Online Feedback Optimizationwith Applications to Power SystemsFlorian DörflerETH ZürichEuropean Control Conference 2020
Acknowledgements
Adrian Hauswirth Saverio BolognaniLukas
Ortmann
CURRICULUM VITAE
Name: Irina Subotić Birth: 22.03.1993, Belgrade, Serbia e-mail: [email protected]
EDUCATION
WORK EXPERIENCE summer 2015 summer 2011, 2013
“Life Activities Advancement Center” • Speech Recognition project
o Software IKARlab o Forensic analysis of speech, speech recognition
• E-speech Therapist o Communication between a patient and a therapist using AdobeMediaInteractivServer o Design of a data base (MySQL DataBase) for patients’ electronic cards
2013-2016 Undergraduate teaching student assistant at Faculty of Electrical Engineering
• Checking preparedness of students for laboratory exercise • Explaining apparatus and demonstrational experiments to students
2011-2016 Tutoring
• Tutored high school pupils/students mathematics, physics • Preparation for faculty admission exams
AWARDS AND SCHOLARSHIPS
Sep 2016- present
Swiss Federal Institute of Technology in Zurich • Master of Science in Robotics, Systems and Control • GPA: - out of 6.0
2012- 2016 School of Electrical Engineering, University of Belgrade, Serbia • Bachelor of Science in Electrical Engineering and Computing Degree • Signals and Systems department • Focused on Systems Control and Signal Processing • GPA: 9.39 out of 10.0
2008-2012 Mathematical Grammar School, Belgrade
• Secondary School Leaving Diploma • Completed experimental final two years of Elementary School at Mathematical Grammar
School, Belgrade, GPA: 5.00 out of 5.00 • 4 years of high school, focused on Mathematic, Physics and Computer Science, GPA 5.00 out
of 5.00 • Honors:
o “Vuk Karadžić diploma“ for overall success in education
• The “Dositeja” reward for studying abroad in the year of 2016 • Scholar of the City of Belgrade in the years of 2012-2016, 2009 and 2010 • The “Dositeja” reward for the extraordinary success in the year of 2009 and 2010 • Second award on the National Physics competition, in Subotica 2010 • Third award on the National Physics competition, in Belgrade 2009
Irina Subotić Gabriela Hug Miguel Picallo Verena Häberle
1 / 31
feedforwardoptimization
Optimization System
d estimate
u
w
y
complex specifications & decisionoptimal, constrained, & multivariablestrong requirementsprecise model, full state, disturbanceestimate, & computationally intensive
vs. feedbackcontrol
Controller Systemr +u
y
w
−
simple feedback policiessuboptimal, unconstrained, & SISOforgiving nature of feedbackmeasurement driven, robust touncertainty, fast & agile response
→ typically complementary methods are combined via time-scale separation
Optimization Controller Systemr +
u
y
−
offline & feedforward∣∣∣ real-time & feedback
2 / 31
Example: power system balancingoffline optimization: dispatch basedon forecasts of loads & renewables
0
50
100
150
200
0 10 20 30 40 50 60 70 80 90 100
mar
gina
l cos
ts in
€/M
Wh
Capacity in GW
RenewablesNuclear energyLigniteHard coalNatural gasFuel oil
online control based on frequencyFrequency
ControlPower
System50Hz +
u
y
frequency measurement
−
re-schedule set-point to mitigate severeforecasting errors (redispatch, reserve, etc.)
more uncertainty & fluctuations→ infeasible& inefficient to separate optimization & control
50 Hz
51 49
generation
load
control [Milano, 2018]
Re-scheduling costs Germany [mio. €]
!!"#$%
!&' !&(
%%(# %%"&
%!()%!*!
!"## !"#! !"#$ !"#% !"#& !"#' !"#( !"#)
[Bundesnetzagentur, Monitoringbericht 2011-2019]
3 / 31
Synopsis & proposal for control architecturepower grid: separate decision layers hit limits under increasing uncertaintysimilar observations in other large-scale & uncertain control systems:process control systems & queuing/routing/infrastructure networks
proposal: open︸ ︷︷ ︸with inputs & outputs
and online︸ ︷︷ ︸iterative & non-batch
optimization algorithm as feedback︸ ︷︷ ︸real-time interconnected
control
optimizationalgorithm
e.g.,
u = −∇φ(y, u)
dynamicalsystem
x = f (x, u, w)y = h(x, u, w)
actuationu
measurementy
operationalconstraints
u ∈ U
disturbance w
4 / 31
Historical roots & conceptually related workprocess control: reducing the effect of uncertainty in sucessive optimizationOptimizing Control [Garcia & Morari, 1981/84], Self-Optimizing Control [Skogestad, 2000], ModifierAdaptation [Marchetti et. al, 2009], Real-Time Optimization [Bonvin, ed., 2017], . . .
