on real-time basis for detecting the
proximity andonset of small-signal instability phenomena,while avoiding thepitfalls
and the shortcomings widely known and reviewed in the literature.The controller then initiates aggressive mitigatory control actions in certain areas
by using commonly available resources such as the SVCs inmodern power networks.
The controller will come into play only under extremely stressed operating conditions
when the damping of some oscillatory modes becomes problematic.Recent blackout events such as the 1996 western American events, 2004
Noreastern disturbance and the 2004 Italian blackout have clearly demon-
strated that the power system will repeatedly find itself in highly stressed
unplanned for operating scenarios. We need to design new automatic control-
lers which are specifically targeted toward such unforeseen operating condi-
tions which mitigate the problem from cascading into big blackouts. This effort
proposes one such ‘‘safety net’’ controller for stabilizing the large system while
facing small-signal instability problems.
Acknowledgments This work was supported by funding from Power Systems EngineeringResearch Center (PSERC) and by Consortium for Reliability Technology Solutions(CERTS), funded by the Assistant Secretary of Energy Efficiency and Renewable Energy,Office of Distributed Energy and Electricity Reliability, and Transmission Reliability Pro-gram of the US Department of Energy under Interagency Agreement No. DE-AI-99EE35075with the National Science Foundation. Partial funding of the work from Bonneville PowerAdministration is also gratefully acknowledged. J. Quintero received partial support fromUniversidad Autonoma de Occidente and Colciencias–Fulbright–Laspau.
Table 5.10 Simulation COI mode results
PSDCb action atPhasecompensation Trigged after f �
No PSDC Action 0.210 �0.0144Mapple Valley(sending Bus)
458 lag
608 lead
Keeler–Allston trips c
Ross–Lexington trips d
Keeller–Allston trips c
Ross–Lexington trips d
0.181
0.183SystemSystem
0.0697
0.0908CollapsesCollapses
Adelanto(receiving Bus)
458 lag
608 lead
Keeler–Allston trips c
Ross–Lexington trips d
Keeler–Allston trips c
Ross–Lexington trips d
0.2400.2200.2290.230
�0.0216�0.02890.02840.0201
aTime window is from 85 to 97s.bSVCs are originally working in voltage regulation control.cFull PSDC function triggers at �15 s.dFull PSDC function triggers at 35.
154 J. Quintero and V. (Mani) Venkatasubramanian
References
1. C. Taylor, D. Erickson, B. Wilson, and K. Martin. (2002, December 16). Wide-AreaStability and Voltage Control System (WACS) Demostration (Living Document). Avail-able: http://www.transmission.bpa.gov/orgs/opi/index.shtm
2. K. Tomsovic, D. Bakken, V. Venkatasubramanian, and A. Bose, ‘‘Designing the NextGeneration of Real-Time Control, Communication and Computations for Large PowerSystems,’’ IEEE Proceedings, vol. 93, pp. 965–979, May 2005.
3. U.S.–Canada Power System Outage Task Force, ‘‘Causes of the August 14th Blackout inthe United States and Canada,’’ Interim Rep., Nov. 2003, ch. 6.
4. A. E. Hammad, ‘‘Analysis of power system stability enhancement by static var compen-sators,’’ IEEE Trans. Power Systems, vol. PWRS-1, pp. 222–227, Nov. 1986.
5. E. Lerch, D. Povh, and L. Xu, ‘‘Advanced SVC control for damping power systemoscillations,’’ IEEE Trans. Power Systems, vol. 6, pp. 524–535, May 1991.
6. R. L. Lee,M. J. Beshir, A. T. Finley, D. R. Hayes, J. C. Hsu, H. R. Peterson, G. L. DeShazo,andD.W. Gerlach, ‘‘Application of static var compensators for the dynamic performance oftheMead-Adelanto andMead-Phoenix transmission projects,’’ IEEE Trans. Power Delivery,vol. 10, pp. 459–466, Jan. 1995.
7. J. R. Smith, D. A. Pierre, D. A. Rudberg, I. Sadighi, A. P. Johnson, and J. F. Hauer, ‘‘Anenhanced LQ adaptive var unit controller for power system damping,’’ IEEE Trans.Power Systems, vol. 4, pp. 443–451, May 1989.
8. K. M. Son and J. K. Park, ‘‘On the robust LQG control of TCSC for damping powersystem oscillations,’’ IEEE Trans. Power Systems, vol. 15, pp. 1306–1312, Nov. 2000.
9. Q. Zhao and J. Jiang, ‘‘Robust SVC controller design for improving power systemdamping,’’ IEEE Trans. Power Systems, vol. 10, pp. 1927–1932, Nov. 1995.
10. X. Yu, M. Khammash, and V. Vittal, ‘‘Robust design of a damping controller for static varcompensators in power systems,’’ IEEETrans. Power Systems, vol. 16, pp. 456–462,Aug. 2001.
11. H. Bourles, S. Peres, T. Margotin, and M. P. Houry, ‘‘Analysis and design of a robustcoordinated AVR/PSS,’’ IEEE Trans. Power Systems, vol. 13, pp. 568–575, May 1998.
12. I. Kamwa, G. Trudel, and L. Gerin-Lajoie, ‘‘Robust design and coordination of multipledamping controllers using nonlinear constrained optimization,’’ IEEE Trans. PowerSystems, vol. 15, pp. 1084–1092, Aug. 2000.
13. J. C. R. Ferraz, N.Martins, andG. N. Taranto, ‘‘Simultaneous partial pole placement forpower system oscillation damping control,’’ presented at the IEEE Winter Power Meet-ing, pp. 1154–1159, Jan. 28 – Feb. 1, 2001.
14. J. J. Sanchez-Gasca, ‘‘Coordinated control of two FACTS devices for damping interareaoscillations,’’ IEEE Trans. Power Systems, vol. 13, pp. 428–434, May 1998.
15. P. Pourbeik and M. J. Gibbard, ‘‘Simultaneous coordination of power system stabilizersand FACTS device stabilizers in a multimachine power system for enhancing dynamicperformance,’’ IEEE Trans. Power Systems, vol. 13, pp. 473–479, May 1998.
16. X. Lei, E. N. Lerch, and D. Povh, ‘‘Optimization and coordination of damping controls forimproving system dynamic performance,’’ IEEE Trans. Power Systems, vol. 16, pp. 473–480,Aug. 2001.
17. I. Kamwa, R. Grondin, and Y. Hebert, ‘‘Wide-area measurement based stabilizingcontrol of large power systems – A decentralized/hierarchical approach,’’ IEEE Trans.Power Systems, vol. 16, pp. 136–153, Feb. 2001.
18. G. E. Boukarim, S.Wang, J. H. Chow,G. N. Taranto, andN.Martins, ‘‘A comparison ofclassical, robust, and decentralized control designs for multiple power system stabilizers,’’IEEE Trans. Power Systems, vol. 15, pp. 1287–1292, Nov. 2000.
19. M. J. Gibbard, N. Martins, J. J. Sanchez-Gasca, N. Uchida, V. Vittal, and L. Wang,‘‘Recent applications of linear analysis techniques,’’ IEEE Trans. Power Systems, vol. 16,pp. 154–162, Feb. 2001.
5 A Real-Time Wide-Area Controller for Small-Signal Instability 155
20. G. R. B. Prony, ‘‘Essai experimental et analytique sur les lois de la dilatalrlite de fluidselastiques et sur cells de la vapeur de l’alcool, a differents tempoeratures,’’ Journal del’Ecole Polytechnique (Parıs), vol. 1, pp. 24–76, 1795.
21. D. J. Trudnowski, J. R. Smith, T. A. Short, and D. A. Pierre, ‘‘An application of Pronymethods in PSS design for multi-machine systems,’’ IEEE Trans. Power Systems, vol. 6,pp. 118–126, Feb. 1991.
22. P. S. Dolan, J. R. Smith, and W. A. Mittelstadt, ‘‘Prony analysis and modeling of a TCSCundermodulation control,’’ inProc. 4th IEEEConference onControl Applications, pp. 239–245,Sept. 1995.
23. M. Amono, M. Watanabe, and M. Banjo, ‘‘Self-testing and self-tuning of power systemstabilizers using Prony analysis,’’ in Proc. IEEEWinter Power Meeting, pp. 655–660, 1999.
24. S.W. Chen, ‘‘A two-stage discrimination of cardiac arrhythmias using a total least squares-based Prony modeling algorithm,’’ IEEE Trans. Biomed. Eng., vol. 47, pp. 1317–1327, Oct.2000.
25. F. M. El-Hefnawi, ‘‘Use of Prony’s method for extracting the poles and zeros yielding awideband window type response of circular antenna arrays,’’ in Proc. IEEE Radio andWireless Conference, pp. 201–204, 1998.
26. K. F. Sabett, L. P. B. Katehi, and K. Sarabandi, ‘‘Wavelet-based CAD modeling ofmicrostrip discontinuities using least square Prony’s method,’’ in Proc. IEEE MTT-SInternational, pp. 1799–1802, 1997.
27. M. Meunier and F. Brouaye, ‘‘Fourier transform, wavelets, Prony analysis: tools forharmonics and quality of power,’’ in Proc. 8th International Conf. Harmonics and Qualityof Power, pp. 71–76, 1998.
28. A. R. Messina and V. Vittal, ‘‘Nonlinear, non-stationary analysis of interarea oscillationsviaHilbert spectral analysis,’’ IEEETrans. Power Systems, vol. 21, pp. 1234–1241,Aug. 2006.
29. D. Ruiz-Vega, A. R. Messina, and M. Pavella, ‘‘Online Assessment and Control ofTransient Oscillations Camping,’’ IEEE Trans. Power Systems, vol. 19, pp. 1038–1047,May. 2004.
30. J. Xiao, X. Xie, J. Han, and J. Wu, ‘‘Dynamic Tracking of Low-frequency Oscillationswith Improved Prony Method in Wide-Area Measurement System,’’ presented at theIEEE Power Meeting, Denver, CO, June 6–10, 2004.
31. R. Zivanovic and P. Schegner, ‘‘Pre-filtering improves prony analysis of disturbancerecords,’’ in Proc. IEE Int. Conf. Developments in Power System Protection, pp.780–783, 2004.
32. M. A. Johnson, I. P. Zarafonitis, and M. Calligaris, ‘‘Prony analysis and power systemstability-some recent theoretical and applications research,’’ inProc. IEEE Power summerMeeting, pp. 1918–1923, 2000.
33. BPA/PNNL Dynamic Systems Identification (DSI) Toolbox: Version 3.0.4, BattelleMemorial Institute, 1998.
34. G. Liu, J. Quintero, and V. Venkatasubramanian, ‘‘OscillationMonitoring System BasedonWide Area Synchrophasors in Power Systems,’’ in Proc. Bulk Power System Dynamicsand Control, Aug. 2007.
35. M.Klein, G. J. Rogers, and P.Kundur, ‘‘A Fundamental Study of Inter-AreaOscillationsin Power Systems,’’ IEEE Trans. Power Systems,vol. 6, pp. 914–921, Aug. 1991.
36. D. N. Kosterev, C. W. Taylor, and W. A. Mittelstadt, ‘‘Model validation for the August10, 1996 WSCC system outage,’’ IEEE Trans. Power Systems, vol. 14, pp. 967–979, Aug.1999.
37. J. F. Hauer, C. J. Demeure, and L. L. Scharf, ‘‘Initial Results in Prony Analysis of PowerSystem Response Signals,’’ IEEE Trans. Power Systems,vol. 5, pp. 80–89, Feb. 1990.
38. D. J. Trudnowski, J. M. Johnson, and J. F. Hauer, ‘‘SIMO System Identification fromMeasured Ringdowns,’’ in 1998 Proc. American Control Conf., pp. 2968–2972.
39. V. Venkatasubramanian and J. R. Carroll, ‘‘Oscillation Monitoring Sytem at TVA’’,presentation at North American Synchrophasor Initiative meeting at New Orleans,
156 J. Quintero and V. (Mani) Venkatasubramanian
LA, March 2008, [Online] Available http://www.naspi.org/meetings/workgroup/2008_march/session_one/tva_oscillation_monitoring_venkatasubramanian.pdf
40. P. M. Anderson and A. A. Fouad, Power System Control and Stability.New York: IEEEPress, 1994.
41. J. Quintero, ‘‘A Real-Time Wide-Area Control for Mitigating Small-Signal Instability inLarge Electric Power Systems,’’ Ph.D. dissertation, Sch. of Electrical Eng. and Comp.Sciences, Washington State Univ., Pullman, 2005 [Online]. Avail.: http://www.disserta-tions.wsu.edu/Dissertations/Spring2005/j_quintero_011905.pdf
5 A Real-Time Wide-Area Controller for Small-Signal Instability 157
Chapter 6
Complex Empirical Orthogonal Function Analysis
of Power System Oscillatory Dynamics
P. Esquivel, E. Barocio, M.A. Andrade, and F. Lezama
Abstract Multivariate statistical data analysis techniques offer a powerful toolfor analyzing power system response from measured data. In this chapter, astatistically based, data-driven framework that integrates the use of complexempirical orthogonal function analysis and the method of snapshots is pro-posed to identify and extract dynamically independent spatiotemporal patternsfrom time-synchronized data. The procedure allows identification of the domi-nant spatial and temporal patterns in a complex data set and is particularly wellsuited for the study of standing and propagating features that can be associatedwith electromechanical oscillations in power systems. It is shown that, in addi-tion to providing spatial and temporal information, the method improves theability of conventional correlation analysis to capture temporal events and givesa quantitative result for both the amplitude and phase of motions, which areessential in the interpretation and characterization of transient processes inpower systems. The efficiency and accuracy of the developed procedures forcapturing the temporal evolution of the modal content of data from timesynchronized phasor measurements of a real event in Mexico is assessed.Results show that the proposed method can provide accurate estimation ofnonstationary effects, modal frequency, time-varying mode shapes, and timeinstants of intermittent or irregular transient behavior associated with abruptchanges in system topology or operating conditions.
6.1 Empirical Orthogonal Function Analysis
Empirical orthogonal function (EOF) analysis is a statistical method of findingoptimal distributions of energy from an ensemble ofmultidimensional measure-ments [1]. The essential idea is to generate an optimal basis for the
P. Esquivel (*)Department of Electrical and Computer Engineering, The Center for Research andAdvanced Studies, Cinvestav, Guadalajara, Mexicoe-mail: [email protected]
A.R. Messina, Inter-area Oscillations in Power Systems,Power Electronics and Power Systems, DOI 10.1007/978-0-387-89530-7_6,� Springer ScienceþBusiness Media, LLC 2009
159
representation of an ensemble of data collected from measurements or numer-ical simulations of a dynamic system.
Given an ensemble of measured data, the technique yields an orthogonalbasis for the representation of the ensemble, as well as a measure of the relativecontribution of each basis function to the total energy with no a priori assump-tion on either spatial or temporal behavior.
The following sections provide a review of some aspects of the qualitativetheory of empirical orthogonal functions that are needed in the analysis of low-dimensional models derived from the technique.
We start by introducing the method in the context of statistical correlationtheory.
6.1.1 Theoretical Development
The proper orthogonal decomposition (POD) method is an optimal techniqueof finding a basis that spans an ensemble of data, collected from an experimentor numerical simulation [1–4]. More precisely, assume that uðxj; tkÞ, j=1,. . .,n,k=1,. . .,N, denotes a sequence of observations on some domain x 2 � where xis a vector of spatial variables and tk½0;T� is the time at which the observationsare made. Without loss of generality, the time average of the time sequence
umðxÞ ¼ u x; tkð Þh i ¼ 1
N
XN
k¼1u x; tkð Þ
is assumed to be zero [3]. Generalizations to this approach are discussed below.The POD procedure determines EOFs, jiðxÞ, i ¼ 1; . . . ;1, such that the
projection onto the first p EOFs (a low-order representation)
uðxj; tkÞ ¼Xp
i¼1aiðtÞjiðxÞ; j ¼ 1; . . . ; n; k ¼ 1; . . . ;N (6:1)
is optimal in the sense that the average least-squares truncation error, "j
"j ¼ uðxj; tkÞ �Xp
i¼1aiðtÞjiðxÞ
�����
�����
2* +; p � N (6:2)
is minimized for any p � N, where :h i denotes the ensemble average,fk k ¼ f; fh i1=2, and :k k denotes theL2norm over�. The ai’s are time-dependent
coefficients of the decomposition to be determined so that (6.1) results in amaximum for (6.2). These special orthogonal functions are called the properorthogonal modes (POMs) of the reduced basis for uðxj; tkÞ.
160 P. Esquivel et al.
Following [1], assume that the field is decomposed into a mean valueum xj; t� �
and a fluctuating part �uðx; tjÞ
uðxj; tkÞ ¼ umðxj; tkÞ þ �uðxj; tkÞ (6:3)
More formally, let L2 denote the space of square integrable functions. It thenfollows that, a normalized basis function j is optimal if the average projectionof uonto j is maximized, i.e., [2]
maxj2L2 ½0;1�ð Þ
jð�uðxj; tkÞ;jÞj2D E
subject to jk k2¼ 1 (6:4)
where the inner product is defined as U;Vh i ¼ pk¼0 UkV
�k ¼ VHU, and
jk k2¼ j;jh i ¼ jTj ¼Xn
j¼1j2j
The optimization problem can be recast in the form of a functional for theconstrained variational problem [3]1
J½j� ¼ ð�u xj; tk� �
;j�� ��2D E
� l jk k2�1� �
(6:5)
A necessary condition for the extrema is that the Gateaux derivative vanishesfor all variations jþ �c 2 L2 ½0; 1�ð Þ, � 2 <. This can be expressed as
dJ
d�jþ �c½ �
�����¼0¼ 0; 8c 2 L2ð�Þ (6:6)
Consider now the Hilbert space of all pairs f; gh i where fand gare functionsof L2½0; 1�, i.e., square integrable functions of the space variable x on theinterval ½0; 1�, where :; :h i denotes the standard inner product on L2 defined by
f; gh i ¼Z 1
0
fðxÞg�ðxÞ dx (6:7)
and
j2�� �� ¼ j;jh i ¼
Z
�
j2 dx (6:8)
where � is the domain of interest over which uðxÞ are j are defined and theasterisk * denotes the conjugate transpose.
1 Given a function to maximize, fðPÞ subject to the constraints gðPÞ ¼ 0, the Lagrangefunction can be defined as FðP; lÞ ¼ fðPÞ � lgðPÞ.
6 Complex Empirical Orthogonal Function Analysis 161
It follows immediately from (6.6) that
dJ
d�½jþ �c�j�¼0¼ 0 ¼ dJ
d��u;jþ �cð Þ jþ �c; �uð Þ � l jþ �c;jþ �cð Þh i½ �
�����¼0
¼ 2Re �u;cð Þ j; �uð Þh i � l j;cð Þ½ �(6:9)
where use has been made of the inner product properties.Noting that
�u;cð Þ j; �uð Þh i � l j;cð Þ ¼Z 1
0
�uðxÞc�ðxÞdxZ 1
0
jðx0Þu�ðx0Þdx0�
� lZ 1
0
jðxÞc�ðxÞdx
¼Z 1
0
Z 1
0
uðxÞu�ðx0Þh ijðx0Þdx0 � ljðxÞ �
c�ðxÞ dx ¼ 0
the condition for the extrema reduces to
Z
�
�uðxÞ�u�ðx0Þh ijðx0Þ dx0 ¼ ljðxÞ (6:10)
Equation (6.10) has a finite number of orthogonal solutions over jiðxÞ (thePOMs) with corresponding real and positive eigenvalues li. They are conse-quently called empirical orthogonal functions.
Defining
Rj ¼Z
�
�uðx; tÞ�u�ðx0; tÞh ijðx0Þ dx0 (6:11)
where
Rðx; x0Þ ¼ 1
N
XN
k¼1�uðx; tkÞ�uðx0; tkÞ (6:12)
the problem of minimizing (6.9) becomes that of finding the largest eigenvalueof the eigenvalue problem Rj ¼ lj, subject to jk k2¼ 1.
In practice, the observations that form the data are only available at discretespatial grid points herein called snapshots. In this case, the kernel Rðx; x0Þ canbe written as [5, 6]
Rðx;x0Þ ¼
Rðx1; x1Þ � � � Rðx1;xnÞ... . .
. ...
Rðxn; x1Þ � � � Rðxn;xnÞ
2664
3775
162 P. Esquivel et al.
where n indicates that the number of measurement points, and
Rðxi;xjÞ ¼1
N
XN
k¼1�uðxi; tkÞ�uðxj; tkÞ; i; j ¼ 1; . . . ; n (6:13)
In other words, the optimal basis is given by the eigenfunctions ji of (6.13)whose kernel is the autocorrelation function R x; x0ð Þ ¼ �u xj; tk
� ��u x0j; tk
� �D E.
6.1.2 Discrete Domain Representation
Time series are usually recorded in discrete form even though the underlyingprocess itself is continuous. For discretely sampled measured data, the integraltime average can be approximated by a sum over the set of sampled data points[1]. In this case, the vectors
xj ¼ �ujðxj; tkÞ ¼ �uðxj; t1Þ �uðxj; t2Þ � � � �uðxj; tNÞ� T
; j ¼ 1; :::; n (6:14)
represent a set of snapshots obtained from the observed data at n locations.The set of data can then bewritten as theN� n-dimension ensemblematrix,X [6]
X ¼ x1 � � � xn½ � ¼
�uðx1; t1Þ � � � �uðxn; t1Þ... . .
. ...
�uðx1; tNÞ � � � �uðxn; tNÞ
2664
3775 (6:15)
where each column corresponds to the response at a specific time.Typically, n 6¼ N, so X is generally rectangular. Under these assumptions,
the actual integral (6.10) can be written as Cj ¼ lj, where
Cij ¼ 1 NPN
k¼1 �uðxi; tkÞ�uðxj; tkÞ.
. Assuming the EOFs to be of the form
ji ¼PN
l¼1 wilxi where wi
l is a coefficient to be determined, the problem
of minimizing (6.2) can be recast as the problem of finding the largesteigenvalue of the linear equation
Cj ¼ lj (6:16)
where C is the autocorrelation (covariance) matrix defined as
C ¼ 1
NXTX ¼ 1
N
xT1 x1 xT1 x2 � � � xT1xn
xT2 x1 xT2 x2 � � � xT2xn
..
. ... . .
. ...
xTn x1 xTn x2 � � � xTn xn
266664
377775¼ 1
N
Xn
i¼1xi � umð ÞT xi � umð Þ (6:17)
6 Complex Empirical Orthogonal Function Analysis 163
The resulting covariance matrix C is a real, symmetric ðCij ¼ CjiÞ positive,and semi-definite matrix. Consequently, it possesses a set of orthogonal eigen-
vectors ji; i ¼ 1; :::; n, i.e.,
jTi jj ¼
�ij; i ¼ j
0; i 6¼ j
�
Using standard linear algebra techniques, the covariance matrix can be
expressed in the form
C ¼ ULVT (6:18)
where U and V are the matrices of right and left eigenvectors and
L ¼ diag½l1 l2 � � � ln�.The eigenvalues computed from (6.18) are real and nonnegative and can be
ordered such that l1 � l2 � � � � � ln � 0. The eigenvectors of Care called the
POMs and the associated eigenvalues are called the proper orthogonal vectors
(POVs).For practical applications, the number snapshots N can be rather large,
leading to a very large eigenvalue problem. There are at least two methods to
solve the eigenvalue problem (6.16) [6]: the direct method and the method of
snapshots.The direct method attempts to solve the eigenvalue problem involving the
N�N matrix directly using standard numerical techniques. This can be
computationally intensive if the number of observations is larger than the
number of observing locations or grid points. Themethod of snapshots, on the
other hand, is based on the fact that data vectors ui and the POD modes jl
span the same linear space. The latter is explored here.The next section describes in more detail the nature of the approximation
employed here to construct the statistical representation.
6.1.3 The Method of Snapshots
The method of snapshots is based on the fact that the data vectors ui and the
POD modes span the same linear space [6]. In this approach, we choose the
eigenfunctions j to be a linear combination of the snapshots:
ji ¼XN
l¼1wilxi (6:19)
where the coefficients wil are to be determined such that u maximizes
164 P. Esquivel et al.
maxj
1
N
Xn
j¼1
xj;j�� ��j;jh i (6:20)
These l functions are assembled into an n�Nmatrix, F, known as the modalmatrix. In matrix form Eq. (6.19) becomes
F ¼ XW (6:21)
where
F ¼" " " "j1 j2 � � � jn
# # # #
2
64
3
75; X ¼" " " "x1 x2 � � � xn
# # # #
2
64
3
75;W ¼" " " "w1 w2 � � � wn
# # # #
2
64
3
75
in which
w1 ¼
w11
w12
..
.
w1N
2666664
3777775; w2 ¼
w21
w22
..
.
w2N
2666664
3777775; . . . ; wn ¼
wn1
wn2
..
.
wnN
266664
377775
Substituting the expression (6.19) into the eigenvalue problem (6.16) gives
CXN
l¼1wilxi ¼l
XN
l¼1wilxi (6:22)
where Cij ¼ ð1=NÞ ð�ui; �ujÞ. This can be written as the eigenvalue problem ofdimension n
CW ¼ LW (6:23)
where
w ¼ w1 w2 � � � wn½ �
and L is a diagonal matrix storing the eigenvalues li of the covariance matrixC.In words, the first-order necessary optimality condition for j to provide amaximum in (6.20) is given by (6.16). This completes the construction of theorthogonal set j1 j2 � � � jnf g.
Once the modes are found using these equations, the flow field can bereconstructed using a linear combination of the modes
6 Complex Empirical Orthogonal Function Analysis 165
ukðxj; tkÞ ¼X1
i¼1aiðtÞjiðxÞ (6:24)
for some aiðtÞ 2 <, where the aiðtÞ are the time-varying amplitudes of the PODmodes jiðxÞ.
The truncated POD of uis
ukðxj; tkÞ ¼Xp
i¼1aiðtÞjiðxÞ þ " (6:25)
where p is the number of dominant modes and " is an error function.Having computed the relevant eigenmodes, the temporal behavior can be
evaluated as the inner product of the eigennmode (the POD mode ji) and theoriginal data. To ensure uniqueness of the solution, the normalization condi-tion of ji;jih i ¼ 1 is imposed.
