1
Situational Awareness:Singular Value Methods for
PMU Data Interpretation
Professor Chris DeMarcoDepartment of Electrical & Computer Engineering
University of [email protected]
PSERC Tele-Seminar SeriesMarch 2, 2010
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 2
Contributors & Acknowledgements
UW-Madison graduate student researcher onthis work: Mr. Rafael Viloria, supported in partby PSERC project S-36, “Use of PMU Data toIncrease Situational Awareness,”
Research has also been supported by theBonneville Power Administration under contractnumber 00037890, “Voltage Stability Controls.”Project lead Dmitry Kosterev; technicalcollaborator at BPA Eric Heredia.
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 3
Problem Motivation & Background
Synchronized “Phasor Measurement Unit”(PMU) familiar to most power engineers.Briefly, a PMU measures a windowed Fouriertransform (“phasor”) of the nominally sinusoidalvoltages, currents, powers throughout grid.
GPS technology has facilitated precise, low-costtime synchronization of signals across largegeographic distances. Today one can collectprecisely synchronized measurements, at 30 or60 Hz sampling rate, on a continental scale.
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 4
Problem Motivation & Background
So… utilities have (or will soon have) hugevolumes of real-time & historic PMU data.
Refrain in our suddenly “Smart” Grid:“How do we extract ‘knowledge’ from data?”
Personally, I prefer less grandiose formulation:How do we compress PMU data, and use itto compute real-time performance metricsthat improve grid control action?
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 5
Problem Motivation & Background
Need for reduction,feature extraction fromvoluminous data hardly unique to power industry.
Geological data processing, gene sequencing,bio-informatics, electronic commerce customerclassification – all problems with similar features.
Long history of successful methods to achievereduction and feature identification in huge datasets, particularly among other branches of theenergy industry (i.e, oil companies).
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 6
Problem Motivation & Background:Late Breaking News
“ Sent: Friday, February 26, 2010 7:29 PMTo: NASPI Work Group MembersSubject: NERC-NASPI SynchroPhasor Data-Sharing Agreements Available
Dear NASPI Colleague --
I am pleased to announce that the new NERC-NASPI SynchroPhasor Data-Sharing Agreements have been completed and are available for yourorganization's review and signature. These agreements have beendeveloped to cover the sharing of real-time synchrophasor data among dataproducers, data users and researchers. The lack of such agreements todate has been a major obstacle inhibiting the collection and sharing ofphasor data for wide-area monitoring and situational awareness to improvebulk power system reliability.
….”
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 7
If it’s Good Enough for Netflix(…it should be good enough for critical national infrastructure)
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 8
Data Algorithms & Prior PMU Work
As suggested in New York Times Netflix article,common “Swiss Army Knife” for treating largedata sets in many fields has been SingularValue Decomposition (SVD) and relatedmethods such as Principal Component Analysis(PCA).
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 9
Prior SVD and PCA Workin Power Systems
SVD methods used for noise reduction in PMUdata, as a pre-filtering step for estimatingoscillatory modes in grid electro-mechanicaldynamics (M. Venkatasubramanian, PSercproject S-29).
PCA methods and its nonlinear variantsemployed for demand prediction in LMP marketrisk management (S. Deng, Pserc project M-17).
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 10
Synopsis of New Contribution
Premise in today’s work: In characterizing quasi-steadystate grid performance, SVD analysis particularly wellsuited to handling large PMU data sets.
Claim: A windowed SVD computation on PMU datatracks a well-established performance metric, whosecomputation traditionally would require state estimationof operating point, with full network and load models.
Method here offers a “model free,” real-time indicator ofquasi-steady state grid performance, particular relevantfor control schemes to guard against voltage instability.
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 11
Singular Value Decomposition:General Background
Refresher on singular value decomposition
J an l xn matrix, rank m, is decomposed as
J = Udiag{σ1, σ2, … σm , 0,…0}V*
where, U and V are unitary matrices (i.e., UU* =Identity Matrix), and σ’s are positive realnumbers (case above illustrates both l, n > m)
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 12
Singular Value Decomposition:General Background
Columns of U serve as orthonormal basisvectors for range space of J;
Columns of V serve as orthonormal basisvectors for domain space of J;
In applications where J plays role of linearoperator from input space to output space (e.g.,multi-input, multi-output transfer function), σ’sdescribe gains along different directions.
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 13
Singular Value Decomposition:General Background
Unit sphere ininput space:
Columns of Vas basis vectors
Ellipse in outputspace, lengths ofaxes set by σ’s.
Columns of U asaxes/basis vectors.
Operate withmatrix J, maps
to…
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 14
Singular Value Decomposition inData Compression & Data Mining
• Long-standing use of SVD in data handling.