extremum-seeking: derivative-free but hard for high dimensions & constraints[Leblanc, 1922], . . . [Wittenmark & Urquhart, 1995], . . . [Krstić & Wang, 2000], . . . , [Feiling et al., 2018]
MPC with anytime guarantees (though for dynamic optimization): real-time MPC[Zeilinger et al. 2009], real-time iteration [Diel et al. 2005], [Feller & Ebenbauer 2017], etc.
optimal routing, queuing, & congestion control in communication networks:e.g., TCP/IP [Kelly et al., 1998/2001], [Low, Paganini, & Doyle 2002], [Srikant 2012], [Low 2017], . . .
optimization algorithms as dynamic systems: much early work [Arrow et al., 1958],[Brockett, 1991], [Bloch et al., 1992], [Helmke & Moore, 1994], . . . & recent revival [Holding & Lestas,2014], [Cherukuri et al., 2017], [Lessard et al., 2016], [Wilson et al., 2016], [Wibisono et al, 2016], . . .
recent system theory approaches inspired by output regulation [Lawrence et al. 2018]& robust control methods [Nelson et al. 2017], [Colombino et al. 2018]
5 / 31
Theory literature inspired by power systemslots of recent theory development stimulated by power systems problems
[Simpson-Porco et al., 2013], [Bolognaniet al, 2015], [Dall’Anese & Simmonetto,2016], [Hauswirth et al., 2016], [Gan &Low, 2016], [Tang & Low, 2017], . . .
1
A Survey of Distributed Optimization and ControlAlgorithms for Electric Power Systems
Daniel K. Molzahn,⇤ Member, IEEE, Florian Dorfler,† Member, IEEE, Henrik Sandberg,‡ Member, IEEE,Steven H. Low,§ Fellow, IEEE, Sambuddha Chakrabarti,¶ Student Member, IEEE,
Ross Baldick,¶ Fellow, IEEE, and Javad Lavaei,⇤⇤ Member, IEEE
Abstract—Historically, centrally computed algorithms havebeen the primary means of power system optimization and con-trol. With increasing penetrations of distributed energy resourcesrequiring optimization and control of power systems with manycontrollable devices, distributed algorithms have been the subjectof significant research interest. This paper surveys the literatureof distributed algorithms with applications to optimization andcontrol of power systems. In particular, this paper reviewsdistributed algorithms for offline solution of optimal power flow(OPF) problems as well as online algorithms for real-time solutionof OPF, optimal frequency control, optimal voltage control, andoptimal wide-area control problems.
Index Terms—Distributed optimization, online optimization,electric power systems
I. INTRODUCTION
CENTRALIZED computation has been the primary waythat optimization and control algorithms have been ap-
plied to electric power systems. Notably, independent systemoperators (ISOs) seek a minimum cost generation dispatchfor large-scale transmission systems by solving an optimalpower flow (OPF) problem. (See [1]–[8] for related litera-ture reviews.) Other control objectives, such as maintainingscheduled power interchanges, are achieved via an AutomaticGeneration Control (AGC) signal that is sent to the generatorsthat provide regulation services.
These optimization and control problems are formulatedusing network parameters, such as line impedances, systemtopology, and flow limits; generator parameters, such as costfunctions and output limits; and load parameters, such as anestimate of the expected load demands. The ISO collects allthe necessary parameters and performs a central computationto solve the corresponding optimization and control problems.