The temporal coefficients are then expressed as
ai ¼ x;jih i= ji;jih i (6:26)
Note that the temporal modes are uncorrelated in time, i.e.,ðaiðtÞ; ajðtÞÞ ¼ �ijlj; where �ij ¼ 1 for i ¼ j; 0 else, and that the system (6.25) isoptimal in the sense that minimizes the error functions
"ðjÞ ¼Xp
l¼1�uðxj; tkÞ �
Xp
i¼1aiðtÞjiðxÞ
�����
����� (6:27)
It should also be stressed that no conditions are imposed on the data set; thedata can be a sample of a stationary process or a sample of a nonstationary process.
Equation (6.24) is called the Karhunen–Loeve decomposition and the set jj
are called the empirical basis [5].
6.1.4 Energy Relationships
The use of the POD method leads naturally to a discussion of truncationcriteria. Several techniques to derive truncated expressions have been proposedin the literature. Here, we choose to reduce the residual terms, R ¼ ", such thatthe mean square value
R ¼ Olpþ1Pni¼1 li
� �(6:28)
be as small as possible.
166 P. Esquivel et al.
Among all linear decompositions, the most kinetic energy possible for aprojection onto a given number of modes. Defining the total energy E, byE ¼ Trace ½Rðxi; xmÞ�; one obtains
E ¼Xp
i¼1li (6:29)
The associated percentage of total energy contributed by eachmode can thenbe expressed as
Ei ¼liPn
i¼1li
(6:30)
Typically, the order p of the reduced basis j such that the predeterminedlevel of the total energy E of the snapshot ensemble is captured, i.e., 99%. Thep-dominant eigenfunctions are then obtained as
Ppi¼1 liPnj¼1 lj
¼ 99% (6:31)
for the smallest integer p where E is an appropriate energy level.The key advantage of this technique is that allows extracting information
from short and often noisy time series without prior knowledge of the dynamicsaffecting the time series.
6.2 Interpretation of EOFs Using Singular Value Decomposition
A useful alternative method for estimating modal characteristics can bedeveloped based on the analysis of the response matrix X in (6.15).
Before outlining the procedure for singular value decomposition (SVD)analysis, we introduce some background information on singular valueanalysis.
6.2.1 Singular Value Decomposition
Let A be a real m � n matrix. The SVD theorem states that A can be decom-posed into the following form [7]:
A ¼ USVT (6:32)
6 Complex Empirical Orthogonal Function Analysis 167
where U ¼ col½u1 u2 � � � um� is an m � m orthonormal matrix VT ¼ U�1,S is an m � n pseudodiagonal and semi-positive definite matrix with diagonal
entries containing the singular values ,V ¼ col½v1 v2 � � � vn� and is an n� n
orthonormal matrix UT ¼ U�1. The columns of U and Vare called the left and
right singular vectors for A.Matrix S has the form
S ¼
�1 0 � � � 0 0 � � � 0
0 �2 0 � � � 0
..
. ... . .
. ... . .
. ...
0 0 � � � �m 0 � � � 0
2
66664
3
77775for m5n
or
S ¼
�1 0 � � � 0
0 �2 � � � 0
..
. ... . .
. ...
0 0 � � � �m
0 0 � � � 0
..
. ... . .
. ...
0 0 � � � 0
26666666666664
37777777777775
for m4n
Throughout this research, we will, consider only the case when m4n. The
diagonal entries of S, i.e., the �ii ¼ �i, can be arranged to be nonnegative and in
order of decreasing magnitude �1 � �2 � � � � � �m � 0.Equivalently, we can express the model as
A ¼ u1 � � � uk j ukþ1 � � � um½ �
�1 � � � 0 0
..
. . .. ..
. ...
0 � � � �k 0
0 � � � � � � 0
2
66664
3
77775
vT1
..
.
vTkvTkþ1
..
.
vTn
2666666666664
3777777777775
(6:33)
or
A ¼ Uj?U�
D 00 0
� �� V?V
�
168 P. Esquivel et al.
where D is the diagonal matrix of nonzero singular values and U and V are the
matrices of left and right singular vectors, respectively, corresponding to the
nonzero singular values.It is clear that only the first r of the u’s and v’s make any contribution to A,
and can be expressed as an outer product expansion
A ¼Xr
i¼1�i uiv
Ti
� �(6:34)
where the vectors ui and vi are the columns of the orthogonal matricesU and V,
respectively. Techniques to compute the POMs based on SVD are next discussed.
6.2.2 Relation with the Eigenvalue Decomposition
An interesting interpretation of the POD modes can be obtained from the
singular value analysis of the response matrix X.Using the notation in Section 6.1.2 let the response matrix be given by
X ¼
�uðx1; t1Þ � � � �uðxn; t1Þ... . .
. ...
�uðx1; tNÞ � � � �uðxn; tNÞ
2
664
3
775 (6:35)
The SVD of the response matrix X may be written in compact form as
X ¼ USVT (6:36)
where U is an orthonormal N � N matrix whose columns are the left singular
vectors of X, S is N � n matrix containing the singular values of X along the
main diagonal and zeros elsewhere, andV is an n� n orthonormalmatrix whose
columns correspond to the right singular vectors of X. The response matrix, X,
is complex and symmetric and possesses a set of orthogonal singular vectors
with positive singular values.In terms of the notation above for SVD, it can be seen directly from (6.18)
that the correlation matrix defined previously is given by
XXT ¼ USVð Þ USVð ÞT¼ US2UT (6:37)
and
XTX ¼ USVð ÞT USVð Þ ¼ VS2VT (6:38)
6 Complex Empirical Orthogonal Function Analysis 169
Hence, (6.32) becomes
XXTU ¼ U
l1l2
. ..
ln
2
66664
3
77775¼ U
�21�22
. ..
�2n
2
66664
3
77775(6:39)
It follows immediately from (6.32), (6.33), (6.34), the singular values ofX are
the square roots of the eigenvalues ofXXT orXTX [2, 8]. In addition, the left and
right eigenvectors of X are the eigenvectors of XXT and XTX, respectively. Also
of interest, the trace of (6.39) is given by
XTX ¼Xr
i¼1�2i
The POMs, defined as the eigenvectors of the sample covariancematrixC are
thus equal to the left singular vectors ofX. The POVs, defined as the eigenvalues
of matrix C are the squares of the singular values divided by the number of
samples N.
6.3 Numerical Computation of POMs
In this section, a step-by-step description of the algorithm used to extract modal
information is presented. The procedure adopted to compute the POMs can be
summarized as follows:
1. Given an ensemble of measurements of a nonstationary process, computethe response matrix X. Form the complex time series matrix X ¼ Xþ jXH,where XH is the Hilbert transform of X
2. Compute the singular vectors U;V and the corresponding singular values �.3. Determine the time evolution of the temporal modes, ai. Extract standing
and propagating features using the complex SVD (singular value decompo-sition) formulation.
Figure 6.1 illustrates the proposed algorithm.For different events recorded at the same location, statistical averaging can
be employed to take advantage of the statistics of the data. In this case, the
snapshots can be thought of realizations of random fields generated by some
kind of stochastic process.The processing steps are detailed in the sections that follow.
170 P. Esquivel et al.
6.4 Complex Empirical Orthogonal Function Analysis
Empirical orthogonal function analysis of data fields is commonly carried outunder the assumption that each field can be represented as a spatially fixedpattern of behavior. This method, however, can not be used to for detection ofpropagating features because of the lack of phase information. To fully utilizethe data, a technique is needed that acknowledges the nonstationarity andbehavior of the time-series data.
Let uðxj; tkÞ be a space–time scalar field representing a time series, wherexj; j ¼ 1; . . . ; n is a set of spatial variables on a space �k and tk; k ¼ 1; . . . ;N isthe time at which the observations are made. Provided u is simple and squareintegrable, it has a Fourier representation of the form [9]
uðxj; tkÞ ¼X1
m¼1½ajðmÞð!Þ cosðm!tkÞ þ bjðmÞð!Þ sinðm!tkÞ� (6:40)
where ajðmÞð!Þ and bjðmÞð!Þare the Fourier coefficients defined as
ajðmÞ ¼1
p
Zp
�p
uðxj; tkÞ cos m!tkð Þ d!
bjðmÞ ¼1
p
Zp
�p
uðxj; tkÞ sin m!tkð Þ d!
Fig. 6.1 Conceptual view of the proposed algorithm
6 Complex Empirical Orthogonal Function Analysis 171
This allows description of traveling waves. Equation (6.40) can be rewrittenin the form
ucðxj; tkÞ ¼X1
m¼1cjðmÞð!Þe�im!tk (6:41)
where cjðmÞð!Þ ¼ ajðmÞð!Þ þ ibjðmÞð!Þ [1] and i ¼ffiffiffiffiffiffiffi�1p
. Expanding (6.41) andcollecting terms gives
Ucðxj; tkÞ ¼X1
m¼1ajðmÞð!Þ cosðm!tkÞ þ bjðmÞð!Þ sinðm!tkÞ� � �
þ iX1
m¼1bjðmÞð!Þ cosðm!tkÞ � ajðmÞð!Þ sinðm!tkÞ� � �
¼ uðxj; tkÞ þ iuHðxj; tkÞ
(6:42)
where the real part of U is given by (6.40) and the imaginary part is the Hilberttransform of uðxj; tkÞ[10]
uHðxj; tkÞ ¼ �1
p
Z1
�1
h u xj; tk� ��
tk � xjdx (6:43)
This represents a filtering operation upon uðxj; tkÞ in which the amplitude ofeach Fourier spectral component remains unchanged while its phase isadvanced by p/2. The eigenvectors here are complex and can be expressedalternatively as a magnitude and phase pair.
In the proposed formulation u can be estimated more efficiently by perform-ing a filtering operation on u itself. Equation (6.43) can be rewritten in the formof a convolution as
uHðxj; tkÞ ¼XL
‘¼�Luðxj; tk � ‘Þhð‘Þ; L ¼ 1 (6:44)
where h sis a convolution filter with unit amplitude response and 908 phase shift.In practice, a simple filter that has the desired properties of approximate unit
amplitude response and p/2 phase shift is given by [11]
hðlÞ ¼2pl sin
2 pl2
� �; l 6¼ 0
0; l ¼ 0
((6:45)
where �L � l � L. As L!1, Equation (6.45) yields an exact Hilberttransform.
172 P. Esquivel et al.
In what follows we discuss the extension of the above approach to compute
standing and propagating features.
6.4.1 Complex EOF Analysis
Drawing on the above approach, an efficient formulation to compute complextime-dependent POMs has been derived. Following Susanto et al. [12] assumethat X is a j (spatial points) by k (temporal points) ensemble matrix. From thepreceding results, it follows that
X ¼ USVH (6:46)
where VH is the conjugate transpose of V, the superscript H denotes a Hermi-tian matrix, and we assume that
UHU ¼ I
VHV ¼ I
Now, it can be easily verified that
XXH ¼ USVH� �
USVH� �H¼ USSTUH
XHX ¼ USVH� �H
USVH� �
¼ VSTSVH(6:47)
where ST denotes the transpose of S. As is apparent from Eq. (6.46), the
columns of U are the eigenvectors of X XH� T
, and that the columns of V are
the eigenvectors of XH� T
X. The n singular values on the diagonal of S are the
square roots of the nonzero eigenvalues of both X XH� T
and XH� T
Xwhere n is
the rank of X.Once the spatial eigenvectors are calculated, their corresponding time evolu-
tion is given by the time series AiðtÞ which is obtained by projecting the timeseries X onto the proper eigenvector ji, and summing over all locations:
AiðtÞ ¼Xn
j¼1Xðxj; tkÞjiðxÞ (6:48)
The original complex data field, Xðx; tÞ, can be reconstructed by adding thisproduct over all modes, i.e.,
Xðx;tÞ ¼Xn
i¼1AiðtÞjH
i ðxÞ (6:49)
6 Complex Empirical Orthogonal Function Analysis 173
Using the complex SVD, it is possible to compute the spatial amplitude andspatial and temporal phase functions as discussed below.
6.4.2 Analysis of Propagating Features
The time-dependent complex coefficients associated with each eigenfunction canbe conveniently split into their amplitudes and phases. From the complex EOFanalysis in (6.49), the ensemble of data can be expressed as the complex expan-sion [13, 14]
Xðx; tÞ ¼Xn
i¼1RiðtÞffqi SiðxÞfffi (6:50)
where RiðtÞ is the complex temporal amplitude function, SiðxjÞ is the complexspatial mode or eigenvector, and q, f are the phase functions corresponding toRiðtÞ and SiðxjÞ. These phase functions describe the propagation characteristicsof the ith mode.
Equation (6.50) can be rewritten as
uðxj; tÞ ¼Xn
i¼1
��RiðtÞ����SiðxÞ
��e j½qRiðtÞþfSi
ðxÞ�(6:51)
This effectively decomposes the data into spatial and temporal modes.Four measures that define possible moving features in uðx; tÞ can then be
defined [9].
1. Spatial distribution of variability associated with each eigenmode2. Relative phase of fluctuation3. Temporal variability in magnitude4. Variability of the phase of a particular oscillation
The following definitions introduce these concepts.
Definition 6.1 (Spatial amplitude function, SiðxÞ) The spatial amplitude func-tion, SiðxÞ, is defined as
SiðxÞ ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffijHi ðxÞjiðxÞ
q(6:52)
This function shows the spatial distribution of variability associated witheach eigenmode.
Definition 6.2 (Spatial phase function, ji) This function shows the relative phasefluctuation among the various spatial locations where u is defined and is given by
174 P. Esquivel et al.
fiðxÞ ¼ tan�1ImfjiðxÞgRefjiðxÞg
� �(6:53)
This measure, for which an arbitrary reference value must be selected,
varies continuously between 0 and 2p. Together, Eqs. (6.52) and (6.53) give
a measure of the space distribution of energy and can be used to identifythe dominant modes and their phase relationships. Further, for each domi-
nant mode of interest, a mode shape can be computed by using the spatial
part of (6.50).
Definition 6.3 (Temporal amplitude function, Ri) Similar to the description ofthe spatial amplitude function in (6.52), the temporal amplitude function, Ri
can be defined as
RiðtÞ ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiAH
i ðtÞAiðtÞq
(6:54)
This function gives ameasure of the temporal variability in themagnitude of the
modal structure of the field u. As described later, the general form of theseequations is very amenable to computational analysis.
Definition 6.4 (Temporal phase function, qi) This temporal variation of phaseassociated with uðx; tÞ is given by
qiðtÞ ¼ tan�1ImfAiðtÞgRefAiðtÞg
� �(6:55)
For the simple case of a sinusoidal wave with fixed frequency and wave
number, �iðtÞ ¼ !t. In the more general (and interesting case), the space
derivative of the phase and frequency of the modal components can be calcu-lated from
ki ¼ dðfiÞ=dðxÞ
wi ¼ dðqiÞ=dðtÞ
ci ¼ wi=ki
(6:56)
where ci is the phase velocity of the function.Equations (6.52), (6.53), (6.54), (6.55), (6.56) provide a complete character-
ization of any propagating effects and periodicity in the original data field
which might be obscured by normal cross-spectral analysis. Finally, it mightbe remarked that, in the special case of real analysis, these expressions simplify
to the normal definitions.
6 Complex Empirical Orthogonal Function Analysis 175
6.5 Application to Time Synchronized Measured Data
To test the ability of the method to analyze complex oscillations, we analyze
data from time-synchronized measurements. The data used for this study were
recorded by phasor measurement units (PMUs) over a 4 s window during a real
event in northernMexico. A brief description of the data follows.More detailed
information on system measurements can be found in [4, 15].At local time 06:27:42 in the early hours of January 1, 2004, undamped
oscillations involving frequency, voltage, and power were observed throughout
the northern systems of the Mexican Interconnected System (MIS). The main
event that originated the oscillations was a failed temporary interconnection of
the Northwest regional systems to the MIS through a 230 kV line betweenMZD
(Mazatlar DOS) and TTE (Tres Estrellas) substations. It is noted that, prior to
this oscillation incident, the northwestern system operated as an electrical island.Oscillations in the northern systems with periods about 0.61, 0.50, and
0.27 Hz persisted for approximately 1.2 s before the northwestern system was
disconnected from the MIS. During the time interval 06:27:42–06:28:54 the
system experienced severe fluctuations in frequency, power, and voltage result-
ing in the operation of protective equipment with the subsequent disconnection
of load, independent generation, and major transmission resources.
Fig. 6.2 Schematic of the MIS system showing the location of the observed oscillations.Measurement locations are indicated by shaded circles
176 P. Esquivel et al.
Figure 6.2 shows a geographical diagram of the MIS showing the PMU
locations and the location of the initiating event. For demonstration purposes,
three buses spread across the system are selected. To allow comparison with
previous work, bus frequency signals are used in the analysis.Figure 6.3 is an extract from PMU measurements of this event showing the
observed oscillations of selected bus frequencies. For the purpose of further
comparison to EOF analysis, the relevant time interval of concern is zoomed in.
Systemmeasurements in this plot demonstrate significant variability suggesting
a nonstationary process in both space and time. The most prominent variations
occur in the interval during which the oscillation start at 06:27:42 and the
interval in which the operating frequency is restored to the nominal condition
(60 Hz) by control actions (06:28:21).As discussed later, during this period the system experiences changes in
frequency (amplitude) content and mode shapes.
Fig. 6.3 Time traces of recorded bus frequency swings recorded on January 1, 2004 and detailof the oscillation buildup
6 Complex Empirical Orthogonal Function Analysis 177
As a first step toward the development of a POD basis, the observed records
are placed in a complex data matrix, Xðxj; tkÞ, as
Xðxj; tkÞ ¼ fHðtÞ; fMZDðtÞ; fTTEðtÞ½ � ¼
uðx1; t1Þ � � � uðxn; t1Þ... . .
. ...
uðx1; tNÞ � � � uðxn; tNÞ
2
664
3
775
where j ¼ 1; 2; 3 is the spatial position index (grid location), tis the time, andN is
the number of data points in the time series. For our simulations, 2021 snap-
shots are available, representing equally spaced measurements at three different
geographical locations.Each time series is then augmented with an imaginary
component to provide phase information and the EOF method is employed to
approximate the original data by a general, nonstationary, and correlatedfrequency model.
The following subsections describe the application of complex empirical
orthogonal function analysis to examine the temporal and spatial variability
of measured data.
6.5.1 Construction of POD Modes via the Method of Snapshots
The method of snapshots was applied to derive a spatiotemporal model of the
oscillations. In order to improve the ability of the method to capture temporalbehavior, the individual time series are separated into their time-varying mean
and fluctuating components. By separating the data into their mean and fluc-
tuating components, EOF analysis is able to selectively determine the temporal
behavior of interest. The method may also help reduce the detrimental effects of
crossing events, spatial aliasing, and random contamination.In the application of the proposed method to measured data, we assume that
each signal can be construed as a superposition of fast oscillations on top of a
slow oscillation (the time-varying instantaneous mean) [16]. The slow oscilla-
tion essentially captures the nonlinear (and possibly time-varying) trend while
the slow oscillations are the fluctuating parts.A two-stage analysis technique based on wavelet shrinkage is proposed to
determine the temporal properties of time-synchronized information. In thefirst step, the original system time histories are decomposed into their time-
varying mean speeds and fluctuating speeds through wavelet shrinkage. More
formally, the recorded time series are decomposed into their time-varying mean
frequencies, umðxj; tkÞ; and nonstationary fluctuating components, �uðxj; tkÞ; asfollows:
uðxj; tkÞ ¼ umðxj; tkÞ � �uðxj; tkÞ (6:57)
178 P. Esquivel et al.
In the second stage, complex EOF analysis is applied to the fluctuating field
to decompose the spatiotemporal data into orthogonal temporal modes.Following Donoho [17], the two-stage approach can be used to reconstruct
an unknown function f from noisy data di ¼ fðtiÞ þ �zi; i¼ 0; . . . ;n� 1;wheredi is the observed data point (the noise-contaminatedmeasurement point), zi is astandard Gaussian white noise, ti are equispaced data points, and �is a noiselevel. The first stage is to find the estimates fð:Þ ¼ Tðy; dyÞ ð:Þ, where Tðy; dyÞ ð:Þis a suitable reconstruction formula with spatial smoothing parameters � and dyis data-adaptive choice of the spatial smoothing parameter. The separatingprocedure is carried out using the wavelet shrinkage method because of itsability to capture fast changes in the data.
The adopted separation/identification procedure can be summarized asfollows:
1. Expand the data uðxj; tkÞ into wavelet series. Using the wavelet decomposi-tion structure, estimate level-dependent thresholds for signal compressionusing suitable thresholding approach
2. Obtain a denoised compressed function, um, representing the instantaneousmean, via the wavelet shrinkage, and
3. Compute the fluctuation of the signal by calculating �uðxj; tkÞ ¼ uðxj; tkÞ �umðxj; tkÞ; where um is the time-varying mean of the signal. Among theexisting techniques, we use the Birge–Massart strategy in [18].
The algorithm is simple to implement and the computational requirementsare small.
6.5.2 Spatiotemporal Analysis of Measured Data
Based on the analytical procedure outlined in Section 6.5.1, the complex EOFmethod was used to determined dynamic trends and to analyze phase relation-
ships. In this procedure, the measured data are augmented with an imaginarycomponent defined by its Hilbert transform, and the temporal patterns areextracted by using the procedure in Section 6.4.
Complex EOF analysis was performed on the original time series, and thenonlinearly detrended (wavelet shrinkage model) time series. The analysis pre-sented here uses the Daubechies wavelet with a fixed decomposition level. Alevel-dependent threshold is then obtained using a wavelet coefficients selection
rule based on the Birge–Massart strategy [18].As seen in Fig. 6.4, this model very effectively describes the long-term
behavior of the data while also capturing transient fluctuations. This, in turn,results in improved characterization of system behavior.
Spectral analysis results for the leading POM in Fig. 6.5 show that the mainpower is concentrated in oscillations with frequencies about 0.61, 0.50, and0.27 Hz, which can be associated with major inter-area modes in the system [4].
6 Complex Empirical Orthogonal Function Analysis 179
Small peaks in Fig. 6.5 may indicate nonlinear interactions between frequency
components. From Fig. 6.5 it is also seen that the time-varying instantaneous
mean (bold gray line) acts as a high-pass filter. This approach can be used to
separate the slow from the fast components as well as to improve the numerical
accuracy of the method. This is discussed further in the next two subsections.Figure 6.6 presents the corresponding eigenspectrum of these modes com-
puted to capture 99.9% of the signal’s energy. In these plots, the horizontal axis
shows the number of modes required to attain 99% of the averaged total
energy; the vertical axis shows the energy in (6.13) captured by each POM.
The data set being analyzed corresponds to bus frequency measurements
observed with the PMUs at various geographical locations.
100 120 140 160 180 200 220 240 260 280 30058.5
5959.5
6060.5
Fre
quen
cy [H
z]
100 120 140 160 180 200 220 240 260 280 30058.5
5959.5
6060.5
Fre
quen
cy [H
z]
100 120 140 160 180 200 220 240 260 280 30059.9
60
60.1
60.2
60.3
Time (ms)
Fre
quen
cy [H
z]
Original signalTime−varying instantaneous mean
Fig. 6.4 Time-varying means and fluctuating components of the recorded bus frequencysignals. The time series are smoothed using wavelet shrinkage
0 0.2 0.4 0.6 0.8 1 1.2
–100
–50
0
50
Frequency (Hz)
Mag
nitu
de (
dB)
POD of original signalsPOD with instantaneous mean removedPOD of instantaneous mean0.61 Hz
0.49 Hz0.27 Hz
Fig. 6.5 Comparison of theFourier transform spectrumof the original signal and thespectra constructed with theinstantaneous meanremoved
180 P. Esquivel et al.
Complex EOF analysis of the synchronized measurements of frequency in
Fig. 6.6 shows that wide-area system dynamics is well represented by three
modes; the two leading modes together account for 96.5% of the total energy.
Individually, these modes account for 72, 24.5, and 3.5% of the energy (see
Fig. 6.6 caption for details).Figure 6.7 compares the reduced solution using the POD basis functions to
the full solution obtained from the measured data for the Hermosillo signal in
1 2 30
10
20
30
40
50
60
70
80
Number of modes
Ene
rgy
(%)
Real EOF formulationComplex EOF formulation
Mode Energy (%) Cumulative energy (%)
1 72 722 24.5 96.53 3.5 100
Fig. 6.6 Energy captured as a function of the number of modes. The percentage of energylocated in the jth mode is measured by E ¼ 100lj 100lj
� Pni¼1 li
120 140 160 180 200 220 24059
59.5
60
60.5
61
Freq
uenc
y (H
z)
120 140 160 180 200 220 24059
59.5
60
60.5
61
Freq
uenc
y (H
z)
120 140 160 180 200 220 24059
59.5
60
60.5
61
Time (ms)
Freq
uenc
y (H
z)
Original time seriesPOM 1
Original time seriesPOM 1 + POM 2
Original time seriesPOM 1+ POM 2 + POM 3
Energy = 44 %Cumulative = 44 %
Energy = 55 %Cumulative = 99 %
Energy = 1 %Cumulative = 100 %
Fig. 6.7 Reconstruction of the original data using the three leading POMs. Solid linesrepresent the original time series and dotted lines represent the composite oscillation obtainedby adding the temporal modes
6 Complex Empirical Orthogonal Function Analysis 181
Fig. 6.4. As observed in this plot, using only three modes we are able toaccurately approximate the measured data over the entire observation period.The agreement between the reduced order model and the observed behaviorillustrates the high degree of accuracy that is possible with a simplified model.
On the basis of these results, we conducted detailed analysis aiming atdisclosing hidden information in the data. For clarity of exposition, the analysisof temporal and spatial patterns will be presented separately.
6.5.3 Temporal Properties
In order to reveal the hidden wave signatures in the time series, we examine bothamplitude and temporal patterns in system behavior in the light of complexorthogonal function analysis. Figure 6.8 illustrates the temporal evolution(amplitude and phase) of the dominant mode.