• Basic idea: consider sequential acquisition(1,2, …k…) of a vector of “l” measurements m[k]
• For running window of length n, construct
M[k] := [m[k-n+1], m[k-n+2], … m[k] ] (so M[k] is dimension l rows, n columns)
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 15
Simplest InterpretationApplying SVD to PMU Data
• View PMU data as a time series “output,” withvector of PMU measurements at time sample kcomprising m[k]. A window of such vectorscomposes matrix M[k].
• VERY simple idea: watch for degradation inoperating condition to show up as changes inσ’s and U. Initial efforts seek relation to anestablished performance metric, hence focus onthe largest singular value and associated vector(but certainly could look at additional σ’s )
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 16
Interpreting PMU Datain SVD Perspective
Consider Quasi-steady-state, input-output view:• In power system, inputs are the continuously
varying P-Q injections. Outputs are δ’s and V’sof PMU data. Mapping between theminfluenced by network switching, componentfailure, other structural changes.
• Injections have slowly varying component (dailyload curve), and small magnitude, faster randomvariation. Random part looks like small variancefiltered white noise(~ 1% demand magnitude).
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 17
Next Level of SophisticationInterpreting SVD for PMU Data
Fast time-scaleRandom
Variations inP-Q Injections
as “Input”
Measured δ’s andV’s of PMU data
as “Output”
Physical PowerSystem provides the
mapping:approximate as
PF Jacobian Inversein Taylor Expansion
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 18
Next Level of SophisticationInterpreting SVD for PMU Data
Variations inP-Q Injections
as “Input”
Measured δ’s andV’s of PMU data
as “Output”
Lightly StressedSystem: modest
sensitivity of δ’s andV’s to injections.
Largest σ will havemoderate magnitude
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 19
Next Level of SophisticationInterpreting SVD for PMU Data
Highly StressedSystem: high
sensitivity of δ’s andV’s to injections.
Largest σ will havevery large magnitude
Variations inP-Q Injections
as “Input”
Measured δ’s ,V’s,flows of PMU data
as “Output”
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 20
Relation to Traditional VoltageInstability Metric
• Voltage instability problem inspired voluminousliterature and range of methods in 80’s and 90’s.
• Details vary, but most shared basic viewpoint –quasi-steady state input changes drive variationin operating point. “Degree of stability”degrades (perhaps as measured througheigenstructure of linearized dynamics), untilstability is lost. “Path” of state divergence afterstability lost often manifested predominantly asvoltage decline - hence VOLTAGE instability.
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 21
Relation to Traditional VoltageInstability Metric
• In mathematical terms, scenario describedabove is a bifurcation.
• Wide range of bifurcation phenomena possible.Simplest appearing to “fit” observed voltageinstability phenomena is saddle node bifurcation– quasi-static motion of operating point forlinearized dynamics looses stability viaeigenvalue passing through zero.
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 22
Relation to Traditional VoltageInstability Metric
Power system linearized dynamics depend on operatingpoint through power flow Jacobian. Hence, subject to allsimplifying assumptions above, performance metricindicating proximity to voltage instability emerges:
Track smallest singular value of power flowJacobian matrix as operating point varies.
Among early works making this observation –A. Tiranuchit, R.J. Thomas, “A Posturing Strategyagainst Voltage Instabilities in Electric Power Systems,”IEEE Trans. Power Sys, 1988.
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 23
Relation to Traditional VoltageInstability Metric
Perhaps obvious, but useful to note:
1/ [Smallest singular value of power flow Jacobian]
= Largest singular value power flow Jacobian inverse
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 24
Recall Conceptual Picture Earlier –PMU Role of Power Flow Jacobian
Fast time-scaleRandom
Variations inP-Q Injections
as “Input”
Measured δ’s andV’s of PMU data
as “Output”
Physical Power System providesthe mapping: approximate as
PF Jacobian Inverse
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 25
Caveats and Practical Issues
• PMU deployment expanding, but still expectonly a modest subset of all phasor angles andvoltages (i.e., PMU’s measurement densitymodest % of all bulk power system buses)
• So mapping we actually get is only a subset ofthe rows of the Power Flow Jacobian Inverse.
(Aside: this framework may also offer a verytractable formulation for optimizingmeasurement placement/observability).