With increasing penetrations of distributed energy resources(e.g., rooftop PV generation, battery energy storage, plug-invehicles with vehicle-to-grid capabilities, controllable loads
⇤: Argonne National Laboratory, Energy Systems Division, Lemont, IL,USA, [email protected]. Support from the U.S. Department of En-ergy, Office of Electricity Delivery and Energy Reliability under contractDE-AC02-06CH11357.†: Swiss Federal Institute of Technology (ETH), Automatic Control Labora-tory, Zurich, Switzerland, [email protected]‡: KTH Royal Institute of Technology, Department of Automatic Control,Stockholm, Sweden, [email protected]§: California Institute of Technology, Department of Electrical Engineering,Pasadena, CA, USA, [email protected]¶: University of Texas at Austin, Department of Electrical and Computer En-gineering, Austin, TX, USA, [email protected],[email protected]. Support from NSF grant ECCS-1406894.⇤⇤: University of California, Berkeley, Department of Industrial Engineeringand Operations Research, Berkeley, CA, USA, [email protected]
providing demand response resources, etc.), the centralizedparadigm most prevalent in current power systems will poten-tially be augmented with distributed optimization algorithms.Rather than collecting all problem parameters and performinga central calculation, distributed algorithms are computedby many agents that obtain certain problem parameters viacommunication with a limited set of neighbors. Depending onthe specifics of the distributed algorithm and the application ofinterest, these agents may represent individual buses or largeportions of a power system.
Distributed algorithms have several potential advantagesover centralized approaches. The computing agents only haveto share limited amounts of information with a subset ofthe other agents. This can improve cybersecurity and reducethe expense of the necessary communication infrastructure.Distributed algorithms also have advantages in robustness withrespect to failure of individual agents. Further, with the abilityto perform parallel computations, distributed algorithms havethe potential to be computationally superior to centralizedalgorithms, both in terms of solution speed and the maxi-mum problem size that can be addressed. Finally, distributedalgorithms also have the potential to respect privacy of data,measurements, cost functions, and constraints, which becomesincreasingly important in a distributed generation scenario.
This paper surveys the literature of distributed algorithmswith applications to power system optimization and control.This paper first considers distributed optimization algorithmsfor solving OPF problems in offline applications. Many dis-tributed optimization techniques have been developed con-currently with new representations of the physical modelsdescribing power flow physics (i.e., the relationship betweenthe complex voltage phasors and the power injections). Thecharacteristics of a power flow model can have a large impacton the theoretical and practical aspects of an optimizationformulation. Accordingly, the offline OPF section of thissurvey is segmented into sections based on the power flowmodel considered by each distributed optimization algorithm.This paper then focuses on online algorithms applied toOPF, optimal voltage control, and optimal frequency controlproblems for real-time purposes.
Note that algorithms related to those reviewed here havefound a wide variety of power system applications in dis-tributed optimization and control. See, for instance, surveyson the large and growing literature relevant to distributedoptimization of electric vehicle charging schedules [9] anddemand response applications [10] as well as work on dis-tributed solution of multi-period formulations for model pre-
Steven Low
Enrique Mallada
John Simpson-Porco
Changhong Zhao
Claudio De Persis
Nima Monshizadeh
Arjan Van der SchaftMarcello Colombino
Emiliano Dall’Anese
Sairaj Dhople
Andrey Bernstein
Krishnamurthy Dvijotham
Andrea Simonetto
Na Li
Sergio Grammatico
Yue Chen
Florian Dörfler
Saverio Bolognani
Sandro ZampieriJorge Cortez
Henrik Sandberg
Karl Johansson
Ioannis Lestas
Andre Jokic
early adoption: KKT control [Jokic et al, 2009]
literature kick-started ∼ 2013 by groups fromCaltech, UCSB, UMN, Padova, KTH, & Groningen
changing focus: distributed & simple→ centralized & complex models/methods
implemented in microgrids (NREL, DTU, EPFL, . . . )& conceptually also in transactive control pilots (PNNL)
6 / 31
Overview
algorithms & closed-loop stability analysisprojected gradient flows on manifoldsrobust implementation aspectspower system case studies throughout
7 / 31
ALGORITHMS & CLOSED-LOOP
STABILITY ANALYSIS
Stylized optimization problem & algorithmsimple optimization problem
minimizey,u
φ(y, u)
subject to y = h(u)
u ∈ U
cont.-time projected gradient flow
u = ΠgU
(−∇φ
(h(u), u
))= Πg
U
(−[∂h∂u
I]∇φ(y, u)
)∣∣∣y=h(u)
Fact: a regular† solution u : [0,∞]→Xconverges to critical points if φ has Lip-schitz gradient & compact sublevel sets.