These functions display a number of interesting features. As discussed below,the analysis identifies two periods of interest; a transient period associated withthe interconnection of the systems (06:27:42–06:28:21) and a nearly stationaryperiod in which the frequency of the interconnected system is restored to itsnormal value (06:28:54–06:29:39).
The first interval manifests particularly strong temporal activity as can be seenin Fig. 6.8a. In interpreting these results, we remark that break or changes in thetemporal functions may signal different physical regimes or control actions.
An examination of the temporal phase in Fig. 6.8b, on the other hand,reveals a nonstationary behavior in which the phase (frequency) contentchanges with time. Here, the slope of the spatial phase function represents theinstantaneous frequency. The slowly increasing trends indicate periods ofessentially constant frequencies.
6.5.4 Frequency Determination from Instantaneous Phases
Additional insight into the frequency variability of the observed oscillations canbe obtained from the analysis of instantaneous frequencies. Recognizing thatthe instantaneous frequency is the time derivative of the temporal phase func-tion, �, the instantaneous frequencies can be estimated from (6.55) for eachmode of concern.
The study focuses on POM 1 which is the mode that captures most of thevariability in the signal. Figure 6.9 gives the instantaneous frequency of POM 1for the interval of interest in this study. Also plotted, is the instantaneous meanfrequency (nonlinear trend) determined using wavelet shrinkage analysis above.
Nonstationary features are evident in this plot. Analysis of these plots showstwo modal components: a 0.27 Hz component associated with the steady-statebehavior of the system, and a 0.64 Hz component associated with the transient
182 P. Esquivel et al.
system fluctuation following the system interconnection. The 0.27 Hz compo-
nent captures the slow ambient swings previous to the onset of system oscilla-
tions and the steady behavior of the system. The results are consistent with
those based on nonstationary analysis of the observed oscillations giving valid-
ity to the results.
120 140 160 180 200 220 2400
0.2
0.4
0.6
0.8
1
1.2
Time (ms)
Mag
nitu
de 06:27:42
06:28:06 06:28:21
06:28:54 06:29:39
a) Temporal amplitude
120 140 160 180 200 220 240−200
−150
−100
−50
0
50
100
150
200
Time (ms)
Phas
e (d
egre
es)
06:27:42
06:28:06
06:28:21
06:28:54
06:29:39
b) Temporal phase
Fig. 6.8 Temporal patterns of variability associated with the dominant mode
6 Complex Empirical Orthogonal Function Analysis 183
6.5.5 Mode Shape Estimation
One of the most attractive features of proposed technique is its ability to detectchanges in the shape properties of critical modes arising from topology changes
and control actions. Changes in mode shape may indicate changes in topologyor changes in load/generation and may be useful for control decisions and the
design of special protection systems. This is a problem that has been recentlyaddressed using spectral correlation analysis [19].
Using the spatial phase and amplitude, the phase relationship between key
system locations (the mode shapes) can be determined. In this analysis, wedisplay the complex value as a vector with the length of its arrow proportional
to eigenvector magnitude and direction equal to the eigenvector phase.Figure 6.10 shows the mode shape for the three intervals of interest above
(06:27:42–06:28:06, 06:28:06–06:28:21, and 06:28:21–06:28:54 ) computed using
the spatial function (6.50). It is interesting to note that the dominant POMmode shape changes with time. The effect is more pronounced for the time
interval 06:28:06–06:28:21 in which several control actions take place in the
system. This information may be useful to identify the dominant generatorsinvolved in the oscillations, and ultimately devise control mechanisms to damp
the observed oscillations.These results are in general agreement with previously published results
based on real EOF analysis and Prony results [4]. The new results, however,
provide clarification on the exact phase relationships between key systemmeasurements as a function of time.
120 140 160 180 200 220 240−1
−0.5
0
0.5
1
1.5
2
Time (ms)
Freq
uenc
y (H
z)Instantaneous frequencyTrend
0.64 Hz
0.27 Hz
Fig. 6.9 Instantaneous frequency of POM 1. Time interval 06:27:42–06:29:39
184 P. Esquivel et al.
6.5.6 Energy Distribution
In the previous section it was shown that a linear combination of an individual
eigenmode can accurately reconstruct the temporal behavior of simultaneous
measurements at different geographical (spatial) locations. A key related ques-
tion of interest is that of finding a small number of measurements that will
provide a good estimation of the entire field of interest.Based on the decomposed EOFs, complex EOF analysis was used to deter-
mine the locations with the most energy. Figure 6.11 shows the participation of
each location to the total energy of the record. The x-axis shows spatial sensor
location and the y-axis shows the energy value. From Fig. 6.11, it is evident that
0.5
1
30
210
60
240
90
270
120
300
150
330
180 0
0.5
1
30
210
60
240
90
270
120
300
150
330
180 0
0.5
1
30
210
60
240
90
270
120
300
150
330
180 0
a) 06:27:42−06:28:06 b) 06:28:06−06:28:21
c) 06:28:21−06:28:54
Hermosillo
Mazatlan Dos
Tres Estrellas
Fig. 6.10 Mode shape of POM 1 for various time intervals of interest
Fig. 6.11 Energy distribution
6 Complex Empirical Orthogonal Function Analysis 185
modes 1 and 2 are quite prominent at the Mazatlan Dos and Hermosillosubstations while mode 3 is more strongly evident at the Tres Estrellas substa-tion. This is consistent with conventional analysis (not shown). However, theproposed approach provides an automated way to estimate mode shapes with-out any prior information of the time intervals of interest.
6.6 Concluding Remarks and Directions for Future Research
In this chapter, a new method of temporal representation of nonstationaryprocesses in power systems has been presented. Complex empirical orthogonalfunction analysis provides an efficient and accurate way of looking at thetemporal variability of transient processes while at the same time providingspatial information about each one of the dominant modes with no a prioriassumption on either spatial or temporal behavior. The main advantage of thisapproach is its ability to compress the variability of large data sets into thefewest possible number of temporal modes.
Complex empirical orthogonal function analysis is shown to be a usefulmethod for identifying standing and traveling patterns in wide-area systemmeasurements. Using wide-area frequency information, spatiotemporal analy-sis of time-synchronized measurements shows that transient oscillations maymanifest highly complex phenomena, including nonstationary behavior.Numerical results show that the proposed method can provide accurate estima-tion of nonstationary effects, modal frequency, time-varying mode shapes, andtime instants of intermittent transient responses. This information may beimportant in determining strategies for wide-area control and special protectionsystems. The identified systemmodes from the decomposition may also serve toreveal relevant, but unexpected structure hidden in the data such as that result-ing from short-lived transient episodes. Other issues such as the effect ofnumerical approximations on modal estimates will be investigated in futureresearch.
A generalization of this theory is also needed to treat statistical data from anensemble of nonstationary oscillations. This is an aspect that warrants furtherinvestigation. Finally, the generalization of the proposed technique to deter-mine the most suitable locations for phasor measurement devices and theanalysis of modal coherency are topics worthy of further investigation.
References
1. Philip Holmes, et al., Turbulence, coherent structures, dynamical systems and symmetry.New York: Cambridge University Press, 1996.
2. Gaetan Kerschen, et al., ‘‘The method of proper orthogonal decomposition for dynamicalcharacterization and order reduction of mechanical systems: An overview,’’ NonlinearDynamics, vol. 41, 2005, p. 147.
186 P. Esquivel et al.
3. S. S. Ravindran, ‘‘Reduced-order controllers for control of flow past an airfoil,’’ Interna-tional Journal for Numerical Methods in Fluids, vol. 50, 2006, p. 531.
4. A. R. Messina, et al., ’’Extraction of Dynamic Patterns from Wide-Area Measurementsusing Empirical Orthogonal functions,’’ IEEE Trans. on Power Systems, vol. 22, no. 2,May 2007, p. 682.
5. G. Kerschen, et al., ‘‘The proper orthogonal decomposition for characterization andreduction of mechanical systems,’’ Nonlinear Dynamics, vol. 41, 2005, p. 147.
6. Lawrence Sirovich, Turbulence and the dynamics of coherent structures,’’ Quarterly ofApplied Mathematics,vol. XLV, no. 3, October 1987, p. 561.
7. Dan Kalman, ‘‘A singularly valuable decomposition: The SVD of a matrix,’’ The CollegeMathematics Journal, vol. 27, no. 1, January 1996, p. 2.
8. G. Kerschen, et al. ‘‘Physical interpretation of the proper orthogonal modes using thesingular value decomposition,’’ Journal of Sound and Vibration, vol. 249, no. 5, 2002,p. 849.
9. T. P. Barnett, ‘‘Interaction of the Monsoon and pacific trade wind system at interannualtime scales. Part I: The Equatorial zone,’’Monthly Weather Review, vol. 111, 1983, p. 756.
10. S.L. Hahn, Hilbert transforms in signal processing, The Artech House, Signal ProcessingLibrary, 1996.
11. A. V. Oppenheim, and R. W. Shafer, Discrete-Time Signal Processing, 2nd Edition,Prentice Hall, 1998.
12. R. Dwi Susanto, et al. ‘‘Complex Singular Value Decomposition Analysis of EquatorialWaves in the Pacific Observed by TOPEX/Poseidon Altimeter,’’ Journal of Atmosphericand Oceanic Technology, vol. 15, 1998, p. 764.
13. J. D. Horel, ‘‘Complex principal component analysis: Theory and examples,’’ Journal ofclimate and Applied Meteorology, vol. 23, 1984, p. 1660.
14. James K. Kaihatu, et al., ‘‘Empirical orthogonal function analysis of ocean surfacecurrent suing complex and real-vector methods, ‘‘Journal of Atmospheric and OceanTechnology,’’ August 1998, p. 927.
15. A. R.Messina, et al., Leading-Edge Electric Power System Research, CianM. O’Sullivan(editor), Nova Science Publishers, Inc., 2008. New York.
16. L. Chen, et al.,’’ Proper orthogonal decomposition of two vertical profiles of full-scalenonlinearity downburst wind speeds,’’ Journal of Wind Engineering and Industrial Aero-dynamics, vol. 93, 2005, p. 187.
17. Davd L. Donoho, and Iain M. Johnstone, ‘‘Ideal spatial adaptation by wavelet shrink-age,’’ Biometrika, vol. 81, no. 3, 1994, p. 425.
18. M. Misiti, et al. ‘‘Wavelet Toolbox for use with Matlab,’’ The Mathworks Inc., 1996.19. Daniel J. Trudnowski, ‘‘Estimating Electromechanical Mode Shape from Synchrophasor
measurements,’’ IEEE Trans. on Power Systems, vol. 23, no. 3, August 2008, p. 1188.
6 Complex Empirical Orthogonal Function Analysis 187
Chapter 7
Detection and Estimation of Nonstationary Power
Transients
Gerard Ledwich, Ed Palmer, and Arindam Ghosh
Abstract This chapter looks at issues of non-stationarity in determining whena transient has occurred and when it is possible to fit a linear model to anon-linear response. The first issue is associated with the detection of loss ofdamping of power system modes. When some control device such as an SVCfails, the operator needs to know whether the damping of key power systemoscillation modes has deteriorated significantly. This question is posed here asan alarm detection problem rather than an identification problem to get a fastdetection of a change. The second issue concerns when a significant disturbancehas occurred and the operator is seeking to characterize the system oscillation.The disturbance initially is large giving a nonlinear response; this then decaysand can then be smaller than the noise level ofnormal customer load changes.The difficulty is one of determining when a linear response can be reliablyidentified between the non-linear phase and the large noise phase of thesignal.The solution proposed in this chapter uses ‘‘Time-Frequency’’ analysis tools toassistthe extraction of the linear model.
7.1 Introduction
For much of the time a power system operates in near steady state with gradualchanges in the overall loading and pattern of generation. There are continualchanges as different customers switch loads ON and OFF but when the systemamalgamates the loads of millions of customers the effect is one of continual smallperturbations. Occasionally, dramatic disturbances happen, lightning causes afault on a line which is then tripped, a boiler tube failure causes a generator togo off-line or a secondary systems fault causes a static VAR compensator (SVC) togo off-line. Even rarer is the cascading failure such as in the USA/Canada in 2003.
G. Ledwich (*)Faculty of Built Environment and Engineering, Queensland University of Technology,Brisbane, Australiae-mail: [email protected]
A.R. Messina, Inter-area Oscillations in Power Systems,Power Electronics and Power Systems, DOI 10.1007/978-0-387-89530-7_7,� Springer ScienceþBusiness Media, LLC 2009
189
Here lines sagging into trees tripped, which shifted load onto other lines whichthen tripped under overload, generators became isolated, and there was massiveloss of supply for millions of customers. For a correct response to a transient, itcritical to know how severe it is and whether drastic response is required.
To ensure that the control response is appropriate the operators need toknow whether a change in performance has occurred and what are the condi-tions of the new level the system is moving to. This chapter explores two aspectsof small-signal models: the detection of change in the system and the identifica-tion of the new operating parameters.
7.2 Modal Damping Change Detection
When a power system is operating normally there are continual customer loadchanges which excite modal responses. If the damping parameters of the powersystem suddenly deteriorate due to the loss of a control item such as an SVCwith stabilizer signals, then the response of the modes will change. In particular,a reduction of damping will increase the level of response in one or more modes.While the level of customer changes exciting the system will change for thedifferent portions of the day, it is not expected that there will be significantshort-term variations in the level of customer variation so an increase of energyof a mode can be used to signal a reduction of damping. If there was a majorpower system transient that stayed in the linear domain then Prony techniqueswould be applicable to perform fast identification of the modal parameters. Ifthe loss of the control unit itself did not generate a significant transient, then theonly route to modal identification would be from the response of the systemdynamics to the background customer load variations. For accurate identifica-tion of the system, changes using this technique can take from 10 to 30 min ofdata to reach different levels of confidence in the identification. If the systemchange was to a marginally stable operating point this would be too long beforedetection. The approach in this section of the chapter is to develop a fastdetection of significant change rather than precise estimation of the parameters.This will be able to alert operators to the deterioration while more preciseevaluation of the change is being identified.
In this section, there are two approaches to this detection of change. The firstapproach is to monitor the energy of the signal around the known systemmodes. When the relative energy level of the mode rises beyond some threshold,it indicates a significant loss of damping relevant to that mode. Because thedamping level can be mapped to the magnitude of response of a mode thenthreshold levels can be identified and alarming processes set.
The second approach is based on Kalman filtering. If the system has beenoperating in one condition for an extended time then the modal parameters canbe identified and an ideal whitening filter can be identified. The whitening filterwould process measurements such as angle difference from center of area, and
190 G. Ledwich et al.
when the modes are identified, the error between the measurement and theKalman filter prediction of measurement would be random white noise. Nobetter predictor of the next measurement would be possible. Now, if there hasbeen a significant reduction of modal damping, the whitening filter would nolonger be perfect and the residual error would no longer be white. In particular,when amodal response magnitude increased due to loss of damping, the residuewould show a rise of energy at this frequency.
There is a strong interrelation between the two approaches but the theore-tical advantage of the Kalman approach is that the properties of the white noiseresidue are well characterized so the degree of confidence in showing thatdamping change has occurred with a precise degree of confidence or specifiedprobability of false alarms becomes standard statistical analyses.
7.2.1 Energy Detection Approach
This energy detection approach seeks to characterize the expected energy in theresponse of the power system and thus detect sudden changes in damping. Thework presented here is drawn from [1].
7.2.1.1 Theory
This section of the chapter presents formulae for the probability density func-tion (PDF) of the energy of x(n) under normal operating conditions. Theseoperating conditions can be determined by the types of techniques described in[2]. The availability of the PDF enables reliable thresholds to be set so that onecan create alarms if the energy deviates too much from the normal operatingconditions. Typically, one will set the threshold for a false alarm rate of 10% orless. Because energy can be calculated fairly rapidly, the computation does notprovide a significant barrier to rapid detection of modal deterioration.
Once an alarm has been ‘‘raised’’ it is necessary to furthermonitor the energy.A series of sequential data windows are collected and statistical comparisonsare made with the energy PDF. Consistently high-energy readings will triggercorrective action.
7.2.1.2 PDF Derivation
Typically customer induced changes in the load are unpredictable in the shortterm. The power is modeled as the integral of white noise as shown in Fig. 7.1.The power PDF is first derived as shown below. The energy PDF is subse-quently determined from the power PDF via a simple scaling of the axes. Basedon the model in Fig. 7.2, the discrete spectrum of the output signal is
XðkÞ ¼ HðkÞWðkÞ (7:1)
7 Detection and Estimation of Nonstationary Power Transients 191
and the output power is
P ¼XN�1
k¼0XðkÞj j2 ¼
XN�1
k¼0HðkÞj j2 WðkÞj j2 (7:2)
whereH(k) is the discrete Fourier transform (DFT) of h(n) andW(k) is the DFT
of w(n). Note that W(k) is a complex random variable (RV) with real and
imaginary parts
W kð Þj j2¼ Real W kð Þf g2þImag W kð Þf g2 (7:3)
Now if the variance of w(n) is �2, then the left-hand side of (7.3) is a �2 RV
with two degrees of freedom and variance, �2/N, that is.
f W kð Þj j2n o
¼ N
2�2e�xN2�2 (7:4)
Fig. 7.1 Model for quasi-continuous modal disturbances in a power system
Fig. 7.2 Equivalent model for quasi-continuous modal oscillations in a power system
192 G. Ledwich et al.
Using (7.1) and (7.4) the PDF of X(k) can be deduced to be
fx Xkð Þ ¼ 1
HðkÞj j2
�����
�����fwx
HðkÞj j2
!¼ N
2 H kð Þj j2�2e�x
N
2 H kð Þj j2�2(7:5)
From (7.2) it is evident that the power is obtained by summingNRVs whereeach RV represents the ensemble power at discrete frequency k. Furthermore,these RVs have PDFs given by (7.5). The PDF of the sum (i.e., of the totalpower) is obtained by convolving the PDFs of all the RVs being summed. Thatis, the PDF of the total power is given by
fX xð Þ ¼ fXN�1 xN�1ð Þ � fXN�2 xN�2ð Þ � � � � � fX0x0ð Þ (7:6)
Expanding, we get
fX xð Þ ¼ N
2�2H 0ð Þj j�2e�xN2 H 0ð Þj j�2��2 � H 1ð Þj j�2e�xN2 H 1ð Þj j�2��2
� � � � � H N� 1ð Þj j�2e�xN2 H N�1ð Þj j�2��2
" #(7:7)
where * denotes convolution.The PDF of the energy that will correspond to a particular observation
interval is then obtained from the power PDF in (7.7) by simply rescaling theaxes by a factor equal to the observation length.
From the PDF, the threshold for detection of change can be formulated.Establishing say the 10% false alarm rate is via the cumulative summation ofthe PDF area until the 90% point is determined.
7.2.1.3 PDF Verification
To verify the theoretically determined system output PDF, simulations wereconducted with known modal parameters. These simulations created a data-base of outputs that were then formulated into a histogram. The simulatedhistogram and the theoretical PDF were then compared directly.
A two-mode system with modal parameters
Mode 1. !1 = 1.7 rad/s, �1 = –0.4 s–1
Mode 2. !1 = 2.7 rad/s, �1 = –0.52 s–1
generated the PDF shown in Fig. 7.3. It is based on a 20 s data acquisitionwindow. The variance of w(n), the disturbance excitation, was set to 1.0.The verification procedure involved creating a histogram from10,000 simulation runs of random noise feeding the known modal system,based on a 20 s data window with a sampling rate of 5 Hz. Statisticalcharacteristics of the simulated histogram and of the theoretical PDFwere also calculated and compared. The statistical characteristics
7 Detection and Estimation of Nonstationary Power Transients 193
examined were the first three central moments: mean, variance, and skew.
The percentage errors between the theoretical and simulated PDFs are
shown in Table 7.1. The errors are all comparatively low, inspiring
confidence in the fact that the derived PDF is correct.
7.2.1.4 Results
Simulations for detecting change have demonstrated promising results. The
following simulation is for a two-mode system with the following stationary
modal parameters:
Mode 1. !1 = 1.7 rad/s, �1 = –0.4 s–1
Mode 2. !1 = 2.7 rad/s, �1 = –0.52 s–1
Mode 1 (only) was changed in simulations in the following way:
� No change from normal operation (quiescent damping of �1 = –0.4 s–1)between 0 and 100 min.
� Deteriorating damping from –0.4 to –0.1 s–1 between 100 and 200 min.� Reset damping at –0.2 s–1 for 100 min.
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.040
10
20
30
40
50
60
70
80
90
Energy-Joules
Output half energy histogram vs PDF
Fig. 7.3 Histogram versus PDF
Table 7.1 Comparison of theoretical and simulated moments
Time window 20 s 40 s 60 s
Noise variance 1.0 1.0 1.0
% Error �1 0.94 0.00 0.00
�2 0.28 1.26 0.06
Skewness 3.97 6.38 6.01
194 G. Ledwich et al.
The energy levels for the simulated output are then compared to three setthresholds:
1. 10% false alarm rate.2. 5% false alarm rate.3. 1% false alarm rate.
Simulations were run for two data windows, 20 and 40 s with results shownrespectively in Figs. 7.4 and 7.5.
From these simulation results the following can be observed: when there isno deterioration in the damping then the rate of false alarms falls within anacceptable tolerance of false alarm. In both the 20 and 40 s simulations thealarm rate becomes significantly higher as the damping deteriorates. The alarmrates are shown in Tables 7.2 and 7.3 for different data windows.
It can be seen in the above results that during quiescent operation(0–100 min) the occurrences of false alarms happen at a generally expectedrate. However, as the damping linearly deteriorates (100–200 min), the numberof alarms rises dramatically. The two data window lengths (20 and 40 s) bothexhibit similar responses in alarming to the deteriorating damping, although the40 s data window does provide more tangible evidence of change (a higher rateof alarming). This is due to the increased statistical reliability which occurs withlonger data records. The longer the intended data analysis window, the moretruly Gaussian the formulated PDF tends to become. Hence, the skewed tail ofthe PDF in Fig. 7.3 (for a 20 s window) will be minimized for longer datawindows and alarm thresholds will be lower in respect to the PDF mean.However, as rapid detection is a major requirement for system control [3],
0 50 100 150 200 250 3000
0.005
0.01
0.015
0.02
0.025
0.03
0.035Half Output Energy vs Time
Time-Minutes
Ene
rgy-
Joul
es
Energy10% FAR5% FAR1% FAR
Fig. 7.4 The 20 s data window of energy measurements with false alarm rates
7 Detection and Estimation of Nonstationary Power Transients 195
shorter data analysis windows have been used in this section of the chapter. It
can also be observed in Figs. 7.4 and 7.5 that the alarm rate is very high between
180 and 200 min. The alarming occurs as the deteriorating damping approaches
�1¼ –0.1 s–1. This characteristic is highly desirable in rapid alarming situations
where a large detrimental change occurs within the system dynamics.
0 50 100 150 200 250 3002
4
6
8
10
12
14
16× 10
–3 Half Output Energy vs Time
Time-Minutes
Ene
rgy-
Joul
es
Energy10% FAR5% FAR1% FAR
Fig. 7.5 The 40 s data window of energy measurements with false alarm rates
Table 7.2 The 20 second data window alarm rates
Alarm rates
For threshold set togive false alarmrate¼ 10% at�1¼ –0.4 s–1
For threshold set togive false alarmrate¼ 5% at�1¼ –0.4 s–1
For threshold set togive false alarmrate¼ 1% at�1¼ –0.4 s–1
% Alarms for0–100 min nochange (�1 ¼–0.4 s–1)
6.33 3.33 1.67
% Alarms for100–200 minramped change(–0.4 to –0.1 s–1)
23.33 17.33 9.33
% Alarms for200–300 min setdamping (�1 ¼–0.2 s–1)
28 19.67 9.67
196 G. Ledwich et al.
Another point to note is in the 200–300 min data analysis window, where the50% reduction in Mode 1 damping is also ‘‘detected’’ by the rise in the rate ofalarming with respect to quiescent operating conditions.
7.2.2 Introduction to Kalman Approach
It is proposed that the change detection be performed using measurements frompower systems in normal operation [4]. This section of the chapter initially definesa stochastic model relating random ambient disturbance inputs in the powersystem, such as customer load changes, to the measured system output. Based onthis model a Kalman filter is then set up to estimate the output arising from thedisturbances. The innovation is then determined as the difference between themeasured output and the estimated output. It is well known that the ‘‘innovation’’from aKalman filter is spectrally white as long as the assumedmodel parametersare valid [5, 6]. By monitoring the whiteness of the innovation, therefore, one candetect if there are any changes in the model parameters [6].
7.2.2.1 Theory (See Development in [7])
This section uses the theory of optimal detection of random signals [8] fordetecting changes to individual modes. The impulse response, h(n), is consid-ered to be the sum of h1(n) and h2(n), with h1(n) corresponding to the currentmode of interest, and h2(n) corresponding to the other modal components ofh(n). The output, x(n) is considered to have two components, x1(n) and x2(n),with x1(n) being the output due to the mode of interest and x2(n) being theoutput due to the other modes. The detection algorithm involves passing the‘‘observed signal’’ power spectral density (PSD) and the ‘‘reference signal’’ PSDthrough a whitening filter and cross-correlating the outputs. The algorithm isoutlined in Fig. 7.6. The ‘‘observed signal’’ is considered to be x(n), while the
Table 7.3 The 40 s data window alarm rates
Alarm rates
For threshold set togive false alarmrate¼ 10% at�1¼ –0.4 s–1
For threshold set togive false alarmrate¼ 5% at�1¼ –0.4 s–1
For threshold set togive false alarmrate¼ 1% at�1¼ –0.4 s–1
% Alarms for0–100 min no change(�1 ¼ –0.4 s–1)
10.66 6.67 1.33
% Alarms for100–200 min rampedchange (–0.4 to–0.1 s–1)
34.67 22.67 17.33
% Alarms for200–300 min setdamping (�1 ¼–0.2 s–1)
36.67 27.33 16
7 Detection and Estimation of Nonstationary Power Transients 197
‘‘reference signal’’ is taken to be x1(n). It is assumed that there are Nsamples inthe observation.