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 26
Computational Experiments inSynthetically Generated Data
First: IEEE 14 and 118 bus test systems, in aMATLAB power flow:
• Construct sequential power flow computation;• “Drive” computation by time sampled loads &
generation dispatch, along 24 demand curve, with1% random load variation superimposed(computation to follow uses 15 sec samplinginterval, 5760 samples per 24 hours);
• “Stress” system by randomly chosen switching inand out of lines over 24 study period;
• For (subset of buses) record angles and voltagemagnitudes as hypothetical PMU measurements;
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 27
Computational Experiments inSynthetically Generated Data
For IEEE 14 and 118 bus scenarios described on previouspage, we’ll test our hypothesis by comparing plots of:
• Largest singular value of windowed PMU measurementmatrix (labeled as “Sub-window SingVal” in plots to follow);
• Largest singular value of computed from appropriate rowsof PF Jacobian inverse (labeled as “SingVal inv-Jacobian”in plots to follow)
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 28
IEEE 14 Bus Experiment – IdealizedLimit of PMU at Every Bus
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 29
IEEE 118 Bus Experiment – PMUPenetration 11 out of 118 Buses
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 30
IEEE 118 Bus Experiment – PMUPenetration 11 out of 118 Buses
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 31
Power Flow based ComputationalExperiments for Larger Systems
• Next goal in computational experiments -generate time-sequenced power flows in morerealistic model, with increased system stressand corresponding control actions (e.g., capswitching) for known scenario.
• Efforts under way to modify power flow tools tomore easily accomplish time sequenced studies,with very large number of samples over periodof interest (perhaps down to 2 seconds, so43,200 power flow solutions for 24 hr study).
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 32
Power Flow Based ComputationalExperiment in Larger System
• Bonneville Power Administration collaborator, Mr. EricHeredia, (painstakingly!) constructed 350-sample study(as we await improved time-sequenced solution tools).
• Study system is synthetic WECC case, graduallystressed with increasing north-to-south transfers.
• All tap changers and cap-switching blocked, until a mid-study correction point, at which a number of voltagecontrol actions applied.
• “Stress metric” (i.e., first singular value vs. time) should:gradually increase, drop at midpoint, increase to near-loss of PF solution, & “drop to solvability” at end…
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 33
Power Flow Based ComputationalExperiment in Larger System
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 34
Real-World: SVD tests on historicPMU data sets courtesy of BPA
• Plot to follow shows six eight-hour “days,” withcolor coded largest singular value versus time.
• Vertical axis: Largest singular value computedon windowed measurement set (52 PMUmeasurement channels used, sampled@ 30 Hz, window length ~150 samples).
• Horizontal axis: Time in hours, starting from11:00 AM as “zero hour”
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 35
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 36
Information from Singular Vector:Which Measurements Influential?
• As part of singular value decomposition, one alsoobtains singular vectors: in notation of earlier slides,the columns of the U matrix.
• By definition, each column has unit length (2-normequal to 1). Size of entries of U1 show degree ofcontribution to largest singular value.
• Rough first test: for the 52 measurements in dataset here, flagged a measurement as “influential” ifmagnitude exceeded 1.5/sqrt(52). Result seems tocorrelate well with engineering experience as towhich PMUs are “important” measurement points.
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 37
Interpretation & Questions Raised
• Data scaling: when mixing PMU measurements ofphysically different quantities, their relative scalingstrongly influences their contribution to singular values.
• For example, natural to keep voltages in pu, while forMW line flows standard base of 1000 MVA seemsnatural (but clearly arbitrary!);
• But then angles: Degrees? (no – too large numerically –angles unrealistically dominate significance!);Radians? (better, but initial experience suggests may bea little “small” – less significance than seems realistic);Historic baseline average? (appealing, but subjective)
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 38
Next Steps
Data Scaling Issues:• Model based approach: run SVD analysis on synthetic
PMU data generated from power flow studies. CorrelateSVD scaling and thresholds with trusted analyticindicators available in full power flow (e.g. conditioningand sensitivities from power flow Jacobian)
• Statistical learning approach: use large number ofhistoric data sets, with ranking of degree of systemstress during these periods, to learn statistics of svdbehavior, and thresholds of transitions to unacceptableoperating conditions.
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 39
Next Steps
Data Drop-out Filtering Issues:• Inevitable that a geographically distributed
measurement system (PMUs) will be subject tocommunication loss/data dropouts, other bad data.Seek optimal estimation/filtering to reduce these effects.
• Yet unusual line switching and other transient eventsproduce points that are significant, should be reflectedin performance metric, and perhaps influence control.
• Adds challenge to filtering problem - avoid throwingbaby out with the bath water.
PSERC Seminar, C.L. DeMarco, [email protected]; 3/2/2010. Full slides & audio at www.pserc.org 40
Next Steps
Data Reduction Issues• Common historic use of SVD is as data compression
tool - ignore data associated with singular values belowsome tolerance (aside: this technique historically wascompetitor to jpeg in image data compression).
• Our very crude first tests, with high standards of fidelity,and no filtering to remove outliers, suggest fairly modestcompression gains with real-world PMU data sets.
• But much more could be done in data compression.To best knowledge of this researcher, appears a missedopportunity in design of PMU data concentrators.