projected dynamical system
x ∈ ΠgX [f ](x) , arg min
v∈TxX‖v − f(x)‖g(x)
I domain XI vector field fI metric gI tangent cone TX
all sufficiently regular†
† regularity conditions made precise later8 / 31
Algorithm in closed-loop with LTI dynamicsoptimization problem
minimizey,u
φ(y, u)
subject to y = Hiou+Riow
u ∈ U
→ open & scaled projected gradient flow
u = ΠU(−ε[HTio I
]∇φ(y, u)
)
LTI dynamicsx = Ax+Bu+ Ew
y = Cx+Du+ Fw
const. disturbance w & steady-statemaps
x = −A−1B︸ ︷︷ ︸His
u −A−1E︸ ︷︷ ︸Rds
w
y =(D − CA−1B
)︸ ︷︷ ︸Hio
u +(F − CA−1E
)︸ ︷︷ ︸Rdo
w
ε∫
Uu
B∫
w E A
∇u φ D F
HTio∇y φ
yC
+ x
++
+++
− ++−
9 / 31
Stability, feasibility, & asymptotic optimality
Theorem: Assume thatregularity of cost function φ: compact sublevel sets & `-Lipschitz gradient
LTI system asymptotically stable: ∃ τ > 0 , ∃P � 0 : PA+ATP � −2τP
sufficient time-scale separation (small gain): 0 < ε < ε? , 2τcond(P )
· 1`‖Hio‖
Then the closed-loop system is stable and globally converges to the criticalpoints of the optimization problem while remaining feasible at all times.
Proof: LaSalle/Lyapunov analysis via singular perturbation [Saberi & Khalil ’84]
Ψδ(u, e) = δ · eTP e︸ ︷︷ ︸LTI Lyapunov function
+ (1− δ) · φ(h(u), u
)︸ ︷︷ ︸
objective function
with parameter δ∈(0, 1) & steady-state error coordinate e=x−Hisu−Rdsw
→ derivative Ψδ(u, e) is non-increasing if ε ≤ ε? and for optimal choice of δ
10 / 31
Example: optimal frequency control
dynamic LTI power system modelpower balancing objective
control generation set-points
unmeasured load disturbances
measurements: frequency + constraint variables (injections & flows)
I linearized swing dynamicsI 1st-order turbine-governorI primary frequency droopI DC power flow approximation
optimization problem
→ objective: φ(y, u) =cost(u)︸ ︷︷ ︸economic generation
+ 12‖max{0, y − y}‖2Ξ + 1
2‖max{0, y − y}‖2Ξ︸ ︷︷ ︸
operational limits (line flows, frequency, . . . )
→ constraints: actuation u ∈ U & steady-state map y = Hiou+Rdow
→ control u = ΠU (. . .∇φ) ≡ super-charged Automatic Generation Control
11 / 31
Test case: contingencies in IEEE118 systemevents: generator outage at 100 s & double line tripping at 200 s
0 50 100 150 200 250 3000
2
4
6
Time [s]
Power Generation (Gen 37) [p.u.]
Setpoint Output
12 / 31
How conservative is ε < ε? ?still stable for ε = 2 ε?
−5
0
5
·10−2 Frequency Deviation from f0 [Hz]
System Frequency
0 5 10 15 20
0
1
2
3
Time [s]
Line Power Flow Magnitudes [p.u.]
23→26 90→26 flow limit other lines
unstable for ε = 10 ε?
−2
0
2
4
Frequency Deviation from f0 [Hz]
System Frequency
0 5 10 15 20
0
2
4
Time [s]
Line Power Flow Magnitudes [p.u.]