The detection algorithm uses a threshold which is determined from a PDF ofthe cross-correlation test statistic. The availability of the PDF enables reliablethresholds to be set so that one can create alarms if the modal response deviatestoo much from the normal operating conditions. Typically, a threshold for afalse alarm rate of would be 10% or less. In practice, the monitoring processinvolves applying the detection algorithm to all the modes individually.
Because the detection algorithm can be calculated fairly rapidly, the compu-tational overhead does not provide a significant barrier to monitoring all theindividual modes. It should be noted though that an n-mode power systemwould require n parallel detectors to monitor each mode.
Once an alarm has been ‘‘raised’’ it is necessary to further monitor the modes.A series of sequential data windows are collected and statistical comparisonsare made with the stationary condition PDFs. Consistently high readings willtrigger corrective action on the deteriorating mode.
7.2.2.2 Individual Mode Test Statistic Details
The following quantities are first defined:
xðnÞ)= XðkÞ ¼W kð ÞH kð Þ (7:8)
h1ðnÞ)=H1ðkÞ (7:9)
h2ðnÞ)=H2ðkÞ (7:10)
x1ðnÞ)=X1 kð Þ ¼W kð ÞH1 kð Þ (7:11)
x2ðnÞ)=X2 kð Þ ¼W kð ÞH2 kð Þ (7:12)
where = indicates discrete Fourier transformation [9].
Observation PSD Whiten
Reference PSD Whiten
Cross-correlate
Threshold and Detect
Fig. 7.6 The detection algorithm
198 G. Ledwich et al.
Now to detect a change in Mode 1 (the mode of interest), we choose X1(k) asthe (frequency domain) reference signal. The remainder of the frequencydomain observation (X2(k)) becomes the interference signal. A whitening filteris created to whiten the interference according to
Hwh kð Þ ¼ H�12 ðkÞ (7:13)
This whitening filter is applied to both the reference and observation signals.The corresponding PSDs are then determined:
PSDref kð Þ ¼ H1 kð Þj j2 Hwh kð Þj j2E WðkÞj j2n o
(7:14)
PSDobsðkÞ ¼ jX kð Þj2jHwh kð Þj2
¼ jWðkÞj2jHobsðkÞj2jHwhðkÞj2(7:15)
where Efg denotes the expected value.Now cross-correlate (7.7) and (7.8) to obtain a test statistic �
� ¼XN�1
k¼0PSDrefðkÞPSDobsðkÞ (7:16)
To practically apply the test statistic (7.16) in a detection process, a thresholdlevel must be determined. To intelligently set the threshold, a PDF of theexpected test statistic is required. A threshold can then be set based on thePDF at a desired level of confidence.
7.2.2.3 PDF Derivation
The formulation of theMode 1 test statistic PDF is as follows. To derive the teststatistic PDF, (7.16) can be expanded using (7.14) and (7.15) to give
� ¼XN�1
k¼0E W kð Þj j2n o
H1 kð Þj j2 H kð Þj j2 HwhðkÞj j4 W kð Þj j2 (7:17)
The above equation can then be rewritten as
� ¼XN�1
k¼0jZðkÞj2jWðkÞj2 (7:18)
7 Detection and Estimation of Nonstationary Power Transients 199
where Z(k) is defined as
Z kð Þ ¼ jH1 kð ÞjjH kð ÞjjHwhðkÞj2ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiE W kð Þj j2n or
(7:19)
Now the expression in (7.18) containsW(k) which is a complex RV with realand imaginary parts. Furthermore, the square magnitude of |W(k)|2 is
W kð Þj j2¼ Real W kð Þf g2þImag W kð Þf g2 (7:20)
Where Real{} and Imag{} denote the real and imaginary parts, respectively.
If the variance of w(n) is �2, then the left-hand side of (7.20) is a �2 RV with twodegrees of freedom and variance, �2/N. That is,
f W kð Þj j2n o
¼ N
2�2e�xN2�2 (7:21)
Using (7.18) and (7.21), the PDF of |Z(k)|2|W(k)|2 at discrete frequency k can
be deduced to be
f ZWð Þ ¼ 1
ZðkÞj j2
�����
�����fwx
ZðkÞj j2
!¼ N
2 Z kð Þj j2�2e�x
N
2 Z kð Þj j2�2(7:22)
From (7.18) it is evident that the test statistic is obtained by summingNRVs.
Furthermore, these RVs have PDFs given by (7.22). The PDF of the sum(i.e., of the test statistic) is obtained by convolving the PDFs of all the RVsbeing summed. That is, the PDF of the test statistic for Mode 1 is given by
f� zwð Þ ¼ fZWN�1 zwN�1ð Þ � fZWN�2 zwN�2ð Þ
� � � � fZW0zw0ð Þ
(7:23)
f� zwð Þ ¼ N
2�2
Z 0ð Þj j�2e�xN2 Z 0ð Þj j�2��2
� Z 1ð Þj j�2e�xN2 Z 1ð Þj j�2��2
� � � � � Z N� 1ð Þj j�2e�xN2 Z N�1ð Þj j�2��2
2664
3775 (7:24)
where * denotes convolution.
200 G. Ledwich et al.
From the PDF in (7.24), the threshold for detection of change can be
formulated. To establish the 10% false alarm rate, the cumulative summation
of the PDF area is taken until the 90% point is determined.
7.2.2.4 Results
In practical systems; modal parameters are quasi-stationary and the values
determined via modal estimate algorithms [10, 11] applied to data measured
at optimal points within a power system as outlined in [12]. However current
algorithms for modal parameter estimation require significant data lengths for
accurate estimates. Therefore, to demonstrate the rapid detection ability of this
section’s algorithm, a 40 s data analysis window will be applied.The simulation in this section is for a two-mode system with the following
stationary modal parameters:
Mode 1. !1 = 1.7 rad/s, �1 = –0.4 s–1
Mode 2. !1 = 2.7 rad/s, �1 = –0.52 s–1
Mode 1 (only) was changed in the simulations in the following way:
� No change from normal operation (quiescent damping of �1 = –0.4 s–1)between 0 and 100 min.
� Deteriorating damping from –0.4 to –0.1 s–1 between 100 and 200 min.� Reset damping at –0.2 s–1 for 100 min.
The tests statistics for the simulated output are then compared to three set
thresholds
1. 10% false alarm rate.2. 5% false alarm rate.3. 1% false alarm rate.
Simulations were run for the 40 s data windows, involving change detection
of Mode 1 andMode 2. The results are shown respectively in Fig. 7.7–Table 7.4
and Fig. 7.8–Table 7.5.It can be seen in the above results that during quiescent operation
(0–100 min) the occurrences of false alarms are at a generally expected rate
for both Mode 1 (Fig. 7.6) and Mode 2 (Fig. 7.7), respectively. As the damping
of Mode 1 linearly deteriorates (100–200 min) the number of Mode 1 alarms
rises dramatically (Fig. 7.6). Importantly, it is shown in Fig. 7.4 and in Table 7.2
that the quiescent Mode 2 false alarms do not rise.It can also be observed in Fig. 7.7 that the alarm rate is quite high between
180 and 200 min. The alarming occurs as the deteriorating damping approaches
�1 = –0.1 s–1. It clearly indicates the Mode 1 change whilst Fig. 7.8 indicates
that there is no deterioration toMode 2. This characteristic is highly desirable in
rapid alarming situations where a large detrimental modal change occurs within
the system and the aberrant mode requires identification.
7 Detection and Estimation of Nonstationary Power Transients 201
Another area of interest is within the 200–300 min data analysis window
where there is a 50% reduction in Mode 1 damping. The analysis provides
significant alarming to this event (Fig. 7.7), whilst still maintaining no signifi-
cant deterioration has occurred in Mode 2 (Fig. 7.8).
0 50 100 150 200 250 3000
5
10
15
20
25
30Neta vs Time
Time-Minutes
Test
Sta
tistic
Net
a
Neta10% FAR5% FAR1% FAR
Fig. 7.7 The 40 s data window of Mode 1 test statistic with false alarm rates shown
Table 7.4 The 40 s data window Mode 1 alarm rates
Mode 1 alarm rate
For threshold set togive false alarmrate¼ 10% at�¼ –0.4 s–1
For threshold set togive false alarmrate¼ 5% at�¼ –0.4 s–1
For threshold set togive false alarmrate¼ 1% at�¼ –0.4 s–1
% Alarms for0–100 min (nochange)
11.33 4.67 1.33
% Alarms for100–200 min Mode1: ramped change(–0.4 to –0.1 s–1)
34.67 30.67 20.67
% Alarms for200–300 min Mode1: set damping(�1 ¼ –0.2 s–1)
44.00 36.00 22.00
202 G. Ledwich et al.
Other simulations (not shown here) involving three mode systems have
displayed similar successes in identifying individual modal change.
0 50 100 150 200 250 3001.5
2
2.5
3
3.5
4
4.5
5Neta vs Time
Time-Minutes
Tes
t Sta
tistic
Net
a
Neta10% FAR5% FAR1% FAR
Fig. 7.8 The 40 second data window of Mode 2 test statistic with false alarm rates shown
Table 7.5 The 40 s data window Mode 2 alarm rates
Mode 2 alarm rate
For threshold set togive false alarmrate¼ 10% at�2¼ –0.52 s–1
For threshold set togive false alarmrate¼ 5% at�2¼ –0.52 s–1
For threshold set togive false alarmrate¼ 1% at�2¼ –0.52 s–1
% Alarms for0–100 min (nochange)
7.33 2.00 1.33
% Alarms for100–200 min Mode1: ramped change(–0.4 to –0.1 s–1)
2.67 2.00 0.67
% Alarms for200–300 min Mode1: set damping(�1 ¼ –0.2 s–1)
4.67% 0.67% 0.67%
7 Detection and Estimation of Nonstationary Power Transients 203
7.2.3 Application to Real Data [7]
Data were obtained from the Australian power system, comprising voltageangle measurements at the Adelaide, Melbourne, and Sydney measurementsites from 22:00 on 09/04/2004 to 03:05 on 10/04/2004. As mentioned earlier,voltage angle measurements were used rather than power signals because thepotential for modal information extraction is greater than when using powersignals [13]. It is generally understood that inter-area modes have frequencies inthe range 0.1–0.8 Hz and so this will be the region of focus in the PSD [14].
To formulate the state–space model and consequently the system Kalmanestimator, a knowledge (or at least an accurate estimate) of the power systemtransfer function must be available. Accordingly, for the work reported in thissection, a long-term estimator (LTE) was applied to estimate this transferfunction [15]. This LTE was determined using a 45 min window. Short-term(change) detection was then continuously applied to the PSD of the Kalmaninnovation.
The real data analysis was conducted in two parts. Part I closely examined305 min of the Melbourne measurement data difference from the center of areaof the connected system (called Melbourne-COA) and Part II briefly analyzedthe data collected at the Sydney and Adelaide sites. Part II also examined theopportunity for combining the multisite innovation power spectrum data toenhance the detector performance.
7.2.3.1 Part I: Analysis of the Melbourne Data
The LTE was determined from the data between 120 and 165 min after the startof the measurement record. The LTE quasi-stationary modal estimates wereobtained using the technique in [15] and are listed in Table 7.6. The LTE alsoprovides an estimate of the measurement site transfer function in Laplace form.As a result, an estimate of the system quasi-stationary frequency response forMelbourne-COA can be observed in Fig. 7.9. In Fig. 7.9, Mode 1 peak is quiteapparent at 0.33 Hz, whileMode 2 peak (estimated to be at 0.59 Hz) is harder todistinguish due to the relatively heavy modal damping.
Once the long-term estimate of system characteristics was established, theremaining 140 min were examined in 1 min intervals for any sudden detrimentalchanges to system modes. To demonstrate the significant information theinnovation sequence contains, one only has to compare the differentiatedangle measurements with the normalized innovation obtained after application
Table 7.6 Damping and frequency for longtime estimates over 120–135 min
Mode Damping Frequency (Hz)
1 0.2913 0.33
2 1.0083 0.59
204 G. Ledwich et al.
of the Kalman estimator, shown in Fig. 7.10. The comparison appears to
support the result noted by Kailath in [5] – namely, that the Kalman filter
innovation sequence contains the same information as the system output
sequence, but in a less correlated form.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.90
0.5
1
1.5
2
2.5
1
3
× 10–3
Hz
Melbourne Frequency Response
Mag
nitu
de (
units
)
Fig. 7.9 Melbourne frequency response estimate from LTE at 165 min (portions reprinted,with permission, from [16]. # 2007 IEEE)
Fig. 7.10 Comparison of (a) system output and (b) normalized innovation (portions rep-rinted, with permission, from [16]. # 2007 IEEE)
7 Detection and Estimation of Nonstationary Power Transients 205
It is apparent from inspection of Fig. 7.10 that the system output and the
innovation both demonstrate a deviation from a quasi-stationary operation
within the 197–201 min interval. To lend some numerical support to this visual
inspection a 60 s analysis windowwas applied to the innovation PSD before and
during the 197–201 min interval. The results of the analysis are provided in
Figs. 7.11, 7.12, and 7.13. The innovation is shown in Fig. 7.11(a) with the
196–197 min interval marked off by two vertical lines. The spectrum of the
196–197 min segment of innovation is shown in Fig. 7.11(b). To account for
the fact that with real data, there is less certainty than there was with simulated
data, a different threshold level will be used for the detection. Accordingly, the
99.999%CI threshold is shown as a dashed horizontal line. Even though such a
threshold could be regarded as high, the main focus is to only detect large
detrimental changes and minimize false alarms. In the single site analysis, the
data does exhibit a wide variance and hence a high threshold is required to
minimize the false alarms. However in the following section, this issue is
addressed and the ability to have more acceptable threshold CIs is presented
along with a measured number of false alarms.
Fig. 7.11 (a) Innovation sequence �(n) and (b) innovation PSD at 196–197 min (portionsreprinted, with permission, from [16]. # 2007 IEEE)
206 G. Ledwich et al.
In examining Fig. 7.11, no part of the spectrum crosses the threshold.Figure 7.12 depicts the innovation segment and associated spectrum correspond-ing to the 197–198 min segment. For this segment the innovation spectrum crossesthe threshold. Moreover, the threshold is crossed at the 0.59 Hz frequency posi-tion, indicating a loss of damping for Mode 2. Figure 7.12 shows the innovationsequence and associated spectrum corresponding to the 198–199min segment. Forthis segment the innovation spectrum crosses the threshold in an even morepronounced way than it did in Fig. 7.13(b). Again, the threshold is crossed atthe 0.59 Hz frequency position, indicating a damping change for Mode 2.
Further analysis was performed which showed that the loss of damping wastemporary. By the 201st minute of the data record, the modes reset to theiroriginal characteristics.
7.2.3.2 Part II: Combining Multisite Data for Enhanced SNR and Detection
Similar data analysis was conducted for the other two measurement sites inSydney and Adelaide (with COA correction). A comparison of the results forthe 196–197 and the 198–199 min time frame can be seen in Figs. 7.14(a) and
Fig. 7.12 (a) Innovation sequence �(n) and (b) innovation PSD at 197–198 min (portionsreprinted, with permission, from [16]. # 2007 IEEE)
7 Detection and Estimation of Nonstationary Power Transients 207
7.15(a), respectively. Figure 7.14(a) demonstrates a spectrum for Sydney and
Adelaide, which does not cross the threshold prior to the event. Figure 7.15(a)
shows a detection of damping deterioration centered at 0.59 Hz at both the
Sydney and Adelaide sites. These results confirm the detection registered at the
Melbourne site for the same analysis window.Within these plots, all sites exhibit similar responses to detection of damping
deterioration. Therefore, a combinationof the innovation spectrum was exam-
ined to assess opportunities for an enhanced detector. The combined spectrum
was obtained by adding the complex innovation spectrums from all three sites.The threshold for the combined innovation spectrum will need to be set
differently to that for the individual innovation spectra. Ideally, if the individual
normalized innovation spectra are all uncorrelated with one another then the
samples of the combined innovation spectrum will again have a �2 PDF with 28of freedom, �2, but with three times the variance (assuming individual normal-
ized innovations have unity variance) [15]. In the case of the innovation spec-
trums from the multisite data as examined in this paper, the innovation spectra
will not be strictly independent, as the power system is interconnected.
Fig. 7.13 (a) Innovation sequence �(n) and (b) innovation PSD at 198–199 min (portionsreprinted, with permission, from [16]. # 2007 IEEE)
208 G. Ledwich et al.
Nonetheless, empirical experiments have indicated that a �2 distribution, withvariance 3�2 is a suitable way to model the PDF.
In the general case whereXmeasurement sitesmust be combined, the samplesof the combined normalized innovation spectrum would have a �2 PDF with 2degrees of freedom [17, 18], and variance X�2. Thus using the normalizedinnovation the generalized ensemble frequency PDF for a combination of X,N-point innovation spectra would be
f � kð Þf g ¼ N
Xe�� kð ÞNX (7:25)
Hence for the three measurement sites, X =3, (7.25) simplifies to
f � kð Þf g ¼ N
3e�� kð ÞN3 (7:26)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1–50
–40
–30
–20
–10
0(b)Combination Normalised Innovation Spectrum
Frequency Hz
dB/H
z
PSD @ 196 mins
Innovation CombinationThresholdest Mode 1 frequencyest Mode 2 frequency
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1–50
–40
–30
–20
–10
0(a)Individual Normalised Innovation Spectrum
Frequency Hz
dB/H
z
SydneyMelbourneAdelaideThreshold
Fig. 7.14 Normalized (a) Individual innovation PSDs for Sydney, Melbourne, and Adelaide(detection threshold and modal frequency estimates also shown) (b) Combination PSD at196–197 min showing new threshold with false alarm rate (FAR) of 99.9% (portions rep-rinted, with permission, from [16]. # 2007 IEEE)
7 Detection and Estimation of Nonstationary Power Transients 209
where (7.26) is the expected ensemble PDF of the three combined spectra
assuming the power system is quasi-stationary and the resulting innovation, white.The resulting combined innovation spectra prior and during the disturbance
are shown in Figs. 7.13(b) and 7.14(b), respectively. It is important to note two
significant outcomes attributed to the combining of the normalized innovation
spectra. First, the combination of the three sites has lead to an improvement in
the signal-to-noise ratio (SNR) for detectable signals. The SNR analysis results
are shown in Table 7.7, in which the bold letters emphasize the key results. The
improved SNRs exhibit comparable values to an ideal theoretical improvement
of 4.72 dB, whereby an ideal theoretical improvement is one that would be
expected if three identical deterministic signals, with independent, equal var-
iance Gaussian white noise were spectrally combined. Hence, the resulting
theoretical expected improvement under these criteria for three sites is
SNRimprovement ¼ 10 log10 3sitesð Þ ¼ 4:77 dB (7:27)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1–50
–40
–30
–20
–10
0(a)Individual Normalised Innovation Spectrum
Frequency Hz
dB/H
z
SydneyMelbourneAdelaideThreshold
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1–50
–40
–30
–20
–10
0(b)Combination Normalised Innovation Spectrum
Frequency Hz
dB/H
z
PSD @ 198 minsAlarm @ 0.58333Hz
est Mode 1 frequencyThresholdInnovation Combination
est Mode 2 frequency
Fig. 7.15 Normalized (a) Individual PSD and (b) combination PSD at 198–199 min showingnew threshold with FAR of 99.9% (portions reprinted, with permission, from [16]. # 2007IEEE)
210 G. Ledwich et al.
Although the result in (7.27) was verified both analytically and empirically,in practice however, the responses from the three sites are not perfectly inde-pendent and the detection signals exhibit similar statistical properties. Hence,the combining of the innovation spectra will exhibit slightly less than thetheoretically maximum improvement in SNR.
7.2.3.3 Summary of Kalman Approach
The simulations indicate that the method proposed in this section can effec-tively monitor individual modal changes in power systems. It does so with apredetermined level of confidence that sets the alarm at a desired false alarmrate. Although not shown here, longer time widows exhibit a greater degree ofconfidence in a detected change, and can be used in conjunction with the shortertime windows to confirm or disaffirm a need for remedial action.
In practice, the monitoring is done for all modes concurrently. It should benoted that the methods outlined in this section only apply to multimodesystems. Rapid detection of detrimental mode change within single-mode sys-tems can use the method discussed in [7].
7.3 Estimation ofModal Parameters fromNonstationary Response
7.3.1 Introduction
The estimation of power system modes, in particular mode shapes andeigenvalues, is of importance in the design of stabilization controls such aspower system stabilizers. The use of eigenvalues to describe modes is derivedfrom the concept of a linear model for a power system. A power systemhowever, like many systems is actually nonlinear and so the linear model isat best only a good approximation when the system is not subjected to largedisturbances.
In recent years, much research has focused on both active [19, 20], andpassive [21], estimation of modes using parametric, and in some cases, non-parametric signal processing techniques, as detailed in [22]. Active estimationhas involved subjecting the system to a disturbance and analyzing the powerflows that result using a number of linear estimation techniques. It is the active
Table 7.7 SNR improvement through combination of site analysis
Analysis window during alarm signal (min)
Parameter 197–198 198–199 199–200 200–201
Average single-site SNR (dB) 9.60 7.60 10.82 11.64
Combined SNR (dB) 13.98 12.00 15.27 15.80
SNR improvement (dB) 4.38 4.40 4.45 4.16
7 Detection and Estimation of Nonstationary Power Transients 211
estimation of power system modes using large disturbances that this paper willfocus on. The results are also applicable to recorded large disturbance eventsthat may naturally occur within a power system from time to time.
The estimation of modes presupposes a linear model for a power system.Any disturbance applied to a power system as part of an active modalestimation study runs the risk of exciting nonlinear system behavior. Becausea linear model for a power system is basically a small-signal model theprobability of nonlinear behavior increases with the size of the applieddisturbance. The purpose of this section is to present a range of techniqueswhich can
1. readily detect if any nonlinear behavior is present in a disturbance record,and
2. if any nonlinear behavior is present which decays with time, obtain anaccurate modal estimate.
7.3.2 Estimation of Linear Power System Models
In state–space form the standard homogeneous linearized model of a powersystem is of the form
_x ¼ Ax (7:28)
where x is a state vector consisting of machine angles, velocities, andother machine-related variables such as fluxes and A is the state matrix.The eigenvalues of A give information on the damping and frequency ofeach mode, while the eigenvectors give information related to the parti-cipation of each state in the various modes. The latter is usuallyexpressed in terms of participation factors pki which represents the parti-cipation of the kth state in the ith mode. These are computed from theleft and right eigenvectors of A according to [23, p. 715].
From the model in (7.28), the oscillatory time-domain response of thelinearized power system with M complex eigenvalues of the form lI = �I �j2pfi is
y tð Þ ¼XM M2= 2
i¼1Ai exp �itð Þ cos 2pfitþ �ið Þ (7:29)
The estimation of the eigenvalues involves the estimation of the dampingfactors �i, modal frequencies fi, amplitudes Ai, and phase angles �i.
The estimation of the eigenvalues of A may be performed either parametri-cally or nonparametrically. For details of some of the nonparametric techni-ques that have been proposed, see [9, 24,25].
212 G. Ledwich et al.
A parametric technique used by many workers in the analysis of powerswings following large disturbances is Prony’s method, [11, 19, 26–28]. Theuse of Prony’s method assumes a linear time-invariant system and using a linearprediction models fits a set of P complex exponential functions to a given datarecord. The linear predictor model of a set of N data samples of a discretizedsignal y(n) is
y nð Þ ¼XP
i¼1Ai exp �i þ j2pfið Þ n� 1ð ÞTþ j�ið Þ (7:30)
which maybe written as
y nð Þ ¼ �XP
i¼1a ið Þy P� ið Þ (7:31)
The discrete time eigenvalues zk are then found as the roots of the character-istic equation
� zð Þ ¼XP
m¼1a mð Þzp�mk (7:32)
The estimation procedure which is detailed in [11] is as follows:
1. Find the coefficients a(k) that best fit the data record.2. Form and determine the roots of (7.32).3. Determine the damping factors, modal frequencies, modal amplitudes and
modal phases using (7.33) below
�k ¼ln zkð ÞT
fk ¼tan�1 Im zkð Þ
Re zkð Þ
h i
T(7:33)
where T is the sampling period.
7.3.3 Time–Frequency Representations
The use of Prony’s method and other linear parametric/nonparametric methodsassumes the absence of nonlinear and time-varying dynamics. Time–frequencyrepresentations (TFRs) originally designed to analyze nonstationary signals area logical choice as a method of testing this assumption. Examples of nonsta-tionary signals occur in radar, swept sine wave testing, communications, and
7 Detection and Estimation of Nonstationary Power Transients 213
biomedical phenomena. A simple piece of music is another example of a non-stationary signal.
The response of a linear system to a disturbance will consist of a linearcombination of modal frequencies which do not change with time. This is nottrue of a nonlinear system, and in many cases the response will be a nonsta-tionary signal.
The spectral variation of a nonstationary signal is measured by its instanta-neous frequency (IF). The IF is a measure of how the frequency of a nonsta-tionary signal changes with time. The IF fi is defined by Eq. (7.34),
fi ¼1
2pd�ðtÞdt
(7:34)
where z(t) = z(t)ej�(t) is the analytic version of the measured signal y(t)computed by taking its Hilbert transform, see [29, Chapter 1] and �(t) isknown as the instantaneous phase. If the signal is a constant frequencytone then �(t) = 2pf. A simple example of a nonstationary signal is a chirp,whose IF is a linear function of time. The analytic version of a chirp signalwould be given as
zðtÞ ¼ ej2p�t2
2 þf0t� �
and would have an IF given by fi(t) = � t + f0. An example of a linear chirpsignal where �= 50 and f0 = 0 Hz is illustrated in Fig. 7.16.
A time–frequency distribution (TFD) seeks to display both temporal andspectral data on a three-dimensional plot. The short-time Fourier transform(STFT) is perhaps themost intuitive of TFDs. The STFT computes a successionof spectra localized about a time t by first multiplying the signal by a windoww() centered about t = . Thus the STFT S(t,f) is computed as (7.35)
S t; fð Þ ¼Z 1
�1s ð Þw t� ð Þe�j2pf d (7:35)
The STFT, like all TFRs, represents a compromise between time-domainresolution and frequency-domain resolution. The longer the window w(), thebetter the frequency resolution but at the expense of reduced time-domainresolution and vice versa. It has been shown in [29, Chapter 2] that an optimalrectangular window length is
� ¼ffiffiffi2p dfi
dt
����
�����1
2
(7:36)
where fi is the IF of the signal. Because this is unknown a priori, the STFT islimited in its usefulness.
214 G. Ledwich et al.
The Wigner–Ville distribution (WVD) W(t, f), of a signal S(t) is defined by(7.37)
Wðt; fÞ ¼Z1
�1
w
2
� �w��2
� �s tþ
2
� �s� t�
2
� �exp �j2pfð Þ d (7:37)
where w(t) is the time-domain window applied to the signal. W(t, f) is basicallythe Fourier transform of the autocorrelation function of a signal in the lagvariable.
A discrete version of theWVD,W(n, k) where n is the time index and k is thefrequency index may be computed as
Wðn; kÞ ¼ 2XM�1ð Þ=2
m¼�ðM�1Þ=2w mð Þw� mð Þs nþmð Þs� n�mð Þe�j4pmk=M (7:38)
where w is an M point window centered at m ¼ 0. Algorithms for determiningthe WVD can be found in [29, Chapter 6].
TheWVD has the property of giving the optimal trade-off between temporaland frequency resolution for a linear FM signal, that is, a signal whose IF is alinear function of time. In the case of a chirp signal or a sinusoid that hasconstant IF the continuous time WVD consists of a series of delta functionslying along the IF law.
An example of the WVD plot for a chirp signal is shown in Fig. 7.17. The IFis described by a ridge in the time–frequency plane.
Fig. 7.16 Linear chirp signal with � ¼ 50 and f0 ¼ 0 Hz
7 Detection and Estimation of Nonstationary Power Transients 215
Where there is only one component it is straightforward to compute theinstantaneous frequency as
fiðtÞ ¼ f :Wðt; fiðtÞÞ ¼ maxf
Wðt; fÞj j (7:39)
The WVD being a quadratic TFD suffers, however, from the presence ofcross-terms when more than one component is present in the signal. In otherwords, if s(t) ¼ s1(t) + s2(t), then the WVD of s(t) is
Wss t; fð Þ ¼Ws1s1ðt; fÞ �Ws2s2ðt; fÞ �Ws2s2ðt; fÞ �Ws2s1ðt; fÞ
These cross-terms appear at intermediate frequencies and may obscureaspects of the time–frequency behavior of a system. This effect can in part bemitigated by replacing the signal in (36) by its Hilbert transform [30].
To reduce these effects further, a number of TFRs have been proposed basedon filtered versions of the WVD. The general form of these other TFRs is givenin (7.39).
W t; fð Þ ¼Z1
�1
G t; ð Þ �ts tþ
2
� �s� t�
2
� �exp �j2pfð Þ d (7:40)
where �t denotes convolution in the time variable such that
Fig. 7.17 WVD of linear chirp signal shown in Fig. 7.16
216 G. Ledwich et al.
G t; ð Þ �ts tþ
2
� �¼Z1
�1
G l; ð Þs t� lþ 2
� �dl
and G(t,) is a function of time and lag called the time–lag kernel. The time–lagkernel can be engineered to filter out the effect of the highly oscillatory cross-terms. For further details of this, the reader is referred to [29, Chapter 2].Various toolboxes exist for computing the WVD and other TFRs. The onesused in this paper may be downloaded from http://tftb.nongnu.org/ asMATLAB ‘.m’ files and form a package of time–frequency routines developedby Francois Auger, Oliver Lemoine, Paulo Gonclaves, and Patrick Flandrinunder the auspices of the Centre National de la Recherche Scientifique. Wheremore than one component is present and there is more than one phase functionthe meaning of IF needs to be refined. Some workers in the field use thefollowing definition, see [31]:
fi ¼
R1
�1fW t; fð Þ df
R1
�1W t; fð Þ df
(7:41)
which for monocomponent signals is the same as (7.34).
7.3.4 Application to Transient Stability Swings
In the presence of a small disturbance a power system should behave in a linearfashion characterized by a number of modes with time-invariant parameters. Inthe case of a large disturbance, however, the linear assumption is not valid andnot only do the modes lose their decoupling but variation in frequency of eachmode may occur. The coupling phenomenon that occurs was investigated in[32], while the effect onmodal estimates was been investigated in [27]. In [27] thenonlinearities of the transmission system, excitation system current limiting,load models and field saturation were investigated and shown to affect eigen-value estimates.
In this chapter, the transmission system nonlinearity is the only nonlinearitythat is focused on. While this was found to be not most significant nonlinearityin the cases considered in [27] it can be the most significant nonlinearity whenthe system is subjected to a severe disturbance. This is indeed the case with thesystem considered in this chapter.
This is illustrated for the 21 bus, six-machine system shown in Fig. 7.18 anddescribed in [33, 2]. The model was originally designed to model a typicalAustralian network which in general are not meshed networks and consequen-tial have very well-defined modes. The machine models used were fifth order
7 Detection and Estimation of Nonstationary Power Transients 217
and the loads are modeled as constant impedances. Further details of the
parameters of this system may be obtained from either [33] or [34].A modal analysis of this system reveals the mode shapes detailed in Table 7.8.
The mode shapes were determined by examining the phase angles of the angular
velocity components of the eigenvectors corresponding to electromechanicalmodes.There are five electromechanical oscillation modes associated with this
system. Three of these modes are local modes only involving two machines.
The other two are inter-area involving coherent groups of generators. It is the
inter-area modes that are of principal interest as they typically have lower
damping than the faster local modes.A disturbance in the form of a three-phase fault to ground is applied at bus
14 which excites inter-area modes 4 and 5. This means that components of both
modesmay be present inmeasured data depending onwhere the data is taken. If
the data are taken from machine 5 terminals only one mode, mode 5, will be
present, while at the terminals of machine 3 both modes 4 and 5 will be present.
The local modes 1–3 will not be excited by this fault.If the fault is short this corresponds to a small disturbance whilst a prolonged
fault constitutes a large disturbance likely to excite transmission system non-
linearities. Ringdown plots of machine 5 angular velocities are shown in
68
75
192017
16
92
121
1514
1312
103
11
4
18
Fig. 7.18 Twenty-one bus test system
Table 7.8 Eigenanalysis of six-machine system
Mode no. Description Eigenvalue
1 Machine 3 vs. 4 �1:25� j16:1812 Machine 1 vs. 2 �0:86� j12:82
3 Machine 5 vs. 6 �1:33� j12:01
4 Machines 1 and 2 vs. 3 & 4 �0:37� j10:89
5 Machines 5 and 6 vs. rest �0:13� 6:41
218 G. Ledwich et al.
Figs. 7.19 and 7.21. The variation in frequency with time for a large disturbance
is apparent from Fig. 7.21 but to properly quantify this, a TFD is necessary.
WVDs were computed for each case using the entire downsampled analytic
version of the original signal. The preprocessing of the data involved:
1. Downsampling the signal from a 50Hz sampling rate to a 5Hz sampling rate2. Computing the Hilbert transform of the downsampled signal to obtain the
analytic version of the downsampled signal.
The effect of the size of the disturbance on the time–frequency spectra of the
angular velocity of machine 5 is illustrated by Figs. 7.20 and 7.22. Figure 7.20
shows a constant modal frequency following a short disturbance indicating that
a linear model is an accurate representation of the system in this case over the
entire transient period.Figure 7.22, on the other hand, shows that a time-varying frequency appears
following a large disturbance eventually settling down to a constant frequency.
This phenomenon cannot be explained by a simple linear model and so it is clear
that the transmission system nonlinearities are having a significant effect. These
effects, however, become less significant as the angular velocity swings decay.
Hence, the IF in the large disturbance case eventually reaches a constant
frequency which is consistent with the predictions of the linear model.The advantage of the WVD over the STFT in representing the time-varying
frequency is illustrated in Fig. 7.23 that shows the STFT of the signal whose
WVD is shown in Fig. 7.22. It is clear from comparing Figs. 7.22 and 7.23 that
the WVD plot gives a sharper image of the variation of modal frequency with
time In other words, the WVD has better time and frequency resolution than
Fig. 7.19 Machine 5 ringdown following small disturbance (clearing time ¼ 0.20 s)
7 Detection and Estimation of Nonstationary Power Transients 219
the STFT. In doing so, this enables better the user to estimate the time variation
of the modal frequency.In Table 7.8, it is shown that of the two inter-area modes, machine 5 only
participates in mode 5. Hence, theWVD computed from the angular velocity of
machine 5 consists of only one component and so cross-terms are not of
concern. The local mode in which machine 5 participates is not excited by the
applied disturbance. For initially illustrating the technique presented in this
section, the discussion will concentrate on the case where a single mode is
Fig. 7.20 WVD of machine 5 angular velocity – small disturbance
Fig. 7.21 Machine 5 ringdown after a large disturbance (clearing time ¼ 0.85 s)
220 G. Ledwich et al.
dominant in the response. For this reason, the discussion will initially focus on
the response of machine 5.Application of Prony’s method to the entire data set obtained from the large
disturbance case will clearly produce erroneous answers, as is shown in Table 7.9.
As expected, the error incurred in the case of a large disturbance is significant.
Fig. 7.22 WVD of machine 5 angular velocity – large disturbance (clearing time ¼ 0.85 s)
Fig. 7.23 STFT of machine 5 angular velocity – large disturbance (clearing time ¼ 0.85 s)
7 Detection and Estimation of Nonstationary Power Transients 221
7.3.5 Error Reduction by Time-Domain Windowing
As expected, the modal estimates obtained in the presence of the nonlinear
effect of the transmission system are erroneous, and this nonlinear effect is
clearly illustrated by theWVD. It is seen from theWVD that system does attain
a linear state with constant frequency once the amplitude of the oscillations
have died down. An obvious way of removing the nonlinear effects is to
determine a region in time where the nonlinear effects as described by the
WVD are negligible and only perform Prony analysis (or any other linear
modeling technique) over that region.The instantaneous frequency may be estimated by peak detection in the
time–frequency plane, in accordance with Eq. (7.38). Using this technique, it
is possible to generate a plot of IF over the length of the transient (Fig. 7.24).To determine the boundary of this region we note from Fig. 7.24 that the
time-varying frequency reaches 95% of the steady-state value after about 10 s.
Clearly, it is easier to determine the point where the modal frequency has
reached steady state using the WVD than the more intuitive STFT on account
of its superior time and frequency resolution.It should be noted that after about 25 s the transient has decayed below the
noise floor and the IF estimate from that point onward is meaningless.
Table 7.9 Prony analysis
True eigenvalue Small disturbance estimate Large disturbance estimate
–0.13 � j6.41 –0.1 � j6.28 –0.19 � j5.35
Fig. 7.24 Instantaneous modal frequency, fi
222 G. Ledwich et al.
If the signal is truncated so that only the section between 10 and 25 s after the
onset of the disturbance is retained, then the Prony modal estimate is as shown
in Table 7.10.This process is further validated by projecting backward the predictions of a
linear model fitted to the portion of the transient following the point where the
IF has reached a constant value. It is seen from Fig. 7.25 that the linear model
fails to accurately predict the transient response in the period before the IF
becomes constant but is a good predictor over the latter part of the transient
when the nonlinear effects have become negligible.As can be seen there is a dramatic improvement in the Prony estimate by
simply windowing the signal between the time when the nonlinear aspect has
sufficiently decayed and the time when the signal has decayed to the noise floor.
Clearly, the discarding of data must be done with some degree of caution, and
so, with this in mind varying window sizes were investigated.Defining fpu(t) = fi(t)/fss to be the per unit instantaneous frequency relative
to the steady state frequency fss, the percentage error in eigenvalue estimates for
windows commencing at different values of fpu is depicted in Fig. 7.26.The plot in Fig. 7.26 clearly illustrates the improvement in the eigenvalue
estimate error from nearly 16% for an estimate based on the entire data set to
less that 2 % when the window commences at a time when the per unit
instantaneous frequency has reached 0.95.
Table 7.10 Windowed eigenvalue estimates
True eigenvalue Unwindowed eigenvalue estimate Windowed eigenvalue estimate
–0.13 � j6.41 –0.19 � j5.35 –0.13 � j6.33
Fig. 7.25 Back prediction of linear model. ***, prediction of linear model; —, machine 5angular velocity)
7 Detection and Estimation of Nonstationary Power Transients 223
7.3.6 Discussion
A power system is by nature multimodal. Discussion so far has only concen-trated on a signal containing a single mode. A disturbance will in general exciteseveral modes and these will be apparent in the signal record depending uponthe location of the measurement site. In the 21-bus system considered here,there are two inter-area modes excited by the applied fault. Machine 5 onlyparticipates in one of these modes and hence there is only onemodal componentpresent in its angular velocity signal. Machine 3, on the other hand, participatesin both inter-area modes as is evident from the WVD plot of the angularvelocity for machine 3 shown in Fig. 7.27. Here the cross-term effect discussedin Section 7.2.3 is apparent. This may be removed at the expense of some loss ofresolution using a reduced interference TFR such as the Choi–Williams Dis-tribution (see [29, Chapter 2]). However, in this case it is obvious that compo-nent lying in frequency midway between modes 4 and 5 is the cross-term effect.
Now it should be noted that the damping factor of the two modes aresignificantly different, the damping of the higher frequency mode being larger.The windowed Prony technique can be used provided both modes have left thenonlinear region. If however, one mode has decayed significantly by the timethe other component has left the nonlinear region, then it would be preferable toapply different windows to the two components. This would avoid the loss ofinformation about the more heavily damped mode, which may occur by trun-cating the signal at the point in time where the lightly dampedmode’s IF law hasreached a constant value. This would necessitate the filtering off of the longerlasting component in order to identify the more heavily damped one.
Fig. 7.26 Error in eigenvalue estimate versus per unit instantaneous frequency at the start ofthe window
224 G. Ledwich et al.
Also as the higher frequency mode is more heavily damped the effectiveduration of that component of signal is less, which in turn creates a greaterspread in the time–frequency plane of that component. This is evident inFig. 7.26 and would make the determination of an IF law for the more heavilydamped component more difficult.
If an analysis is made of the data from machine 3 it is found that the error inboth inter-area eigenvalue estimates is much greater when the entire data set isused. If a window is used which is truncated to begin at the point where aconstant IF is attained, errors in the identification of both modes of less than3% are obtained. This is quite a good performance and tends to validate theymethod for well-separated modes. The effect of varying window start points isshown in Figs. 7.28 and 7.29.
Of course, not all modes are within the resolution limit of the DFT. In suchcases if a time-varying IF law is not obvious from the standardWVD one couldattempt to use a high-resolution WVD as described in [35]. The high-resolutionWVD relies upon the use of high-resolution spectral estimation techniques ofthe sort described in [22] in place of the FFT in computing Eq. (7.37).
The computational load of such high-resolution processing is quite intensive;however, with modern computing power it may well be possible to pursue thisapproach further. Work is continuing in addressing the issue of multiple modes.
Another factor that could be of assistance in identifying the modal para-meters where multiple modes are excited is the use of multiple measurementsites. Some sites will bemore sensitive to certain modes than others, and it mightbe possible to use different sites for measurement of modes of interest. Methodsfor choosing such sites are described in [2].
Cross -Term
Mode 4
Mode 5
Fig. 7.27 WVD – machine 3 angular velocity – multimodal example
7 Detection and Estimation of Nonstationary Power Transients 225
Finally, it may be possible to extend this work by developing a means of
modal identification directly from the WVD. While estimation of modal fre-
quency is reasonably straightforward provided the signal is not too highly
damped, the estimation of damping presents challenges.
Fig. 7.28 Error in eigenvalue estimate versus per unit frequency at the start of the window –lightly damped mode
Fig. 7.29 Error in eigenvalue estimate versus per unit frequency at the start of the window –heavily damped mode
226 G. Ledwich et al.
In conclusion, the method should prove useful in a real power systemwhere asingle dominant mode is excited by a disturbance in
1. Check to see if a nonlinear effect is present2. If one is present, then place a limit on the range of the record that can be used
with any linear estimation technique.
Finally it should be noted that TFDs may be used in any application where atime-varying frequency or transient event may be present. For this reason,discrete control events such as line switchings or load shedding may be detect-able used time–frequency analysis of machine or bus angle variations. Similarly,inter-modal coupling as described in [32] should also be observable and quanti-fiable using a TFD. In this case, a reduced interference TFD would be recom-mended in order to remove cross-term effects.
In terms of computational burden, an N length data sequence computing aWVD requires o(8N2 log2 2N) real multiplications and o(10N2 log2 2N) real addi-tions, [36], thus giving a total of o(18N2 log2 2N) floating point operations.Depending on the number of windows used in computing the STFT the numberof floating point operations involved is o(MN log2N) whereM is the number ofwindows used. Hence, it can be seen that the computation of a WVD is morecomputationally intensive. However, this must be weighed against the improve-ment in time–frequency resolution that results. It is this improvement that makestheWVD a better tool for estimating the point in a transient where the nonlineareffects may be discounted and a linear model may be fitted with confidence.
7.3.7 Recommendations for Time–Frequency
TheWVD has been shown to be of benefit in estimating power systemmodes incases where large disturbances have been applied to the power system. It hasbeen shown that the application of a linear estimation technique such as Prony’smethod over the entire disturbance record can yield very inaccurate results onaccount of system nonlinearities being excited.
Provided modes are separable by the DFT, it should be possible to identifymodal parameters more accurately using the WVD to design a time domainwindow over which to apply a linear technique such as Prony (windowed Prony).
7.4 Conclusions
Time-varying effects are normal for power systems but make the application ofstandard analysis tools more difficult. The first contribution for this chapterwas to define when a significant change had occurred in the dynamics of thesystem by analyzing the dynamic responses to the continuous load changes.Using detection theory rather than identification, the probability of false alarms
7 Detection and Estimation of Nonstationary Power Transients 227
was directly set. The second issue was the identification of system modes
following a major system change. Initially the disturbance is large and linear
identification would be flawed. When the disturbance becomes smaller linear
theory applies but the signal is now more difficult to distinguish from the
background noise. The contribution is to show how time–frequency analysis
can yield improved modal estimates in this time-varying context.
References
1. R. Wiltshire, P. O’Shea, and G. Ledwich ‘‘Rapid Detection of Deteriorating ModalDamping in Power Systems’’ AUPEC, 2004.
2. E. W. Palmer and G. Ledwich, ‘‘Optimal placement of angle transducers in powersystems,’’ IEEE Transactions on Power Systems, Vol. 11, pp. 788–793, 1996.
3. T. George, J. Crisp, and G. Ledwich, ‘‘Advanced tools to manage power system stabilityin the National Electricity Market,’’ AUPEC, 2004.
4. D. J. Trudnowski, ‘‘Estimating electromechanical mode shape from synchrophasormeasurements,’’ IEEE Transcations on Power Systems, Vol. 23, No. 3, pp. 1188–1195,August 2008.
5. T. Kailath, ‘‘An innovations approach to least-squares estimation – Part I: Linear filter-ing in additive white noise,’’ IEEE Transactions on Automatic Control, Vol. 13,pp. 646–655, 1968.
6. R. K. Mehra and J. Peschon, ‘‘An innovations approach to fault detection and diagnosisin dynamic systems,’’ Automatica, Vol. 7, pp. 637–640, 1971.
7. R. A. Wiltshire, P. O’Shea, and G. Ledwich, ‘‘Monitoring of Individual Modal DampingChanges in Multi-Modal Power Systems,’’ AUPEC, 2004.
8. H. VanTrees, Detection, Estimation and Modulation Theory, Part 1. New York: JohnWiley, 1968.
9. P. O’Shea, ‘‘A High Resolution Algorithm for Power System Disturbance Monitoring’’IEEE Transactions on Power Systems, Vol. 17, No. 3, pp. 676–680, Aug. 2002.
10. V. Vittal, ‘‘Consequence and impact of electric utility industry restructuring on transientstability and small-signal stability analysis,’’ Proceedings of the IEEE, Vol. 88,pp. 196–207, 2000.
11. J. F. Hauer, ‘‘Application of Prony analysis to the determination of modal content andequivalent models for measured power system response,’’ IEEE Transactions on PowerSystems, Vol. 6, pp. 1062–1068, 1991.
12. N. Uchida and T. Nagao, ‘‘A new eigen-analysis method of steady-state stability studiesfor large power systems: S matrix method,’’ IEEE Transactions on Power Systems, Vol. 3,pp. 706–714, 1988.
13. G. Ledwich and E. Palmer, ‘‘Modal estimates from normal operation of power systems,’’IEEE Power Engineering Society Winter Meeting, 2000.
14. M.Klein, G. J. Rogers, and P. Kundur, ‘‘A fundamental study of inter-area oscillations inpower systems,’’ IEEE Transactions on Power Systems, Vol. 6, pp. 914–921, 1991.
15. P. Z. Peebles, Probability, random variables, and random signal principles, 4th Ed. NewYork: McGraw Hill, 2001.
16. R. A. Wiltshire, G. Ledwich, and P. O’Shea ‘‘A Kalman Filtering Approach to RapidlyDetecting Modal Changes in Power Systems’’ IEEE Transactions on Power Systems, Vol.22, No 4, Paper No: TPWRS-2007.907529.
17. H. Urkowitz, ‘‘Energy detection of unknown deterministic signals,’’ Proceedings of theIEEE, Vol. 55, pp. 523–531, 1967.
228 G. Ledwich et al.
18. R. A. Wiltshire, P. O’Shea, and G. Ledwich, ‘‘Monitoring of Individual Modal DampingChanges inMulti-Modal Power Systems,’’ Journal of Electrical & Electronics EngineeringAustralia, JEEEA, Vol. 2, pp. 217–222, 2005.
19. J. F. Hauer, C. J. Demuere, and L. J. Scharf, ‘‘Initial Results in Prony analysis of powerSystem Response Signals’’, IEEE Transactions on Power Systems, Vol. 5, No. 1,pp. 80–89, Feb. 1990.
20. N. Zhou and J. W. Pierre, ‘‘Electromechanical Mode Estimation of Power Systems fromInjected Probing Signals using a Subspace Method’’, Proceedings of the 35th NAPSConference, Rolla Mo, Oct. 2003.
21. J. W. Pierre, D. J. Trudnowski, andM. K. Donnelly, ‘‘Initial results in ElectromechanicalMode Identification fromAmbient Data’’, IEEE Transactions on Power Systems, Vol. 12,No. 3, pp. 1245–1251, Aug. 1997.
22. S. M. Kay, Modern Spectral Estimation, Englewood Cliffs, NJ: Prentice-Hall, 1988.23. P. Kundur, Power System Stability and Control, New York: McGraw-Hill, 1994.24. K. Poon and K. Lee, ‘‘Analysis of Transient Stability Swings in Large Interconnected
Power Systems by Fourier Transformation’’, IEEE Transactions on Power Systems,Vol. 3, No. 4, pp. 1573–1579, Nov. 1988.
25. P. O’Shea, ‘‘The Use of Sliding Spectral Windows for Parameter Estimation in PowerSystem Disturbance Monitoring’’, IEEE Transactions on Power Systems, Vol. 15, No. 4,pp. 1261–1267, Nov. 2000.
26. J. F. Hauer, ‘‘Application of Prony Analysis to the Determination of Modal Content andEquivalent Models forMeasured Power System Response’’, IEEE Transactions on PowerSystems, Vol. 6, No. 3, pp. 1062–1068, Aug. 1991.
27. D. J. Trudnowski and J. E. Dagle, ‘‘ Effects of Generator and Static-Load Nonlinearitieson Electromechanical Oscillations’’, IEEETransactions on Power Systems, Vol. 12, No. 3,pp. 1283–1289, Aug. 1997.
28. (2007) [Online] Available http://www.newcastle.edu.au/service/library/adt/public/adt-NNCU20070504.091618/index.html
29. B. Boashash, Time Frequency Signal Analysis and Processing – A Comprehensive Refer-ence, Oxford: Elsevier, 2003.
30. B. Boashash and P. J. Black, ‘‘An Efficient Real-Time Implementation of the Wigner-Ville Distribution’’, IEEE Transactions on Acoustics, Speech and Signal Processing, Vol.35, pp. 1611–1618, Nov. 1987.
31. I. Kamwa and L. Gerin-Lajoie, ‘‘State-Space System Identification – Toward MIMOModels for Modal Analysis and Optimization of Bulk Power Systems,’’ IEEE Transac-tions on Power Systems, Vol. 15, No. 1, pp. 326–335, Feb. 2000.
32. J. J. Sanchez-Gasca, V. Vittal, M. J. Gibbard, A. R. Messina, D. J. Vowles, S. Liu, and U.D. Annakkage, ‘‘Inclusion of Higher Order Terms for Small-Signal (Modal) Analysis –Task Force onAssessing the Need to IncludeHigher Order Terms for Small-Signal (Modal)Analysis’’, IEEE Transactions on Power Systems, Vol. 20, No. 4, pp. 788–793, Nov. 2005.
33. E. Palmer, Multi-Mode Damping of Power System Oscillations, PhD thesis, The Univer-sity of Newcastle, 1998.
34. E. W. Palmer and G. Ledwich, ‘‘Optimal placement of Angle Transducers in PowerSystems’’, IEEE Transactions on Power Systems, Vol. 11, No. 2, pp. 788–793, May 1996.
35. H. Whitehouse, B. Boashash, and J. Speiser, ‘‘High resolution processing techniques fortemporal and spatial signals’’, in High Resolution Techniques in Underwater Acoustics,Lecture Notes in Control and Information Series, New York-Heidelburg-BerlinSpringer-Verlag, 1989.
36. B. Lovell, R. Williamson, and B. Boashash, ‘‘The Relationship between InstantaneousFrequency and Time-Frequency Representations’’, IEEE Transactions on Signal Proces-sing, Vol. 41, No.3, pp. 1458–1461, Mar. 1993.