23→26 90→26 flow limit other lines
Note: conservativeness problem dependent & depends, e.g., on penalty scalings13 / 31
Highlights & comparison of approachWeak assumptions on plant
internal stability→ no observability / controllability→ no passivity or primal-dual structure
measurements & steady-state I / O map→ no knowledge of disturbances→ no full state measurement→ no dynamic model
Weak assumptions on costLipschitz gradient + properness
→ no (strict/strong) convexity required
Parsimonious but powerful setuppotentially conservative bound, but
→ minimal assumptions onoptimization problem & plant
robust & extendable proof→ nonlinear dynamics→ time-varying disturbances→ general algorithms
take-away: open online optimizationalgorithms can be applied in feedback
→ Hauswirth, Bolognani, Hug & Dörfler (2020)“Timescale Separation in Autonomous Optimization”→ Menta, Hauswirth, Bolognani, Hug & Dörfler (2018)
“Stability of Dynamic Feedback Optimizationwith Applications to Power Systems”
14 / 31
Nonlinear systems & general algorithmsgeneral system dynamics x = f(x, u) with steady-state map x = h(u)
incremental Lyapunov functionW (x, u) w.r.t error coordinate x− h(u)
W (x, u) ≤ −γ ‖x− h(u)‖2 ‖∇uW (x, u)‖ ≤ ζ ‖x− h(u)‖
variable-metric Q(u) ∈ Sn+ gradient flow
u = −Q(u)−1∇φ(u)
examples: Newton method Q(u)=∇2φ(u)
or mirror descent Q(u)=∇2ψ(∇ψ(u)−1)
stability condition: ζ`γ·supu‖Q(u)−1‖ < 1
Similar results for algorithms with memory:momentum methods (e.g., heavy-ball)
(exp. stable) primal-dual saddle flows
non-examples: bounded-metricor Lipschitz assumption violated
0 10 20 30 40 500
5
10
15
20Cost Value
Dynamic IC
Algebraic IC
0 20 40 60 80 10010
-10
10-5
100
105
1010
Cost Value
Dynamic IC
Algebraic IC
co
st
valu
e
algebraic plant
dynamic plant
algebraic plant
dynamic plant
discontinuous subgradient
Nesterov acceleration
15 / 31
Highly nonlinear & dynamic test case
Nordic system: case study known forvoltage collapse (South Sweden ’83)
(static) voltage collapse: sequenceof events→ saddle-node bifurcation
high-fidelity model of Nordic systemI RAMSES + Python + MATLABI state: heavily loaded system & large
power transfers: north→ centralI load buses with Load Tap ChangersI generators equipped with Automatic
Voltage Regulators, Over ExcitationLimiters, & speed governor control
g15
g11
g20
g19
g16
g17
g18
g2
g6
g7
g14
g13
g8
g12
g4
g5
g10 g3g1
g9
4011
4012
1011
1012 1014
1013
10221021
2031
cs
404640434044
40324031
4022 4021
4071
4072
4041
1042
10451041
4063
40611043
1044
4047
4051
40454062
400 kV
220 kV
130 kV synchronous condenserCS
NORTH
CENTRAL
EQUIV.
SOUTH
4042
2032
41
1 5
3
2
51
47
42
61
62
63
4
43 46
3132
22
11 13
12
72
71
16 / 31
Voltage collapse
event: 250 MW load rampfrom t = 500 s to t = 800 s
unfortunate control response:non-coordinated + saturation
I extra demand is balanced byprimary frequency control
I cascade of activation ofover-excitation limiters
I load tap changers increasepower demand at load buses
bifurcation: voltage collapse
very hard to mitigate viaconventional controllers
→ apply feedback optimizationto coordinate set-pointsof Automatic Voltage Controllers
17 / 31
Voltage collapse averted !distance-to-collapse objective : φ = −log det
(power flow Jacobian
)
18 / 31
PROJECTED GRADIENT
FLOWS ON MANIFOLDS
Motivation: steady-state AC power flowstationary model
Ohm’s Law Current Law
AC power
AC power flow equations
(all variables and parameters are -valued)imagine constraints slicing this set⇒ nonlinear, non-convex, disconnected
additionally the parameters are ±20%
uncertain . . . this is only the steady state!
graphical illustration of AC power flow
[Hiskens, 2001]
[Molzahn, 2016]19 / 31
Key insights on physical equality constraint
1.5
1
0.5
q2
0
-0.5
-11.51
0.5
p2
0-0.5
-1
1.2
1
1.4
0.8
0.6
v2
vdc
idc
m
iI
v
LI
CI GI
RI
τm
θ, ω
vf
v
if
τe
is
Lθ
Mrf
rs rs
v
iTLT
CG GqC v
RTiI
AC power flow is complex but takesthe form of a smooth manifold
→ local tangent plane approximations,local invertibility, & generic LICQ
→ regularity (algorithmic flexibility)
→ Hauswirth, Bolognani, Hug, & Dörfler (2015)“Fast power system analysis via implicitlinearization of the power flow manifold”
→ Bolognani & Dörfler (2018)“Generic Existence of Unique LagrangeMultipliers in AC Optimal Power Flow”
AC power flow is attractive steadystate for ambient physical dynamics
→ physics enforce feasibility even fornon-exact (e.g., discrete) updates
→ robustness (algorithm & model)
→ Gross, Arghir, & Dörfler (2018)“On the steady-state behavior ofa nonlinear power systemmodel”
20 / 31
Feedback optimization on the manifold
challenging specificationson closed-loop trajectories:1. stay on manifold at all times2. satisfy constraints at all times3. converge to optimal solution
feedbackoptimizationalgorithm
x = ΠgX (−gradφ(x))
physical steady-statepower system
(AC power flow)
Sk + wk =∑
`1
zk`
∗Vk(V∗k − V∗` )
renewablesloads w
generationsetpoints
measurements
prototypical optimal power flow
minimize φ(x)
subject to x ∈ X =M∩K
φ : Rn → R objective functionM⊂ Rn AC power flow manifoldK ⊂ Rn operational constraints
v
TxX
X
projection of trajectory on feasible cone21 / 31
Simple low-dimensional case studies . . .. . . can have simple feasible sets . . . or can have really complex sets
v0 = 1
slack bus generator
qG ∈ [q, q]
vref = 1
load
pL(t)
pG
1j
θ0 = 0
0-2
0.5
3
v
1
2
pG-pL
0
qG
10
2 -1
application demands sophisticated level of generality !22 / 31
Projected dynamical systems on irregular domains
Theorem: Consider a Carathéodory solutionx : [0,∞)→ X of the initial value problem
x = ΠgX (−gradφ(x)) , x(0) = x0 ∈ X .