7 Detection and Estimation of Nonstationary Power Transients 229
Chapter 8
Advanced Monitoring and Control Approaches
for Enhancing Power System Security
Sergio Bruno, Michele De Benedictis, and Massimo La Scala
Abstract Enhancements of power system security can be achieved by develop-
ing an online environment where control center operators have the capability to
monitor in real time the power system dynamic behavior, recognize threats to its
integrity, evaluate, and implement suitable control actions. Themethodological
approaches proposed here are respectively a spectral analysis based on wavelet
transform and a response-based wide-area control for improving power system
dynamic behavior on the transient timescale. Themonitoring technique consists
in processing real-time data coming from phasor measurement units by adopt-
ing an approach based on wavelet spectral analysis. The final result is a diagram
that shows the amplitude of different modes usually associated to different
phenomena (electromechanical modes, inter-area oscillations, etc.) and their
related dampings. The methodology proposed for control assessment is based
on the solution of a dynamic optimization problem whose basic variables are
acquired through a wide-area measurement system.
8.1 Introduction
In recent years, drastic changes in the utility business and an intense sequence
of large blackouts on several major power systems occurred. Many major
countries involved in the restructuring of electric power industry have faced
emergency conditions and blackouts affecting large portions of national and
transnational grids.In the last decades, the management of the transmission systems was based
on well-known system behaviors, full identification of load/generation patterns
and configurations. In a restructured system, the environment is more uncertain
and the system operator (SO) has a scarce direct control over the generation
S. Bruno (*)Dipartimento di Elettrotecnica ed Elettronica (DEE), Politecnico di Bari, Bari, Italye-mail: [email protected]
A.R. Messina, Inter-area Oscillations in Power Systems,Power Electronics and Power Systems, DOI 10.1007/978-0-387-89530-7_8,� Springer ScienceþBusiness Media, LLC 2009
231
side. National and transnational grids have beenmanaged with lack of data andin the presence of a large number of new uncertainties.
The main side effect of the electricity industry restructuring was pushinginvestments toward the generation side, core of the electrical energy business,neglecting security and transmission systems. The protection system, the silentsentinel that guarantees the integrity of the network and provides elementaryconditions for executing any financial and physical power exchange, often isnot upgraded taking advantage of the most updated technologies. In mostcases, transmission networks are exploited close to their capacity limits, withlack of transfer capacity due to the continuous growth of energy demand and todifficulties in building new transmission lines.
In some cases, transmission capability is not limited by localized systemresponse but by wide-region dynamic constraints (i.e., transient stability limits,inter-area oscillations, and voltage collapse). Because the time response is muchfaster than the one associated with thermal overloads, predicting, detecting, andcontrolling such phenomena requires real-time sensors, fast computing, fastcommunication units, and fast actuators sufficiently disseminated all across thegrid.
There are several countermeasures and approaches to system securitythat can be adopted: methodologies for automation and control [1], dynamicsecurity assessment (DSA) and communication systems for real-time dataexchanges [2], adaptive relays [3], FACTS (flexible AC transmission system)and HVDC (high-voltage direct current) technology, and real-time measure-ments and control systems (wide-area measurement systems – WAMS andwide-area measurement and control systems – WAMC) [3, 4]. By means ofphasor measurements units (PMUs) and Global Positioning Systems (GPS)technologies, it is possible to get a state estimate which is a true snapshot of thepower system. Real-time calculations can be performed for a correct estimationof grid capacity on critical sections. Other advanced technologies (FACTS orHVDC)may guarantee the development of flexible transmission systems givingthe SO more tools for controlling power and enhancing the capabilities oftransmission lines.
In the following developments, it is shown how the potentials of wide-areamonitoring and control can be exploited in at least two approaches oriented toimprove power system dynamics on the transient timescale.
WAMS can be adopted in power system monitoring, covering what is arelevant aspect of power system operation and control. For instance, by adoptingsimultaneous synchronizedwide-areameasurements, it is possible to perform real-time monitoring of power system dynamics. An effective monitoring of powersystem dynamic performances can be assessed through wide-area informationand distributed synchronized PMUs. In [5], the IEEE PES Power SystemDynamic Performance Committee had indicated, among the main causes for thepast cascading events, the lack of reliable real-time data oriented to themonitoringof system dynamics. In general, possible failures can be avoided or mitigated byperforming an accurate monitoring of the system [5], keeping track of oscillation
232 S. Bruno et al.
modes during the evolution of operating conditions and detecting the undesirableonset of poorly damped oscillations. Advanced technologies like WAMS, cansupport power system monitoring, providing real-time information about thestate of large portions of national and transnational grids [6].
The usual approach adopted for applying corrective control consists inarming corrective action schemes (remedial action schemes or RAS) whichare triggered only after that the occurrence of a specific fault is detected(event-driven actions). The arming of RAS is based on off-line (preventive)evaluations. Differently, response-based control schemes apply correctiveactions, basing on real-time measurements and on the actual system responseto disturbances and control actions [7].
Ongoing communication technology permits to monitor the status of largeportions of the network, whereas feedback control signals may permit to applyremedial actions in fractions of a second.With this response time, it is conceivableto control a significant number of unstable cases in potentially dangerous areas(critical corridors) [8, 9]. In order to implement these remedial actions, fastactuators are needed. Classically, load/generation shedding can be fast enoughto correct the undesirable transient behavior. More advanced fast devices such asFACTS or line switching can be adopted for the same purpose. In recent years,unified power flow controller (UPFC) has been recognized useful for theenhancement of the power system dynamic behavior [10, 11].
8.2 Monitoring Power System Oscillations by Wavelet Analysis
and Wide-Area Measurements
8.2.1 Approaches for Monitoring Power System Oscillations
Monitoring power system oscillations can be accomplished by means of severalmethodological approaches. Undoubtedly, the eigenvalue analysis is the mosteffective method for studying power system small-signal stability. Major limita-tions of this approach are related to the computational burden in the presenceof large networks.Moreover, the effectiveness of this method can be affected bythe unavailability of a detailed representation or by the onset of nonlinearbehaviors (saturations or relays’ triggering) during stressed conditions or inpresence of large disturbances.
In [12], data from transient stability simulations are processed by meansof Hilbert spectral analysis, in order to characterize the time evolution ofnonstationary power system oscillations, and eventually tracking the dynamics ofcritical system modes. Other feasible approaches for the identification of powersystemmodes of oscillation are based on the direct spectral analysis of power systemresponse.
The Fourier transform is largely adopted for studying stationary signals,where the properties of signals do not evolve in time. Nevertheless, this method
8 Advanced Monitoring and Control Approaches 233
is less efficient in tracking nonstationary signals, where any abrupt change mayhave effect on the whole frequency axis [13]. A usual solution is the short-timeFourier transform (STFT) that employs a time window for mapping a signalinto a two-dimensional function of time and frequency. The drawback is thatonce a particular size for the time window is chosen, that window width has tobe kept constant for all frequencies.
In [14], a technique for analyzing system oscillations by employing the Fouriertransformation is presented. The major frequency components are identified inthe frequency domain and the corresponding damping constants are determinedcomparing the magnitudes of the respective frequency components in differentsubsequent time windows. The Fourier transform is commonly employed inpower network analysis and has the advantage of not requiring a heavy compu-tational effort. On the other hand, this method can be efficient only when aimedat finding a mode of oscillation whose frequency is already known. Finding anunknown frequency can be difficult and require several attempts with windowshaving different widths [15].
The Prony method decomposes time-domain signals into a sum of dampedsinusoids, each characterized by four parameters: frequency, damping, ampli-tude, and phase. This method is able to deal with the nonlinearities of powersystems and can also be applied to field measurements. The size of the model isnot limited by the technique because only the output is analyzed [16]. A scientificvalidation of the Pronymethod as a valuable tool in estimating themodal contentof power system oscillations and its methodological improvements have beenshown in [17, 18].
A new approach to online assessment and control of transient oscillationsis proposed in [19] taking into account Prony analysis and improving itsapplicability and reliability. Comparisons of Prony and eigenvalue analysis, high-lighting potentials and drawbacks of eachmethod, have been presented in [16, 20].In [21], a technique for extracting power system modes based on spectral analysishas been proposed. The approach, adopting the Z-transform identification algo-rithm of Corinthios, permits the identification of the damping factor and thefrequency of signal containing multiple modes.
Wavelet spectral analysis is based on a variable-sized windowing technique.A major advantage of wavelets is the ability to perform local analysis, i.e., toanalyze a localized area of a larger signal. Wavelet transform (WT) algorithmsprocess data with different scales, stretching or compressing the basic waveletfunction, providing a multiresolution analysis in both frequency and time.Wavelet analysis allows using respectively long time intervals for capturinglow-frequency information, and shorter windows for high-frequency informa-tion so that the characteristics of typical nonstationary power disturbances canbe efficiently captured and assessed.
WAMS technology, applied to real-time supervision of power systems is atechnology at hand, but security monitoring of system dynamic performances iscurrently a challenge in terms of operator interface implementation. From anoperator viewpoint, an overall perspective of relevant processes is important
234 S. Bruno et al.
and useful in order to achieve the online supervision of power system dynamicperformances. A contribution to this issue is given in [22], where direct spectralWT-based analysis is proposed as man–machine interface (MMI) for dynamicperformance monitoring and DSA. Analogously, in [23], a WT-based spectralanalysis approach has been applied on real data measured by PMUs duringan experiment on large scale in the Russian Far East Interconnected PowerSystems. The approach allows the detection of the disturbance and themonitoringof the behavior of principal modes of oscillation.
8.2.2 Morlet-Based Wavelet Analysis
Wavelet analysis can be performed adopting different formulations of the WT.Each formulation is characterized by specific features (and suitable applica-tions); for example, the WT can be continuous or discrete, real or complex, andis characterized by the shape given to the mother wavelet function.
The continuous wavelet transform (CWT) offers a good resolution, althoughit requires a significant computational burden and memory usage. Very often inpractical application, the discrete wavelet transform (DWT) can be adopted. Itpermits to discretize scale and time parameters, and its transform is not acontinuous function of the frequency. The DWT is fast and reduces redundantinformation and computational efforts.
Real and complex wavelets give, in general, different results. A real waveletproduces high values of magnitude when the oscillation reaches a maximum (ora minimum), or in the presence of sharp discontinuities. On the other hand, thecomplex wavelet is able to produce magnitude all along the duration of anoscillating signal.
In the proposed approach, the complex continuous wavelet can be adoptedbecause fast filtering and signal reconstruction that usually require high compu-tational efforts are not necessary.Moreover, the availability of continuous valuesof magnitude all along the duration of the signal, increases the feasibility of theapproach in terms of MMI, avoiding the discontinuity created by real CWTs.
Regarding the formulation of the mother wavelet, several functions havebeen proposed in the literature (Gauss,Morlet,MexicanHat, Gabor, etc.), eachhaving its suitable applications [9, 24]. In this approach, the continuous com-plex Morlet WT has been adopted because it offers a very good resolution inboth time and frequency: its frequency resolution is adequate even at the lowestwave number. Furthermore, as shown in the followings, the Morlet waveletpreserves the information on damping in the time–frequency domain [25–27].
The CWT of a time domain signaly tð Þis defined as
CWTcy b; að Þ ¼ 1ffiffiffi
ap
Zþ1
�1
yðtÞc� t� b
a
� �dt (8:1)
8 Advanced Monitoring and Control Approaches 235
where c� tð Þ is the complex conjugate of the analyzing (or mother) wavelet c tð Þ,a is the dilation parameter, whereas b is the parameter that sets the waveletfunction in the time domain.
The complex Morlet wavelet, defined in its curtailed form [27], is
cðtÞ ¼ e�t2
N e j!0t (8:2)
where !0 is the central frequency (wave number) of the wavelet and N (withN > 0) is a parameter that controls the shape of the mother wavelet.
The following discussion is aimed at showing how the Morlet WT is able topreserve damping information in the frequency–time domain.
A general harmonic signal can be formulated as
yðtÞ ¼ AðtÞ cos ’ðtÞð Þ (8:3)
where AðtÞ and ’ðtÞ are time-varying envelope and phase functions.In [27], it has been shown how, if the signal expressed in (8.3) decays to zero
at +1, and therefore,
Zþ1
�1
yðtÞj j2 dt5þ1 (8:4)
then its WT given by Eqs. (8.1) and (8.2) is
CWTcy ðb; aÞ ffi
ffiffiffiap
2
ffiffiffiffiffiffiffiNpp
AðbÞe�N4 a _’ðbÞ�!0ð Þej’ðbÞ (8:5)
A particular application of given formulas is the calculation of the WT of asignal y(t), representing the response of a system governed by a second-orderlinear differential equation:
yðtÞ ¼ A0e��t cos !dtþ ’0ð Þ (8:6)
where � ¼ �!n is the damping defined by the product of the natural pulsation !nand damping ratio �, !d ¼ !n
ffiffiffiffiffiffiffiffiffiffiffiffiffi1� �2
pis the pseudopulsation, and ’0 is the
initial phase value.In this case, for a fixed value of the dilation parameter a = ai, modulus and
phase of the CWT of the signal represented in (8.6) are
CWTcy ðb; aiÞ
������ ¼
ffiffiffiffiaip
2
ffiffiffiffiffiffiffiNpp
AðbÞe�N4 ai!d�!0ð Þ2 (8:7)
ff CWTcy ðb; aiÞ
h i¼ !dbþ ’0 (8:8)
236 S. Bruno et al.
By fixing two time instants b1 and b2, from (8.7) and (8.8) one derives:
CWTcy ðb1; aiÞ
������
CWTcy ðb2; aiÞ
������¼ A0e
��b1
A0e��b2¼ e� b2�b1ð Þ (8:9)
and hence:
� ¼ 1
ðb2 � b1Þlogn
CWTcy ðb1; aiÞ
������
CWTcy ðb2; aiÞ
������
(8:10)
Equation (8.10) shows how the WT of a signal y tð Þ allows identifying the
damping associated to each oscillation mode contained into the signal.The normalized dB value of the CWT magnitude can be expressed as
CWTðb; aiÞ ¼ 20Log10CWTðb; aiÞj jCWTmax
(8:11)
Considered two time instants b1 and b2, with b2> b1, themagnitude variation
�dB ¼ CWTðb2; aiÞ � CWTðb1; aiÞ can be expressed as
�dB ¼ 20Log10CWTðb2; aiÞj jCWTðb1; aiÞj j ¼
20
logn 10logn
CWTðb2; aiÞj jCWTðb1; aiÞj j (8:12)
Considered that logða=bÞ ¼ � logðb a= Þ, from (8.12) one derives:
lognCWTðb1; aiÞj jCWTðb2; aiÞj j ¼ �
�dB logn10
20(8:13)
If the analyzed mode can be represented as in Eq. (8.3) and has an envelope
decreasing exponentially with time, (8.13) can be substituted in (8.10), obtaining
Eq. (8.14):
� ¼ ��dB logn 10
20ðb2 � b1Þ¼ ��dB logn 10
20�b(8:14)
The wavelet function proposed in (8.2) has also the property of linearity and
therefore it is easy to extend the same WT approach to a signal yðtÞ composed
by a sum of different signals.
8 Advanced Monitoring and Control Approaches 237
8.2.3 A WAMS-Based Monitoring Architecture
The proposed monitoring architecture is aimed at detecting large disturbancesoccurring in a control area or in external systems, evaluating the damping ofrelevant oscillation modes related to electromechanical transients.
A pictorial representation of the proposed centralizedmonitoring architectureis shown in Fig. 8.1.
PMUs, located in strategic points of the power system (substations, genera-tion buses, or important interconnectors), provide real-time measurements ofvoltage and current phasors (positive sequence). GPS provides synchronizingsignals for voltage and current samples. The PDC block represents the phasordata concentrator that collects, synchronizes, and normalizes all data sent byPMUs through the communication system, and exports them as a data stream.Dotted lines in Fig. 8.1 represent the communication links: for this architecture,both wired and wireless communication links can be adopted. In [28], an over-view of communication alternatives is given, assessing the related time delayswhich are in the range 100–700 ms.
As soon as measured trajectories are received in the Control Center, a WTdiagnosis is carried out for detecting and analyzing themodes of oscillation of thepower system. The final result is a wavelet mapwhich shows, in a suitable scale ofcolors, the amplitude of different frequencies (modes) usually associated todifferent phenomena (electromechanical modes, inter-area oscillations, etc.)and an assessment of the related damping. The proposed approach providesresults in a compact and concise form that can be adopted as useful tool forMMI. The same interface can be adopted in an off-line mode through a suitableOTS (operator training simulator) environment. By this tool, operators can betrained to identify and handle dynamic threats.
Fig. 8.1 Representation of the proposed monitoring architecture
238 S. Bruno et al.
The supervision of the system can be performed by analyzing only signals
provided by several PMUs located in the network. Real-time voltage trajectories,
differently from system state variables, can be directly measured and then adopted
formonitoring dominant oscillationmodes. In general, it is easy to imagine that the
choice of the PMUs’ location should be made basing on off-line calculations of
mode observability matrices or on the experience in system operation. Such selec-
tion is aimed at monitoring only those voltages that proved to be able to better
observe the oscillation modes that must be kept under surveillance.
8.2.4 Test Results A: Monitoring System Responseto Small Perturbations
This section is aimed at showing how principal modes of oscillation can be identified
bymeans ofWT-based spectral analysis and how the complexMorletWT preserves
the main properties of these modes (frequency and damping) in the time–frequency
domain. More explicitly, eigenvalue analysis is compared to the WT-based direct
spectral analysis in a test case where small-signal analysis is performed.The eigenvalue analysis and the dynamic simulation were carried out on a
test grid obtained bymodifying the IEEE 30-bus ‘‘New England’’ Dynamic Test
Case [29]. The main differentiation from the latter is that the adopted grid has
six generators instead of two generators and four synchronous compensators
(Fig. 8.2). Each generator is represented with a fourth-order model and is
equipped with a governor and an exciter.The eigenvalue analysis of the system for a specific operating point led to the
identification of several modes of oscillation. In Table 8.1, selected modes of
oscillation and their characteristics are reported.
Fig. 8.2 Scheme of the modified IEEE 30-bus test grid
8 Advanced Monitoring and Control Approaches 239
In order to study the dynamic response to small signals, the system has been
perturbed by applying a 0.1% load increase at bus #18. The Complex Morlet
WT has been employed for analyzing the rotor angle trajectories subsequent to
the disturbance. Figures 8.3 and 8.4 show the wavelet chart of the rotor angle
signals for two generators, at bus #8 and at bus #11, respectively. TheWTof the
rotor angle signals showed that three modes of oscillation, namely, modes #3,
#5, and #6, are dominant. Mode #4 is characterized by frequency and damping
values close to mode #3.A comparison between the eigenvalue analysis and the spectral analysis can be
performed through the participationmatrix. In fact, the participationmatrix, obtained
by combining the left and the right eigenvectors, is able to measure the coupling
betweenmodes of oscillation and state variables. The elements of thismatrix are called
participation factors, and eachof themgives ameasure of the relative participationof a
state variable to a mode and vice versa. In Table 8.2, the participation factors of
principal modes, with respect the rotor angle state variables, are reported.
Table 8.1 Features of principal modes of oscillation
Mode EigenvalueFrequency(Hz)
Dampingratio � !n (rad/s)
1 –1.35� j 9.18 1.46 0.145 9.282 –1.20� j 8.07 1.28 0.147 8.163 –0.23� j 7.61 1.21 0.030 7.614 –0.38� j 7.30 1.16 0.052 7.315 –0.56� j 5.56 0.88 0.100 5.596 –0.18� j 3.89 0.62 0.046 3.897 –6.29� j 1.76 0.28 0.963 6.538 –6.23� j 1.73 0.28 0.964 6.47
Fig. 8.3 Wavelet spectral analysis of the rotor angle trajectory at bus #8
240 S. Bruno et al.
Thewavelet chart of a signal can be represented in a three-dimensional space:
frequency,time,andmagnitude.Mode#5,characterizedbya0.88Hzfrequency,isclearlyvisible inFig. 8.3 (rotorangle trajectoryatbus#8) and in theotherwavelet
charts (not shown). InFig. 8.4,mode#5 is less discernible, according to the lower
participationfactorwithrespect togeneratoratbus#11.Modes #3 and #6 are mostly related to the behavior of generator at bus #11,
and are neatly discernible only in Fig. 8.4. This is confirmed by the participation
factors that, for this generator, differ by an order of magnitude from the ones
related to the other generators. In particular, mode #6, being characterized byvery low participation factors with respect to the rotor angle and by high
participation factors with respect to regulators’ state variables, is related to
the local control of the machine at bus #11.The same oscillation modes can be identified by applying the WT to the
voltage magnitude signals. Figures 8.5 and 8.6 show the wavelet chart of the
voltage magnitude signals at the generator at bus #8 and the one at bus #11,
respectively.In Figs. 8.5 and 8.6, the same dominant oscillation modes can be identified.
The possibility of discerning amode, observing a specific trajectory, depends on
Fig. 8.4 Wavelet spectral analysis of the rotor angle trajectory at bus #11
Table 8.2 Participation factors of the principal modes with respect to the rotor angle statevariables
General bus Mode #3 Mode #5 Mode #6
#2 0.0287 0.0765 0.0012
#5 0.0248 0.0732 0.0015
#8 0.0496 0.1838 0.0036
#11 0.4858 0.0513 0.0439
#13 0.0824 0.0978 0.0022
8 Advanced Monitoring and Control Approaches 241
the corresponding values in the mode observability matrix. In this case, mode
#6, previously hardly discernible by observing rotor angle trajectories at bus #8,
can be clearly identified by analyzing voltage trajectories at the same bus.By ‘‘cutting’’ the three-dimensional representation in correspondence of the
frequency of a specific mode, it is possible to represent the magnitude of that
mode as a function of time. Figure 8.7 has been obtained in correspondence of
0.88 Hz (the frequency of mode #5) by the WT previously shown in Fig. 8.4.As shown in Fig. 8.7, the magnitude of mode #5 increases in the first seconds
and, having reached a maximum, decreases exponentially with time (in the
Fig. 8.5 Wavelet spectral analysis of the voltage magnitude at bus #8
Fig. 8.6 Wavelet spectral analysis of the voltage magnitude at bus #11
242 S. Bruno et al.
figure the decrease is linear because the dB scale is a logarithmic scale).The
increasing behavior is due to the necessity for the algorithm to acquire the main
modes; consequently, damping evaluated in this phase cannot be utilized for
stability assessment.Measuring the steepness of the curve, it is possible to derive
the damping of the mode #5, already known by its eigenvalue.As demonstrated in Eqs. (8.11, 8.12, 8.13), if the mode decreases exponen-
tially, having a magnitude variation �dB in the time interval �b, the damping
can be evaluated with relation (8.15):
� ¼ ��dB logn 10
20�b� � 0:115�dB
�b(8:15)
Figure 8.7 shows how, when the mode decreases exponentially, a reduction of
14.6 dB is obtained in 3 s (from 4 to 7 s). From Eq. (8.15), the damping �=0.56.
This value is identical to the one evaluated through the eigenvalue analysis
(Table 8.1), confirming the fact that the complex Morlet WT preserves informa-
tion on oscillation modes’ damping.
8.2.5 Test Results B: Influence of Random Fluctuations Dueto Operating Conditions
Previous simulationswere performedon an ideal systemmodel not affected by noise:
in such case, all system parameters are assumed constant and deterministically
known. In real systems, operating conditions are in general not constant, as they
are affected by random fluctuations. It would be profitable, in this study, to explore
the performances of the proposed monitoring tool in presence of these random
fluctuations denoted here as ‘‘system noise.’’In general, representing the random variations of all system parameters is
not an easy task and many examples can be found in literature [30]. System
noise in the following tests has been represented simply by the fluctuation of
0 1 2 3 4 5 6 7 8 9Time
–25
–20
–15
–10
–5
0
dBN
orm
aliz
ed
–6.39 dB
–21.03 dB
Fig. 8.7 Spectral WT analysis of the rotor angle trajectory at bus #8, at 0.88 Hz
8 Advanced Monitoring and Control Approaches 243
load demand and generation profiles. This choice derives by the rationale that
load demand and generation are major source of randomness in power systems.According to [31], load has beenmodeled as the sumof two terms: a deterministic
component plus a stochastic additive disturbance term. Even if this assumption
might seem too simplifying, certain stability properties can be preserved if the
disturbance term is represented by a stochastic process close to white noise
[31–33]. This is also true in the context of large deviation phenomena.In the following tests, active power demand has been represented as the sum
of a deterministic voltage-dependent term plus an additive stochastic time-
dependent term, modeled as Gaussian white noise. In practice, at each time
step of the integrating routine (0.02 s), each load has been considered affected,
conservatively, by a random load variation having Gaussian distribution and
variance 0.1% with respect to the deterministic average load term (evaluated at
1 pu). Reactive power is recalculated considering a constant power factor.In order to simulate the dynamic response of a full-scale power system, a
representation of the Italian national grid, characterized by an adequate degree
of detail, was adopted for these tests. Themodel includes detailed information on
the external systems and is characterized by about 1,333 nodes, 1,762 lines,
273 generators, and 769 transformers. The steady-state condition refers to a
system configuration during aWednesday (peak load day) at the first daily peak.The test consisted in the simulation of the transient subsequent to the
tripping of one of the two 400 kV circuits of the transmission line Latina–Gar-
igliano in Central Italy (see Fig. 8.8). The transmission line, at the moment of
the tripping, was carrying about 350 MW.The spectral analysis of the voltage trajectory at the Garigliano bus (Fig. 8.9)
shows how the insurgence of the event can be easily identified. The effect of system
noise is negligible with respect to voltage fluctuations due to the line tripping.
Fig. 8.8 Portion of the Italian 400 kV transmission system
244 S. Bruno et al.
The spectral analysis has been performed on the voltage trajectory at the
20 kV generating bus inGarigliano. All simulations start at time t¼ 0 s, whereas
the line tripping occurs at the time instant t¼ 10 s. Therefore, in the first 10 s the
system dynamic is affected only by the system noise. In Fig. 8.9, dominant
oscillation modes are clearly visible. In particular, the spectral analysis at
1.10 Hz, represented in two dimensions in Fig. 8.10, shows amode of oscillation
characterized by an exponential decreasing behavior and damping that, accord-
ing to Eq. (8.15), can be estimated in
Fig. 8.9 Spectral WT analysis of the voltage magnitude trajectory at the Garigliano bus(400 kV)
Fig. 8.10 Spectral WT analysis of the voltage magnitude trajectory at the Garigliano bus(400 kV) at 1.10 Hz
8 Advanced Monitoring and Control Approaches 245
� � � 0:115ð�2Þ13:87� 12:79
� 0:213
Other tests, not represented here for the sake of brevity, showed howdisturbances like the tripping of a transmission line can be detected by analyzingthe voltage magnitude behavior at almost any node of the Italian power grid,with no regard to the actual geographical localization of the disturbance. Forexample, the tripping of a 400 kV transmission line in Central Italy (Villano-va–Larino), almost unloaded and carrying less than 50 MW, was found to berecognizable by performing the WT spectral analysis of the voltage magnitudeat the Priolo bus in Sicily (see Fig. 8.8).