If φ has compact sublevel sets on X , then x(t)
converges to the set of critical points of φ on X .
Hidden assumption: existence, uniqueness, &completeness of Carathéodory solution x(t) ∈ Xin absence of convexity, Euclidean space, . . . ? X =
{x : ‖x‖22 = 1 , ‖x‖1 ≤
√2}
regularity conditions constraint set vector field metric manifoldexistence of Krasovski loc. compact loc. bounded bounded C1
existence of Carathéodory Clarke regular C0 C0 C1
uniqueness of solutions prox regular C0,1 C0,1 C1,1
→ Hauswirth, Bolognani, & Dörfler (2018)“Projected Dynamical Systems on Irregular
Non-Euclidean Domains for Nonlinear Optimization”
→ Hauswirth, Bolognani, Hug, & Dörfler (2016)“Projected gradient descent on Riemanniann manifolds with
applications to online power system optimization”23 / 31
ROBUST IMPLEMENTATION ASPECTS
Robust implementation of projectionsprojection & integrator→ windup→ robust anti-windup approximation→ saturation often “for free” by physics
K
∫PU
k(·, u) x = f(x, ·)
+
−
u
PU (u)
−+
u = ΠU [k(x, ·)](u)K → ∞
disturbance→ time-varyingdomain
f(x)
ΠtX f(x)
X (t)
X (t + δ)
I temporal tangentcone & vector field
I ensure suff. regularity& tracking certificates
→ Hauswirth, Dörfler, & Teel (2020)“Anti-Windup Approximations of Oblique Projected Dynamical
Systems for Feedback-based Optimization”
handling uncertainty when enforcingnon-input constraints : x ∈ X or y ∈ Y
I cannot measure state x directly→ Kalman filtering: estimation& separation
I cannot enforce constraints on y=h(u)by projection (not actuated & h(·)unknown)
→ soft penalty or dualization + grad flows(inaccurate, violations, & strong assumptions)
→ project on1st order predictionof y=h(u)
y+ ≈ h(u)︸︷︷︸measured
+ ε ∂h∂u︸︷︷︸
steady-stateI/O sensitivity
w︸︷︷︸feasible descent
direction
⇒ global convergence to critical points
→ Häberle, Hauswirth, Ortmann, Bolognani, & Dörfler (2020)“Enforcing Output Constraints in Feedback-based Optimization”
→ Hauswirth, Subotić, Bolognani, Hug, & Dörfler (2018)“Time-varying Projected Dynamical Systems with Applications. . . ”24 / 31
Tracking performance under disturbances
G1
G2 C1
C3 C2
W
S
Generator
Synchronous Condensor
Solar
Wind
G
C
S
W
5.3.2 30 Bus Power Flow Test Case
To investigate the capabilities of our new scheme under time-varying generation limits and fluc-tuating load conditions, we consider a power system setup based on the IEEE 30 bus power flowtest case, where wind and solar generation has been added, similar to the one adopted in [13].The grid topology is shown in Figure 5.4, where the controllable units along with their gener-ation limits and the operational constraints of the associated buses are listed in Table 5.3. Inparticular, the upper power generation limit of the solar and the wind farm are time-varying, dueto the fluctuating nature of the corresponding primary sources. This results in a time-varyingconstraint set U of the controllable variables. Additional operational constraints that need to besatisfied include line current limits for different branches. The total generation cost � we aim tominimize is composed of the costs of each generator in [$/h], given as aip
2i +bipi, where ai, bi > 0
are constant cost-coefficients provided in Table 5.3. The marginal operating cost of the solar andthe wind farm is set to zero.
We simulate 24 hours of operation and run Algorithm 1, where the controller receives fieldmeasurements of the system state z every minute. The demand profile is shown in Figure 5.4,which exhibits an abrupt demand reduction of approximately 20% between 20:30 and 21:30 atseveral system buses.
G1
G2 C1
C3 C2S
Generator
�������������� � �
Solar
Wind
G
C
S
W
W
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 2470
80
90
100
110
120
Figure 5.4: Modified IEEE 30 bus power flow test case.
Unit Type Set-points ai bi pgi pgi
qgi qgi
vi vi
G1 Gen.1 V✓, v 0.1 0.9 75 0 50 -50 1.06 0.94freq. ctrl.
G2 Gen.2 PV p, v 0.04 0.5 60 0 50 -50 1.06 0.94C1 Cond.1 PV v 0 0 0 0 50 -50 1.06 0.94C2 Cond.2 PV v 0 0 0 0 50 -50 1.06 0.94C3 Cond.3 PV v 0 0 0 0 50 -50 1.06 0.94S Solar PQ p, q 0 0 ps(t) 0 50 -50 1.06 0.94W Wind PQ p, q 0 0 pw(t) 0 50 -50 1.06 0.94
line 1-2 line 6-8 line 12-15
Table 5.3: Cost coefficients a and b in [$/MW2h] and [$/MWh], respectively. Active powergeneration limits in [MW] and reactive power generation limits in [MVAr], and bus voltagelimits in [p.u.]. The system base power is fixed to 100MVA.
41
net demand: load, wind, & solar (discontinuous)
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
-50
0
50
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
-50
0
50
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0.95
1
1.05
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0.95
1
1.05
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
0.1
0.2
0.3
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
0.1
0.2
0.3
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240
100
200
300
(a) Simulation results of controlled 30 buspower system with exact Jacobian matrixru,yF (u, y, w).
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240
100
200
300
(b) Simulation results of controlled 30 buspower system with constant approximation ofthe Jacobian matrix.
Figure 5.5: Simulation results of controlled 30 bus power system for the exact Jacobian matrixru,yF (u, y, w) and a constant approximation thereof. The dashed lines represent the constraintsand the colors are the same as in Table 5.3.
43
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
-50
0
50
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
-50
0
50
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0.95
1
1.05
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0.95
1
1.05
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
0.1
0.2
0.3
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
0.1
0.2
0.3
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240
100
200
300
(a) Simulation results of controlled 30 buspower system with exact Jacobian matrixru,yF (u, y, w).
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240
100
200
300
(b) Simulation results of controlled 30 buspower system with constant approximation ofthe Jacobian matrix.
Figure 5.5: Simulation results of controlled 30 bus power system for the exact Jacobian matrixru,yF (u, y, w) and a constant approximation thereof. The dashed lines represent the constraintsand the colors are the same as in Table 5.3.
43
25 / 31
Optimality despite disturbances&uncertainty
transient trajectory feasibilitypractically exact tracking ofground-truth optimizer(omniscient & no computation delay)
robustness to model mismatch(asymptotic optimality under wrong model)
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
-50
0
50
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
-50
0
50
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0.95
1
1.05
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0.95
1
1.05
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
0.1
0.2
0.3
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
0.1
0.2
0.3
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240
100
200
300
(a) Simulation results of controlled 30 buspower system with exact Jacobian matrixru,yF (u, y, w).
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240
100
200
300
(b) Simulation results of controlled 30 buspower system with constant approximation ofthe Jacobian matrix.
Figure 5.5: Simulation results of controlled 30 bus power system for the exact Jacobian matrixru,yF (u, y, w) and a constant approximation thereof. The dashed lines represent the constraintsand the colors are the same as in Table 5.3.