Results are consistent with the experience and the modus operandi of controlcenter operators of the Italian transmission system, that usually, in order todetect the onset of large oscillations, keep under surveillance few selected nodesof the grid.
8.2.6 Test Results C: Influence of Measurement Noise
In this section, further simulations are presented, aiming at showing theperformances of the proposed method in the contemporaneous presence of‘‘system noise’’ and measurement noise.
The accuracy of a PMU can be estimated to be about 0.1% for voltagemagnitude [33]. Hypothesizing that voltage trajectories were measured by aPMU, simulated voltage trajectories were ‘‘spoiled,’’ overlapping the simulatedsignal with white noise, having variance 0.03%, with respect to the nominalmeasured value (around 1 pu).
The simulated trajectory at the 400 kV Garigliano bus, analyzed in theprevious section, was modified obtaining what is shown in Fig. 8.11. Thespectral WT analysis of the ‘‘measured’’ signal, shown in Fig. 8.12, is able tohighlight the dominating oscillationmodes (1.10 and 0.74Hz), showing how thedamping information is preserved.
The evaluation of the damping can be performed also in this case,representing the spectral analysis at 1.10 Hz, in two dimensions (Fig. 8.13).The represented mode is not characterized by an exponential decreasing beha-vior all through the transient. Therefore, the damping has been evaluated infirst part of the trajectory were the trajectory decreases linearly in the logarithmscale.
According to Eq. (8.15), the damping calculation provides a result close tothe one obtained in the previous section:
� � � 0:115ð�2Þ13:82� 12:78
� 0:221
246 S. Bruno et al.
8.3 Response-Based Wide-Area Control Approach
In this section, the effects of the implementation of a real-time corrective
control architecture is discussed. The scheme of the proposed control architec-
ture is quite simple, consisting of aWAMS, a Control Center, and a network of
fast actuators, all connected through a suitable communication system.
Fig. 8.12 Spectral WT analysis of the voltage magnitude trajectory at the Garigliano bus(400 kV), in the presence of measurement noise
Fig. 8.11 Voltage magnitude trajectory at Garigliano bus (400 kV), in the presence ofmeasurement noise
8 Advanced Monitoring and Control Approaches 247
The main outline of such architecture is inspired by the special protection
schemes (SPSs) usually adopted in power systems in order to control critical
events. Such systems are designed for applying event-driven corrective actions,
that have been precalculated off-line for a limited number of postulated con-
tingencies, only after the occurrence of the fault. Such systems can be enhanced
taking advantage of real-time measurement system such as the one provided by
WAMS and PMU technology.The main idea is that real-time data acquired by PMUs are employed by the
Control Center in order to implement a response-based control strategy [7]. On
the basis of measured power system trajectories, when network integrity is
threatened, the Control Center elaborates response-based corrective control
actions, and is able to correct the system behavior after the actual occurrence of
large disturbances. Remedial actions may then be applied through any fast
actuator device (e.g., load shedding schemes or FACTS devices).The proposed wide-area control scheme can be defined as generalized
response-based global control architecture. In this context, global control
means that centralized control is performed observing static and dynamic
phenomena on a sufficiently wide-area range, allowing an integrated use of all
available control actions. The necessity of adopting a centralized control struc-
ture is based on the consideration thatmajor electrical infrastructure catastrophic
events are associated to dynamic phenomena (inter-area oscillations, voltage, and
transient instabilities) having wide-area geographical influences.The proposed control scheme can be designed in order to lay on an upper
hierarchical control level, overlapping local control systems. This upper level
works coordinately with lower levels, optimizing their functionalities (i.e.,
modifying protection and local control system set points). The upper control
Fig. 8.13 Spectral WT analysis of the voltage magnitude trajectory at the Garigliano bus(400 kV), at 1.10 Hz, in the presence of measurement noise
248 S. Bruno et al.
level improves performances of local controllers which are not able to evaluatewide-area geographical phenomena.
In this section, amethodology for assessingwide-area real-time corrective controlactions is presented. Such methodology takes advantage of the approaches devel-oped in [34, 35]. The methodology, based on nonlinear programming technique, isaimed at assessing real-time corrective control actions, in order to guarantee powersystem static and dynamic security. Real-time control actions are evaluated exploit-ing measured system trajectories in the solution of a dynamic optimization problemfor nonlinear systems, where static and dynamic inequality constraints are takeninto account. The approach is prone to be applied for improving the systemdynamicbehavior on different timescales. However, in this section we show the potentialsof the approach considering stability problems linked to the transient timescale.
The structure of the mathematical approach hereby presented is based on theapproach developed for event-based evaluation of corrective actions. In theevent-based approach, corrective actions are evaluated off-line on the basis ofprepostulated contingencies and applied after the actual occurrence of thefaults. Differently, in the response-based approach, corrective control has tobe operated online, basing calculations onmeasured values (based on the actualresponse of the system).
8.3.1 Mathematical Formulation
In general, power system behavior on a transient timescale can be describedthrough a set of nonlinear differential and algebraic equations (DAEs):
_xðtÞ ¼ fðxðtÞ;VðtÞ; uðtÞÞgðxðtÞ;VðtÞ; uðtÞÞ ¼ 0
(8:16)
where x is the state vector, u is the control variable vector, andV is the vector ofnodal voltages.
The DAEs set (8.16) can be discretized through a trapezoidal rule andwritten in implicit form as follows:
Hðy; uÞ ¼ 0 (8:17)
where
Hiðyi; uÞ ¼ 0; i ¼ 0; 1; 2; . . . ; nT (8:18)
yi ¼ xTi VTi
� �T(8:19)
y ¼ yT0 yT1 � � � yTi � � � yTnT
� �T(8:20)
H ¼ HT0 HT
1 � � � HTi � � � HT
nT
� �T(8:21)
8 Advanced Monitoring and Control Approaches 249
with yi representing the composition of all state variable and voltage vectors atthe generic ith time step; consequently, y represents the discretization of thewhole trajectory of the system;Hi is the discretization of the DAEs set (8.19) atthe generic ith time step; nT is the total number of time steps relative to theintegration interval [0,T].
The methodology proposed in this chapter takes advantage of the oneadopted in [34, 36]. Corrective control actions are evaluated through theformulation and the solution of a dynamic optimization problem for nonlinearsystems, where static and dynamic inequality constraints are taken intoaccount. In order to solve it, the problem is reformulated in terms of staticoptimization (nonlinear programming) and solved by applying the Lagrangianmultipliers method (nonlinear programming).
Themain difference with regard to [34, 36] is that, instead of formulating andsolving a single optimization problem, this method aims at finding the optimalvalue of the control vector uw that solves the optimization problem given foreach time window tw by the equations
minuw
CU uwð Þ þ CO ywð Þ þ CP ywð Þð Þ (8:22)
subject to
H yw; uwð Þ ¼ 0 (8:23)
and
uw min uw uw max (8:24)
In (8.22), yw represents the system trajectory in the time window tw, CU
represents an objective function aiming at the minimization of the controllingeffort, and CO is the objective function whose scope is to improve the dynamicbehavior of the system. Typically, the objective function can be formulated asan integral norm of deviations of selected variables with respect to a desirablebehavior across time; CP is a penalty function that takes into account inequal-ity constraints. Usually, inequality constraints define a time-varying domainwhere the trajectories of the system should be contained in order to satisfypractical requirements about the dynamic performances of the system.Different objective and penalty functions have been tested during the lastdecade [34–37].
A commonly adopted objective function, aimed at minimizing the transientkinetic in the generic time window tw, can be expressed as follows:
CKTEðywÞ ¼ �KTE
X
i2twVi
KTE (8:25)
250 S. Bruno et al.
where ViKTE denotes the kinetic transient energy evaluated around the center of
inertia (COI) at ith time step, �KTE is a coefficient which takes into account therelative weight of this objective function.
A penalty function that takes into account voltage-related inequalityconstraints can be formulated as follows:
CVðywÞ ¼ �V
Pi2twPn
j¼1Vi;j� �Vlim
�i;j
� 2
ntwn(8:26)
where
�Vlim ¼VM
i;j if Vi;j4VMi;j
Vmi;j if Vi;j5Vm
i;j
Vi;j otherwise
8><
>:and �i;j ¼
Vi;j
1
if Vi;j5Vmi;j
otherwise
In (8.26), ntwrepresents the number of steps in window tw; Vi;j represents thejth bus nodal voltage at the ith time step; Vm
i;j and VMi;j are time-varying thresh-
olds fixed to ensure a desirable voltage transient; �i,j is a weight factor adoptedto avoid dangerous low voltage conditions such as the ones associated totransient voltage instability.
By applying the optimizationmethod of Lagrangianmultipliers, it is possibleto evaluate the solution of the problem stated above:
L ¼ CUðuwÞ þ CO ywð Þ þ CP ywð Þ þ �TH yw; uwð Þ (8:27)
From (8.27), the set of necessary conditions follows:
@L
@yw¼ @ðCO þ CPÞ
@ywþ �T @H
@yw¼ 0 (8:28)
@L
@uw¼ @CU
@uwþ �T @H
@uw¼ 0 (8:29)
@L
@�¼ H yw; uwð Þ ¼ 0 (8:30)
From (8.28) and (8.29) one derives:
@L
@uw¼ @CU
@uw� @ðCO þ CPÞ
@yw
@H
@yw
" #�1@H
@uw(8:31)
8 Advanced Monitoring and Control Approaches 251
When the optimization is operated off-line adopting the event-basedapproach, such optimization problem is solved through an iterative algorithm.The control vector is updated through the equation
uneww ¼ uoldw þ �@L
@uw(8:32)
Then a new trajectory is simulated, and everything starts back from Eqs.(8.28, 8.29, 8.30, 8.31), beginning a recursive approach that stops when thesensitivity term @L=@uw is lower than a specific tolerance limit.
In the response-based approach, proposed here, there are two main differ-ences. First, the trajectory yw is acquired throughWAMSand cannot bemodifiedor resimulated. Thatmeans that the control vector, evaluated for optimizing suchtrajectory, can be applied only afterward the trajectory itself was acquired.Moreover, as calculations of response-based control actions must be performedwithin the shortest time, a research of the optimal solution by means of thesolution of the Eqs. (8.28, 8.29, 8.30, 8.31) is not suitable.
Moreover, it should be remembered that suboptimal solutions are acceptableprovided that stability is ensured. Consequently, at each time window, changeson control variables are evaluated through a simple sensitivity analysis (i.e.,it corresponds to the first step of the iterative process described before) andapplied in the subsequent time window:
uwþ1 ¼ uw þ �@L
@uw(8:33)
where
@L
@uw¼ � @ðCO þ CPÞ
@yw
@H
@yw
" #�1@H
@uw(8:34)
Because a sensitivity approach has been adopted, the function CU gives nocontribution to derivatives in Equation (8.33) at the first iteration. However,neglecting CU is conservative from the security point of view because, in thisway, the control action is evaluated without minimizing the controlling effort.
The basic idea for this approach is that, after calculations, response-basedcorrective control actions calculated with Eqs. (8.33) and (8.34) are applied withan overall time delay �. After the implementation of the control actions, the nextstep of the proposed approach starts when a new piece of trajectory is acquired.If necessary, new corrective actions are evaluated and applied on the next timewindow, and so on. A pictorial representation of this approach is given inFig. 8.14.
The overall time delay � takes into account the time necessary for dataacquisition from WAMS, data transmission to the Control Center, data
252 S. Bruno et al.
synchronization and CPU elapsing time for control action assessment, datatransmission to actuators, and triggering of the corrective control actions. Cleary,this delay is a key parameter for the feasibility of the approach and the roughquantification of its maximum acceptable value is addressed through simulationson a real representation of the Italian power systems.
8.3.2 Test Results
The proposed methodology has been tested simulating its functionality duringoperation of a real power system. Tests were carried out considering the samerepresentation and operating condition described in Sections 8.2.5 and 8.2.6,worsened introducing a 4% uniformly distributed load increase. The base caseconsisted in the simulation of the electromechanical transient subsequent to theoccurrence of a three-phase fault at the 400 kV Aurelia bus in Central Italy,cleared after 0.2 s through the tripping of the 400 kV Aurelia–Roma transmis-sion line.
For tests, two different corrective control schemes were chosen. In test #1control actions were implemented by means of load shedding schemes whereastest #2 has been carried out considering FACTS as actuators of correctivecontrol actions. Test #1 refers to a technology actually implemented on theItalian power system, whereas test #2 hypothesizes the adoption of an advancedtechnology such as the one provided byUPFCs. The adoptedmethodology and
Fig. 8.14 Schematization of the proposed response-based approach
8 Advanced Monitoring and Control Approaches 253
the illustrated mathematical formulation are suitable for both tests. Never-
theless, the methodology implementation and the solution of the optimization
problem are slightly different due to a substantial difference in the composition
of the control variable vector u. Specific considerations for the solution of both
optimization method can be found in [35, 37].As already mentioned, estimating the maximum acceptable delay for
the implementation of response-based corrective control actions has crucial
relevance in assessing the feasibility of the proposed methodology. It also
permits to assess communication technical requirements, computational time,
and actuators’ speed of response. In order to estimate this parameter, in both
tests, simulations have been carried out considering increasing values of the
time delay �.As represented in Fig. 8.15, the chosen test case gives rise to the unstable
behavior due to the fault on the line Aurelia–Roma that produces negative
effects on the whole grid. In particular, generation bus at Termini I. power
station (in Sicily) shows a behavior clearly unstable (Fig. 8.15). The final
undesirable effect is the separation of Sicily from the National Transmission
Grid, and the loss of synchronism of most generators in Southern Italy.In test #1, the proposed algorithm for response-based control of the transient
phenomena allows evaluating the performances of a real-time corrective
control scheme based on load shedding. Beside the base unstable case, four
different simulations were carried out by changing the value of delay �.The values given to � during test #1 are 0 s (ideal case where the control
action is applied instantaneously), 0.2, 0.3, and 1 s. Results of the simulation
performed during test #1 are shown in Fig. 8.15 monitoring the behavior of the
rotor angle at the same generator bus (Termini I.). The generator behavior
when no control is applied, is clearly unstable, as is easily recognizable in the
Fig. 8.15 Rotor angle of a generator at Termini I. (test #1)
254 S. Bruno et al.
picture. The comparison of results, obtained by increasing the � values, showsthat only delays up to 1.0 s can ensure stability (Fig. 8.15).
During test #2, the corrective control actions have been implemented by two
400 kVAUPFCs installed, respectively, on Villanova–Larino and Valmontone–-
Presenzano transmission lines (therefore located along the two critical corridors
for power exchanges between Central and Southern Italy, see Fig. 8.8). The time
delays applied in this case are: 0.02, 0.1, and 0.3 s. Note that the time delay 0.02 s
is an ideal case where the control action is applied almost instantaneously
(with just one time step of delay). Results of the simulations performed during
test #2 are shown inFig. 8.16.Delays up to 0.3 s can ensure stability, whereas with
longer delays the control scheme is ineffective. This result is consistent with other
simulations not reported here for brevity.The two estimated values of the maximum acceptable time delay (1 s for test
#1 and 0.3 s for test #2) permit to compare the performance of the different
actuators adopted. Longer feasible time delays in test #1 can be explained
considering that the load shedding action is a more appropriate and strong
action with regard to the redistribution of power flows operated by the UPFC-
based control actions. The load shedding scheme proves to be an effective, even
if expensive, remedial action which permits to stabilize the unstable behavior of
the system with an overall time delay of about 1 s.The control action actuated byUPFC devices is not so drastic and incisive as the
load shedding in recovering unstable behavior. However, the exploitation ofUPFC
devices allows avoiding generation rescheduling or load curtailments even under
large perturbations by means of a power flows redistribution on the network.
Furthermore, there is no need for minimizing control efforts because variations
on UPFC reference signals do not yield, in general, further operative costs. It needs
just to verify that control actions are compatible with actual machine limitations.
Fig. 8.16 Rotor angle of a generator at Termini I. (test #2)
8 Advanced Monitoring and Control Approaches 255
8.3.3 Computational and Communication TimeDelay Assessment
Previous simulations and test results showed how time delays not exceed-
ing 300 ms are compatible with the stability control of a critical section in
Italy. The feasibility evaluation of the whole proposed control architec-
ture should check if the overall delay due to PMU measurement acquisi-
tion, communication and computational timings do not exceed this
threshold.Even though the computational time might seem the main obstacle, it must
be remembered that, as trajectories are already known by WAMS, corrective
actions are evaluated just through the solution of the linear system (8.34). On a
100 ms time window, this action requires around 400 ms if run on a Compaq
Alpha Server XP1000, 667 MHz, 256 MB RAM.CPU timings are not a real obstacle as they can be made compatible with
response-based control requirements. By adopting more powerful computa-
tional resources (supercomputers, parallel computing environments) it is easily
possible to reduce computational time by a factor 4, or more. As an example, a
speed-up around 9 can be obtained through vectorization and implementing
this problem on a Cray Y-MP8. A speed-up around 6 can be achieved with
message passing machines or distributed machines (e.g., Alpha clusters)
equipped with 32 CPUs.It seems that the bottleneck in the control chain has to be associated with
communication system and actuators’ response. Potentials of WAMS architec-
tures and advanced stability control architectures, with also a special regard on
the performance of the communication system, were investigated during a full-
scale experiment [23, 38, 39]. The experiment program that benefited from the
collaboration of companies involved in PMU manufacturing and WAMS
development dealt with the installation and testing of PMU devices and with
the assessment of SPS performances, including telemetry, monitoring, and
wide-area detection systems. On the basis of data acquired during the above-
mentioned experiment, time performances of the communication
infrastructures have been assessed. The overall time delay for acquiring, trans-
mitting (to the Control Center), and retransmitting (to actuators) data has been
estimated in the range 70–100 ms (25 ms for each one-way data transmission).
This result is also consistent with the estimation of around 100 ms for transfer-
ring data measured by the PMUs to the Control Center in the Terna-WAMS
Italian project [40].Clearly, even though these timings are still on the edge of the feasibility,
technology can provide the right answers for overcoming such limitations. The
implementation of fast actuators such as FACTS devices and high-speed com-
munication infrastructures will certainly allow to meet the strict requirements
imposed by a centralized wide-area response-based control architecture in the
near future.
256 S. Bruno et al.
8.4 Conclusions
The chapter showed how a monitoring architecture based on WAMS technol-
ogy is able to achieve the online supervision of power system dynamic perfor-
mances, identifying poorly damped oscillation modes, and possible threats to
the security of the system. The feasibility of the approach has been proved
through simulations on test grids and on a detailed representation of the Italian
power system.The proposed approach also allowed the recognition of dominant modes of
oscillation. By adopting the complex Morlet WT, it was possible to preserve
damping information in the time–frequency domain. An experimental proof
of this property was obtained through some simple calculations that allowed
damping evaluation on wavelets maps. The feasibility of the approach had
also been proved in the presence of random fluctuations due to operating
conditions and measurement noise. The analysis of both simulated and actu-
ally measured trajectories proved that WT is able to filter out noise. A
suitable MMI for real-time operation and OTS can be addressed by the
proposed approach.An approach for wide-area response-based corrective control of dynamic
security was proposed. On the basis of test results the feasibility of the mathe-
matical approach was showed. Test results led also to the assessment of the
minimum requirements that are requested for this architecture in terms of
overall response time. Basing on performed test results, centralized wide-area
monitoring architectures showed to be quite compatible with response-based
control approach, as the delay associated to data acquisition and control action
implementation is comparable with the maximum acceptable delay.At first glance, it may seem that computation would be the bottleneck.
Nevertheless, because the elapsing time related to dynamic sensitivity calcula-
tions can be drastically reduced through high-performance computing, it is in
our belief that bottlenecks in response time of the control chain are still
associated with communication system and actuators. Fast actuators such as
FACTS devices andmore investments in high-speed communication infrastruc-
tures can provide the right answer to meet the strict requirements imposed by a
centralized response-based control architecture.
Acknowledgments The authors wish to thank the Italian Ministry of the Research forproviding financial support under Interlink Scientific Research Program 2004–2006.
References
1. M. D. Ilic, E. H. Allen, J. W. Chapman, C. A. King, J. H. Lang, E. Litvinov, ‘‘Prevent-ing Future Blackouts by Means of Enhanced Electric Power Systems Control: FromComplexity to Order’’, Proceedings or the IEEE, Vol. 93, No. 11, pp. 1920–1941,November 2005.
8 Advanced Monitoring and Control Approaches 257
2. J. De La Ree, Y. Liu, L.Mili, A. G. Phadke, L. Da Silva, ‘‘Catastrophic Failures in PowerSystems: Causes, Analyses, and Countermeasures’’, Proceedings or the IEEE, Vol. 93,No. 11, pp. 956–964, November 2005.
3. S. H.Horowitz, A.G. Phadke, ‘‘Boosting Immunity to Blackouts’’, IEEEPower &EnergyMagazine, Vol. 1, pp. 47–53, September/October 2003.
4. C. W. Taylor, D. C. Erickson, K. E. Martin, R. E. Wilson, V. Venkatasubramanian,‘‘WACS—Wide-Area Stability and Voltage Control System: R&D and Online Demon-stration’’, Proceedings or the IEEE, Vol. 93, No. 11, pp. 892–906, November 2005.
5. G. Andersson et al., ‘‘Causes of the 2003 Major Grid Blackouts in North America andEurope, and Recommended Means to Improve System Dynamic Performance’’, IEEETransaction on Power Systems, Vol. 20, No. 4, pp. 1922–1928, November 2005.
6. J. S. Thorp, A. G. Phadke, ‘‘Protecting Power Systems in the Post-Restructuring Era’’,IEEE Computer Applications in Power, Vol. 12, No. 1, pp. 33–37, January 1999.
7. C. W. Taylor ‘‘Response-Based, Feedforward Wide-Area Control’’, Position paper forNSF/DOE/EPRI SponsoredWorkshop on Future ResearchDirections for Complex Inter-active Electric Networks, available at: www.transmission.bpa.gov/orgs/opi/Power_Stabil-ity/NSFwideareacontrol.pdf
8. C. W. Taylor, ‘‘The Future in On-Line Security Assessment and Wide-Area StabilityControl’’, Power Engineering Society Winter Meeting, 2000 IEEE, Vol. 1, pp. 78–83,January 23–27, 2000.
9. M. J. Damborg,M.Kim, J. Huang, S. S. Ventaka, A. G. Phadke, ‘‘Adaptive Protection asPreventive and Emergency Control’’, Power Engineering Society Summer Meeting, 2000IEEE, Vol. 2, pp. 1208–1212, July 16–20, 2000.
10. R. Mihalic, P. Zunko, D. Powh ‘‘Improvement of transient stability using Unified PowerFlow Controller’’, IEEE Transaction on Power Delivery, Vol. 11, No. 1, pp. 485–492,January 1996.
11. S. Limyingcharoen, U. D. Annakkage, N. C. Pahalawaththa, ‘‘Effects of unified powerflow controllers on transient stability’’, IEE Proceedings Generation TransmissionDistribution, Vol. 145, No. 2, pp. 182–188, March 1998.
12. R. Messina, V. Vittal, ‘‘Nonlinear, Non-Stationary Analysis of Interarea Oscillations viaHilbert Spectral Analysis’’, IEEE Transaction on Power Systems, Vol. 21, No. 3,pp. 1234–1241, August 2006.
13. S. Huang, C. Hsieh, C. Huang,. ‘‘Application of Morlet wavelets to supervise powersystem disturbances’’, IEEE Trans. on Power Delivery, Vol. 14, No. 1, pp. 235–243,January 1999.
14. K. K.-P. Poon, K.-C. Lee, ‘‘Analysis of transient stability swings in large interconnectedpower systems by Fourier transformation’’, IEEE Transactions on Power Systems, Vol. 3,No. 4, pp. 1573–1581, November 1988.
15. M. Meunier, F. Brouaye, ‘‘Fourier transform, wavelets, Prony analysis: tools forharmonics and quality of power’’, Proceedings of the 8th International Conference onHarmonics and Quality of Power, 1998, Vol. 1, pp. 71–76, October 14–16, 1998.
16. E.Grund, J. J. Paserba, J. F.Hauer, S. L.Nilsson, ‘‘Comparison of Prony and eigenanalysisfor power system control design’’, IEEE Transactions on Power Systems, Vol. 8, No. 3,pp. 964–971, August 1993.
17. J. F. Hauer, C. J. Demeure, L. L. Scharf, ‘‘Initial results in Prony analysis of power systemresponse signals’’, IEEE Transactions on Power Systems, Vol. 5, No. 1, pp. 80–89, February1990.
18. D. J. Trudnowski, J.M. Johnson, J. F.Hauer, ‘‘Making Prony analysismore accurate usingmultiple signals’’, IEEE Transactions on Power Systems, Vol. 14, No. 1, pp. 226–231,February 1999.
19. D. Ruiz-Vega, A. R. Messina, M. Pavella, ‘‘Online assessment and control of transientoscillations damping’’, IEEE Transactions on Power Systems, Vol. 19, No. 2, pp. 1038–1047,May 2004.
258 S. Bruno et al.
20. D. J. Trudnowski, J. E. Dagle, ‘‘Effects of generator and static-load nonlinearities onelectromechanical oscillations’’, IEEE Transactions on Power Systems, Vol. 12, No. 3,pp. 1283–1289, August 1997.
21. M. Bounou, S. Lefebvre, R. P. Malhame, ‘‘A spectral algorithm for extracting powersystemmodes from time recordings’’, IEEE Transactions on Power Systems, Vol. 7, No. 2,pp. 665–683, May 1992.