43
offline optimization feedback optimizationmodel uncertainty feasible ? φ− φ∗ ‖v − v∗‖ feasible ? φ− φ∗ ‖v − v∗‖loads ±40% no 94.6 0.03 yes 0.0 0.0line params ±20% yes 0.19 0.01 yes 0.01 0.0032 line failures no -0.12 0.06 yes 0.19 0.007
conclusion: simple algorithm performs extremely well & robust26 / 31
EXPERIMENTS
Experimental case study @ DTU
EV
SE
1
EV
SE
2
EV
SE
3
EV
SE
4
EV
SE
5
EV
SE
6
EV
SE
7
EV
SE
8
Busbar A
Busbar B
Busbar A
Busbar B
630 kVA100 kVA
Bat
tery
Ext
. 117
-5
Cab
le C
2
Cab
le C
1
CE
E
Ext
. 117
-2
Chg
. pos
t
Cab
le D
1
PV
NE
VIC
Busbar A Busbar B Busbar C
150 kVA
100 kVA
Gaia
Flexhouse
PV
Cable B1
Cable B2
Busbar B Busbar B
Busbar A 200 kVA
Static load
Diesel
CEE
Aircon
Cable A1
Cable A2
PV
Busbar B
Busbar A Busbar B
Cable F1
Flexhouse 2 Flexhouse 3
Cable E1
Cable E2
CEE
CHP
Heatpump 1
BoosterHeater
Cable F1
Crossbar switch
Load conv.
SYSLABbreaker overview
Building 716 Building 715 Building 319
Building 117
Ship
Shore
Mach.set
Con
tain
er 1
Con
tain
er 2
Con
tain
er 3
I
I
I
PCC
v1 v2 v3R1, L1 R2, L2 R3, L3
p1, q1 p2, q2 p3, q3
PV1 PV2 Battery
±8 kVAr
Static load
±6 kVAr±6 kVAr 0 kVAr
10 kW0 kVAr0 kW −15 kW
Voltage[p.u.]
10.99
1.061.05
0.95
21min experiment with eventsI t = 3min: control turned ONI t ∈ [11, 14]min: Pbatt = 0 kW
base-line controllersdecentralizednonlinearproportionaldroop control(IEEE1547.2018)
vi
qiqmaxi
qmini
vmin vmax
qi(t+ 1) = fi(vi(t))
1
comparison of three controllersI decentralized controlI feedforward optimizationI feedback optimization
→ Ortmann, Hauswirth, Caduff, Dörfler, & Bolognani (2020)“Experimental Validation of Feedback Optimization
in Power Distribution Grids”
27 / 31
Decentralized feedback controldecentralized nonlinear proportional droop control
0.97
1
1.031.051.07
−5
0
5
0 5 10 15 20
0.970.980.99
1
Time [min]0 5 10 15 20
0
1
2
0.970.980.99
1
Volta
ge[p.u.]
0
1
2
Reac
tivePo
wer[kV
Ar]Battery
PV2
PV1
constraint violations due to local control saturation & lack of coordination
28 / 31
Successive feedforward optimizationcentralized, omniscient, & successively updated at high sampling rate
0.97
1
1.031.051.07
−5
0
5
0 5 10 15 20
0.970.980.99
1
Time [min]0 5 10 15 20
−2
−1
0
0.970.980.99
1
Volta
ge[p.u.]
−2
−1
0
Reac
tivePo
wer[kV
Ar]Battery
PV2
PV1
performs well but persistent constraint violation due to model uncertainty
29 / 31
Feedback optimizationprimal-dual flow with 10 s sampling time requiring only model I/O sensitivity∇h (or an estimate)
0.97
1
1.031.051.07
−5
0
5
0 5 10 15 20
0.970.980.99
1
Time [min]0 5 10 15 20
−6
−4
−2
0
0.970.980.99
1
Volta
ge[p.u.]
−4
−2
0
Reac
tivePo
wer[kV
Ar]Battery
PV2
PV1
excellent performance & model-free(!) since ∇h(u) approximated by[
1 1 11 1 11 1 1
]30 / 31
CONCLUSIONS
ConclusionsSummary
open & online feedback optimization algorithms as controllers
approach: projected dynamical systems & time-scale separation
unified framework: broad class of systems, algorithms, & programs
illustrated throughout with non-trivial power systems case studies
Ongoing work & open directions
analysis: robustness, performance, stochasticity, sampled-data
algorithms: 0th-order, sensitivity estimation, distributed, minmax
power systems: more experiments, virtual power plant extensions
further app’s: seeking optimality in uncertain & constrained systems
It works much better than it should ! We still need to fully grasp why ?
31 / 31
Thanks !
Florian Dörflerhttp://control.ee.ethz.ch/~floriand
[link] to related publications