22. O. Cocozza, M. Dicorato, M. La Scala, R. Sbrizzai, ‘‘A multiresolution analysis tosupervise power system oscillations’’, Proceedings of Conference on Power Engineering,LESCOPE 2000, Halifax (Canada), pp. 141–145, July 23–26, 2000.
23. M. La Scala, M. De Benedictis, S. Bruno, A. Grobovoy, N. Bondareva, N. Borodina,D. Denisova, A. J. Germond, R. Cherkaoui. ‘‘Development of Applications in WAMSand WACS: an international cooperation experience’’, IEEE PES General Meeting,-Montreal, Quebec, Canada, June 18–22, 2006.
24. P. S. Addison, The illustratedwavelet transformhandbook: introductory theory and applica-tions in science, engineering, medicine and finance, Institute of Physics Publishing Ltd, 2002.
25. M. Ruzzene, A. Fasana, L. Garibaldi, B. Piombo, ‘‘Natural frequencies and dampingsidentification using wavelet transform: application to real data’’,Mechanical Systems andSignal Processing, Vol. 11, No. 2, pp. 207–218, 1997.
26. C.-H. Lamarque, S. Pernot, A. Cuer, ‘‘Damping identification in multi-degree-of-free-dom systems via a wavelet-logarithmic decrement. Part 1: theory’’, Journal of Sound andVibration, Vol. 235, pp. 361–374, 2000.
27. L. Joseph, Ta Minh-Nghi, ‘‘A wavelet-based approach for the identification of dampingin non-linear oscillators’’, International Journal of Mechanical Sciences, Vol. 47,pp. 1262–1281, 2005.
28. B. Naduvathuparambil, M. C. Valenti, A. Feliachi, ‘‘Communication delays in wide areameasurement systems’’, Proceedings of the Thirty-Fourth Southeastern Symposium onSystem Theory, pp. 118–122, March 18–19, 2002.
29. Power System Test Case Archive: IEEE 30 Bus ‘‘New England’’ Dynamic Test Case,http://www.ee.washington.edu/research/pstca/
30. J. R. Hockenberry, B. C. Lesieutre, ‘‘Evaluation of Uncertainty in Dynamic Simulationsof Power SystemModels: The Probabilistic Collocation Method’’, IEEE Transactions onPower Systems, Vol. 19, No. 3, pp. 1483–1491, August 2004.
31. C. L. De Marco, A. R. Bergen, ‘‘A Security Measure for Random Load Disturbancesin Nonlinear Power System Models’’, IEEE Transactions on Circuits and Systems,Vol. CAS-34, No. 12, pp. 1546–1557, December 1987.
32. G. Blankenship, G. C. Papanicolaou, ‘‘Stability and control of stochastic systems withwide-band noise disturbances, Part I,’’ SIAM Journal on Applied Mathematics, Vol. 34,No. 3, pp. 437–476, May 1978.
33. J. Depablos, V. Centeno, A. G. Phadke, M. Ingram, ‘‘Comparative Testing of SynchronizedPhasor Measurement Units’’, IEEE Power Engineering Society General Meeting, Denver,June 6–10, 2004.
34. M. La Scala,M. Trovato, C. Antonelli. ‘‘On-line dynamic preventive control: an algorithmfor transient security dispatch’’, IEEE Transactions on Power Systems, Vol. 13, No. 2,pp. 601–610, 1998.
35. E. De Tuglie, M. Dicorato, M. La Scala, P. Scarpellini. ‘‘A corrective control for angleand voltage stability enhancement on the transient time-scale’’, IEEE Transactions onPower Systems, Vol. 15, No. 4, pp. 1345–1353, November 2000.
36. S. Bruno, E. De Tuglie, M. La Scala, ‘‘Transient Security Dispatch for the ConcurrentOptimization of Plural Postulated Contingencies’’, IEEE Transactions on Power System,Vol. 17, No. 3, pp. 707–714, August 2002.
37. S. Bruno, E. De Tuglie, M. La Scala, P. Scarpellini, ‘‘Dynamic Security Corrective Controlby UPFCs’’, IEEE Transactions on Power Systems, Vol. 16, No. 3, pp. 490–497,August 2001.
8 Advanced Monitoring and Control Approaches 259
38. A. Grobovoy, N. Bondareva, N. Borodina, D. Denisova, A. Germond, R. Cherkaoui, M.La Scala, S. Bruno, M. De Benedictis, ‘‘Full-Scale Tests of the AC Transmission Systemto Examine the New SPS Complexes’’, Proceedings of 8th International ConferenceAC DC Power Transmission,London, UK, March 28–30, 2006.
39. A.Grobovoy, N. Lizalek, N. Bondareva, S. Sirazutdinov, V. Stepanov, E. Atienza,M. LaScala, A. Germond, ‘‘Synchronized Measurement Experiment and Trial WAMS/WACSStructure in the Russian Far East Interconnected Power System’’, International ScientificConference onMonitoring of Power System Dynamic Performance, Moscow, April 25–27,2006.
40. A.DeBellis,M.Delfanti, E.DeBernardis,M. Pezzaglia,M. Pozzi,M. La Scala, S.Massucco,P. Gaggero, ‘‘Controllo del sistema elettrico nazionale’’, AEIT, Vol. 94, pp. 58–73, 2007,ISSN: 1825-828X.
260 S. Bruno et al.
Index
A
Activation deactivation criteria, 139Actuators, 8, 232–233, 247–248, 253–257A-EMD algorithm, 78, 82, 91–92A-EMD method, 90, 96Air-gap power, 142–143Alarm rate, 194–195, 201–202Aliasing problem, 109Alpha clusters, 256Ambient noise, 7, 8, 10–11, 14, 20Analysis techniques, 1, 2, 34, 37, 102, 123Analysis window, 15, 17–18, 38, 117, 120,
195–196, 201, 205–206, 208Analytic signal, 40, 79, 107–109, 111,
117–118, 120Angular velocity, 39, 218–221, 223–225Aurelia–Roma line, tripping, 253–254Autocorrelation function, 7, 163, 215Autoregressive moving average (ARMA), 7Autoregressive moving average exogenous
(ARMAX) model, 8
B
Bang–bang compensators, 128Bimodal test signal, 113Black signal, 23Blackout, 127–130, 139, 154, 231Block processing algorithms, 7Blue signal, 23Bootstrap technique, 22, 26BPA Dittmer Control Center, 150
C
Canonical variate algorithm (CVA), 7Carrier signal, 46Cauchy principal value, 68Cellilo substation, 56
Center of inertia (COI), 150, 251Chirp signal, 214–216Choi–Williams Distribution, 224–225COA correction, 207Compensation control actions, 153Completeness, 71Complex Morlet WT, 235, 239–240,
243, 257Continuous wavelet transform (CWT),
235–237Convolution, 65, 79–80, 82, 102, 106–107,
109, 111, 161, 172, 193, 200, 216filter, 79, 172
Corrective actions, event-basedevaluation, 249
Corrective control actions, 250–254real-time, 249response-based, 248, 252, 254
Correlation matrix, 169Covariance matrix, 164–165, 170Creaker, 3, 119Crest factor, 11–12, 20Crosschecking, 132Cross-spectral density (CSD), 27–29, 31–33Curve-fitting technique, 102
See also Prony analysis
D
Damped sinusoids, 5, 10, 106, 112, 114,121, 234
Dampingcharacterization, 68coefficient, 69–70conditions, 16–17estimation algorithms, 5evaluation of, 246factor, identification of the, 234parameter, 105, 121, 190
261
Damping (cont.)ratio, 7, 26, 70–71, 90, 93, 103, 116, 121,
123, 131–132, 137–140, 150–151,153, 236
Daubechies wavelet, 179Decomposing capability test, 80Detection algorithm, 197Detection theory, 227Differential and algebraic equations
(DAEs), 249Discrete Fourier transform (DFT), 27, 65,
104–106, 109, 192, 198, 225, 227limitations of, 65
Discrete wavelet transform (DWT), 235Distortion, 38, 50, 110, 112, 119Double sideband modulated (DSB) signal, 46Dynamic security assessment (DSA), 232, 235Dynamic System Identification (DSI)
Toolbox, 134, 136, 140, 151–152
E
EDM algorithms, 63Eigenanalysis, 14, 29, 53, 64
six-machine, 218Eigenbasis matrix, 53Eigenspectrum, 180Eigenvalue, 5, 26, 54, 90, 103, 132–133,
149, 162, 164–165, 170, 173,211–212, 217, 222–226,233–234, 239–240, 243
analysis, 54, 90, 149, 233–234,239–240, 243
decomposition, 169–170estimate error, 223parameters, 132ralization algorithm (ERA), 5
Eigenvector, 26, 53, 164, 170, 172–173,212–213, 218, 240
16-machine, 29Electromagnetic stator–rotor interactions, 50Electromechanical dynamic effects, 2Electromechanical modes, 2–3, 5, 10, 26, 34,
218, 231, 238identification, 10
Empirical mode decomposition (EMD),37–40, 63–64, 66, 72, 75, 102
algorithm, 38–39, 77–78, 87, 89, 92, 98,115, 124
technique, 39, 59, 63–65, 95–97, 98Empirical orthogonal function, 159–187
complex analysis, 171discrete domain representation, 163
energy relationships, 166snapshots method, 164theoretical development, 160
Empirical transfer function estimation,10, 52
EOF, see Empirical orthogonal functionEstimating mode shape, 26–28
defining, 26–27estimating, 27–28
Extended Transient Midterm StabilityProgram (ETMSP), 134, 136
F
FACTS, see Flexible AC transmissionsystem
Fast Fourier transform (FFT), 38,43–45, 48–50, 52, 58–59, 75, 78,104–106, 109, 114, 225
Feedback control signals, 233Field current variations, 50–52Filter banks, 64, 79–80Filtering process, 77Flexible AC transmission system, 128,
232–233, 248, 253, 256–257Fourier-based techniques, 67, 128Fourier series, 103, 105Fourier spectral analysis, 64Fourier spectrum, 64, 75, 77–78Fourier technology, 103, 105Fourier transform, 38, 40–41, 66, 79, 102,
128, 180, 214, 233–234Fractional harmonics, 104Frequency deviation recorder (FDR),
55–58Frequency-domain
analysis, 12, 38calculation, 7decomposition (FDD), 8identification, 7
Frequency error, 32Frequency heterodyne technique, 45–48Frequency shifting, 45, 47
G
Garigliano bus, 244–248Gateaux derivative, 161Gaussian load variation, 14Gaussian white noise, 3, 13, 179, 210, 244Global Positioning Systems (GPS), 55,
232, 238Grid capacity, 1, 232
262 Index
H
Hankel total least squares (HTLS), 5, 130Harmonic series, 105Harmonic signal, 236Hermitian matrix, 173Hermosillo signal, 181Heterodyne detection, 46Heterodyne frequency, 46, 48, 59Heterodyne technique, see Frequency
heterodyne techniqueHigh-pass filter, 180High-temperature superconducting (HTS)
motors, 50High-voltage direct current technology, 56,
58, 232line tripping, 58
Hilbert analysis, 74, 78, 90, 97, 101, 103,115–117, 119–121, 123–124
Hilbert transform (HT), 38–42, 46, 49,54, 58–59, 63–64, 67–68, 79–80,82, 84, 98, 101–126, 170, 172,216, 219
Hilbert transform and analytic function, 106implementation, 108instantaneous frequency, 109modal parameters, 107properties, 106
Hilbert transformer, 46, 82Hilbert–Huang technique, 38–39, 48, 51–52,
59–60Hilbert–Huang transform (HHT) 37, 39–41,
48, 50, 52, 59–60, 63–65, 67–69,71–73, 75, 77, 79, 81, 83–87, 89, 93,95, 97–99
empirical mode decomposition, 39Hilbert transform, 40
HTS propulsion motors, 50–52HVDC, see High-voltage direct current
technology
I
Identification techniques, 102–103Fourier methods, 103prony analysis, 102
IF, see Instantaneous frequencyIMF, see Intrinsic mode functionIndex of orthogonality, 72Individual mode test statistic, 198–199Instantaneous damping, 67–71, 83, 93, 109
computaiton, 84–87Instantaneous frequency, 38–41, 49, 52, 54,
59–60, 68, 70, 77–78, 80, 85, 90,
92–93, 96, 109, 130, 182, 214, 216,222, 224
Instantaneous phase, 40–41, 52, 54, 58,182, 214
Inter-area modes, 3, 8, 10–12, 14, 31, 64, 87,90, 93, 109, 128, 130–132, 135, 142,149, 179, 204, 218–219, 220, 224
Inter-area oscillations, 38, 101, 121, 131,231–232, 238, 248
Intrinsic mode function, 37, 39–40, 42–49,51–56, 58–60, 64–68, 71–74, 76–78,80–83, 86–92, 94–98, 112–115,117–118, 120–121, 123
J
Jacobian matrix, 144, 146Jitters, 60
K
Kalman approach, summary of, 211Kalman estimator, 204–205Kalman filter, 190, 197, 205Karhunen–Loeve decomposition, 166Keeler–Allston line, 139–140, 150–151
L
Lagrangian multipliers, 250–251Laguna Verde (LGV) unit, 48, 87–88Lead–lag compensators, 128Leakage, 12, 79, 105Least-mean squares (LMS) method, 8Line switchings, 227Linear analysis method, 112Linear chirp signal, 214–216Linear decomposition, 72, 167Linear differential equation, 236Linear estimation techniques, 211, 227Linear modeling technique, 222Linear system theory, 103Load modeling, 13Load torque variation, 50, 52Load tripping schemes, 2, 26Long-term estimator (LTE), 204–205
M
17-machine system, 14, 17–18Man–machine interface (MMI), 235, 238, 257Maple Valley, 151, 153Masking frequencies, 50–51
Index 263
Masking signal, 44–46, 48–51, 53–54,58–59, 63, 65, 67, 75–78, 93,98, 114
modified, 92Masking technique, 65, 72–80
energy-based, 78Fourier-based, 75–77local Hilbert transform, 79–80standard EMD method, 72–75
MASS software, 149Matlab, 80, 147, 149, 217Matrix pencil method, 5, 130, 154Measurement noise, 3, 16–18, 246–248
influence of, 246–247Measurement-based modal analysis, 5Melbourne-COA, 204Mexican Interconnected System (MIS), 48,
102, 120, 176–177MIMO, see Multiple-input multiple-outputModal damping change detection, 190–211
application to real data, 204–211energy detection, 191–197Kalman approach, 197–203
Modal energy methods, 8Modal estimate algorithms, 201Modal identification, 101, 107–108, 119,
190, 226Modal matrix, 165Modal parameter estimation, 119, 201Mode-damping ratio, 26Mode estimation
algorithms, 2, 7, 17, 19analysis, 13examples, 12–24
field measured data, 19mode-meter performance, 15–19probing test, 20–24ringdown analysis, 14–15simulation system, 13–14
Mode meters, 5Mode observability matrix, 242Model validation and performance
assessment, 24–26model validation, 24–25performance assessment, 25–26
Mode-meter algorithm, 1–2, 8, 15, 17–18Modified Hilbert–Huang technique, 50,
59–60Modified Hilbert–Huang transform, 41–58
frequency heterodyne technique, 45–48limitations of EMD, 42masking signal-based EMD, 44power flow oscillations, 48–49
slow coherency analysis, 53–55torque and field current variations,
50–52wide-area measurement signals, 55–58
Monitoring architecture, 238, 257Monofrequency, 64, 68, 70–71, 74, 84–85, 88
signals, 68, 84Monte Carlo simulations, 16, 18Morlet wavelet, 235–236, 240, 243, 257Multi-Area Small-Signal Stability (MASS)
Program, 136–137, 149Multiple mode signals, 112Multiple-input multiple-output, 3, 64, 128Multi-Prony analysis, 130–135, 137, 150Multi-Prony monitoring, 139Multisine probing signals, 10
N
N4SID algorithm, 7, 8, 10, 17–18National Transmission Grid, 254Natural frequency, 54, 70, 103Natural pulsation, 236Network topology, 3Noise-contaminated measurement
point, 179Noise-corrupted test signal, 116Noise filtering, 119Noise signal, 12Noise tolerance, 116–119Nonlinear programming technique, 249Nonlinear state–space model, 144Nonlinear/Nonstationary signals, 80,
83–84, 86Numerical algorithm for subspace
state–space system identification,see N4SID algorithm
Numerical artifacts, 8
O
Off-line rules, 130, 133, 139Online modal identification technique, 123Online real-time software tools, 2Operator training simulator (OTS),
238, 257Optimization problem, 161, 231, 249–250,
252, 254Orthogonality index, 71–72, 98Orthogonality property, 71Orthonormal matrix, 168–169Oscillation frequencies, 52Oscillation monitoring system (OMS), 130
264 Index
P
Pacific Daylight Time (PDT), 56Pacific DC Intertie (PDCI), 8, 10–11, 19, 23,
25, 56Parameter identification techniques, 106Participation
factors, 55, 212, 240–241matrix, 240
PDC block, 238Performance assessment, 24–26Periodogram averaging methods, 28Phasor data, 32, 238Phasor measurement unit (PMU), 28, 32,
131, 176–177, 180, 232, 235,238–239, 246, 248, 256
Post disturbance setting, 2Power management units, 28Power spectral density, 7, 27–30, 32, 197,
204, 206–210Power swing damping control (PSDC), 130Power System Dynamic Performance
Committee, 232Power system electromechanical modes, 5Power system identification, 8–12
probing signal selection, 10–12Power system measurements, real-time
analysis, 101Power system oscillations, 233–247
approaches for monitoring, 233Morlet-based wavelet analysis, 235WAMS-based monitoring
architecture, 238Power system security, 231Power system stabilizer (PSS), 13, 48, 128, 211Priolo bus, 246Probability density function (PDF), 191Processing gain, 12Prony analysis, 5–6, 14, 19, 64, 83–84, 93, 97,
101–103, 108, 112, 116–117, 119,121, 123–124, 128, 130–133, 135,137, 222, 234
application of, 115matrix extension of, 133
Pronymethod, 102, 124, 127, 129–130, 133, 234application of, 221
Propagating features, 170–175Proper orthogonal decomposition (POD),
160, 164, 169, 178, 181Proper orthogonal vector (POV), 164, 170Propulsion motor, 48, 50–52PSD, see Power spectral densityPSDC compensation, 149, 152PSDC mode, 131, 143, 146–147, 149
Pseudo-DC component, 123Pseudoenergy, 17–18Pseudorandom
noise, 10probing, 20, 23signal, 11
Q
Quality instrumentation, 3Quantification, 253Quasi-stationary operation, 204
R
R3LS, see Recursive least-squaresRandom contamination, 178Random fluctuations, 243, 257Random variable (RV), 192Real-time analysis of power system
measurements, 101Real-time control actions, 249Real-time corrective control actions, 249Real-time sensors, 232Realtime-synchronized measurement
system, 2Realization algorithm, 5Real-zero (RZ) signal, 42Recursive least-squares, 8, 10Reference signal, 30–31, 33, 197–198, 255R-EMD (Refined EMD), 75, 95, 98, 87
algorithm, 77–78, 80, 82, 88–89, 95Remedial action schemes (RAS), 233Response-based control, 233, 248, 252, 254,
256–257Response-based corrective control actions,
248, 252, 254Response-based wide-area, 247–256
computational and communication timedelay assessment, 256
mathematical formulation, 249–253Response matrix, 167, 169–170Ringdown, 3–6, 12, 14, 18–19, 21, 38, 218–219
analysis, 5, 12, 14, 83, 134analyzers, 5
Ripple effect, 79RLS algorithm, 19Root mean square (RMS), 11Ross–Lexington line, 140, 152Rotor angle trajectories, 240, 242RRLS algorithm, 19, 22Russian Far East Interconnected Power
Systems, 235
Index 265
S
Sensor-induced jitters, see JittersSequential tripping, 152
See also TrippingShort-time Fourier transform, 106, 214–215,
233–234, 234Sifting, 60, 64–66, 76–77Signal processing algorithms, 2Signal processing methods for estimating
modes, 5–8mode-meter algorithms, 7–8ringdown algorithms, 5–6
Signal processing techniques, 1, 211Signal-to-noise ratio (SNR), 11–12, 14, 16,
25, 117, 129, 133, 207–211Simulation test system, 13Single-mode signal, 105, 107, 116, 119
application to, 109Single-mode systems, 211Single sideband modulated (SSB), 46Singular value decomposition, 151, 167,
169–170, 174Sinusoidal signal, 75, 85Sinusoidal wave, 175Six-machine system, 53, 21716-machine system, 14–16, 18–19, 21, 29Slow coherency, 48, 53–55
analyzing, 53–55Small-signal stability, 127–130, 233Snapshots, 162, 164, 170, 178
method, 159, 162, 164–166, 170,178–179
Spatial aliasing, 178Spatial amplitude function, 174–175Spatial phase function, 174, 182Spatiotemporal analysis, 179, 186Spatiotemporal model, 178Special protection scheme (SPS), 60,
248, 256Spectral analysis, 26–27, 29, 31–32, 38, 48,
64, 175, 179, 231, 233–235,239–242, 244–246
Spectral correlation analysis, 184Spectral estimation techniques, 225Spline fitting, 59Stabilization controls, 211Standard linear algebra technique, 164State–space model, 10, 144, 204Static VAR compensator (SVC), 128–132,
134, 136, 141–154, 170, 189Statistical correlation theory, 160STFT, see Short-time Fourier transformStochastic additive disturbance, 244
SVC local unit, 141–150classical power system model, 142–143linearized state–space classical model,
143–146SVC rules, 149–150
SVD, see Singular value decompositionSwing curves, 53–55Swing equation, 141–142Sylmar substation, 56Synchronized phasor measurements,
128, 159System identification
algorithms, 11theory, 11
System noise, 25, 243–246
T
Temporal amplitude function, 174–175Temporal behavior, 82, 98–99, 160, 166, 178,
185–186Temporal coefficients, 166Temporal phase function, 174–175, 182Testing mode estimation algorithms, 19Theory of optimal detection of random
signals, 197Three-phase fault, 218, 253Tie-line power flow signal, 115, 121–123Time-dependent decay function, 69Time-domain
algorithms, 7response, 4, 212signals, 234
Time–frequencydistribution (TFD), 214, 215, 219, 227domain, 235, 239, 257representation (TFR), 213–214, 216, 224resolution, 227
Time-synchronized data, 34, 159Time-synchronized information, 178Time-synchronized measured data,
176–186construction of POD modes, 178energy distribution, 185frequency determination, 182mode shape estimation, 184spatiotemporal analysis, 179temporal properties, 182
Time-synchronized measurements, 1, 26, 28,176, 186
phasor measurements, 93Time-varying effects, 227Torque, 48, 50–52
266 Index
Transfer functionequation, 10estimation, 10, 50, 52, 60
Transmission system nonlinearity, 217Tripping, 37, 48–49, 53–54, 56, 244–246, 253
U
Unified power flow controller (UPFC), 233,253, 255
V
Validating criteria, 140–141Vectorization, 256Verification procedure, 193Vibration analysis, free, 70Vibratory system solution, 70Voltage
angle, 17, 136, 204fluctuation, 244magnitude, 134, 137, 144, 241, 246trajectory, 244–245transient, 251vectors, 250
W
WAMS, see Wide-area measurementsystems
Wavelet analysis, 233–235Wavelet shrinkage model, 179
Wavelet transform (WT) algorithms, 234Welch periodogram, 7, 25Western North American power system,
2, 4, 8, 10, 12, 19–20, 23–24, 25,31–33
Western Systems Coordinating Council(WSCC), 130, 138–139, 150–151
White Gaussian noise, 116Whitening filter, 190, 197–199Wide-area frequency information, 186Wide-area measurement and control systems
(WAMC), 232Wide-area measurement signals, 48, 55–58Wide-area measurement technology, 234, 257Wide-area measurement systems, 37, 55–58,
232–234, 238, 247–248, 252,256–257
Wigner–Ville distribution (WVD), 215–217,219–222, 224–227
Windowed Prony, 224, 227Windowing technique, 234wNAPS, see Western North American
power system
Y
Yule–Walker (YW) algorithm, 7–8
Z
Zero padding, 112, 124Z-transform identification algorithm, 234
Index 267
Power Electronics and Power Systems
Series Editors:Series Editors: M. A. PaiUniversity of Illinois atUrbana-Champaign Urbana, Illinois
Alex StankovicNortheastern UniversityBoston, Massachusetts
Continued from page ii
Operation of Restructured Power SystemsKankar Bhattacharya, Math H.J. Bollen, Jaap E. DaalderISBN 978-0-7923-7397-1
Transient Stability of Power Systems: A Unified Approach to Assessment and ControlMania Pavella, Damien Ernst, and Daniel Ruiz-VegaISBN 978-0-7923-7963-8
Maintenance Scheduling in Restructured Power SystemsM. Shahidehpour and M. MarwaliISBN 978-0-7923-7872-3
Power System OscillationsGraham RogersISBN 978-0-7923-7712-2
State Estimation in Electric Power Systems: A Generalized ApproachA. MonticelliISBN 978-0-7923-8519-6
Computational Auction Mechanisms for Restructured Power Industry OperationsGerald B. ShebleISBN 978-0-7923-8475-5
Analysis of Subsynchronous Resonance in Power SystemsK.R. PadiyarISBN 978-0-7923-8319-2
Power Systems Restructuring: Engineering and EconomicsMarija Ilic, Francisco Galiana, and Lester Fink, eds.ISBN 978-0-7923-8163-1
Cryogenic Operation of Silicon Power DevicesRanbir Singh and B. Jayant BaligaISBN 978-0-7923-8157-0
Voltage Stability of Electric Power SystemsThierry Van Cutsem and Costas VournasISBN 978-0-7923-8139-6
Automatic Learning Techniques in Power SystemsLouis A. WehenkelISBN 978-0-7923-8068-9
Energy Function Analysis for Power System StabilityM. A. PaiISBN 978-0-7923-9035-0
Electromagnetic Modelling of Power Electronic ConvertersJ. A. FerreiraISBN 978-0-7923-9034-3
Spot Pricing of ElectricityF. C. Schweppe, M. C. Caramanis, R. D. Tabors, R. E. BohnISBN 978-0-89838-260-0