an optimization-on-manifold approach to the design of...

An optimization-on-manifold approach to the design of distributed feedback control in smart grids Saverio Bolognani

An optimization-on-manifold approachto the design of distributed feedback control

in smart grids

Saverio Bolognani, Florian Dörfler

Automatic Control LaboratoryETH Zürich

ECC 2016 WorkshopDistributed and Stochastic Optimization: Theory and Applications


Future electric power distribution grids

FUTURE ELECTRIC POWER DISTRIBUTION GRIDS



Power distribution grids

transmissiongrid

distributiongrid

TraditionalPower

Generation

I Distribution grid: the“capillary system” ofpower networks

I It delivers power fromthe transmission gridto the consumers.

I Very little sensing,monitoring, actuation.

I The “easy” part of thegrid: conventionallyfit-and-forget design.



New challengesI Distributed microgenerators (conventional and renewable sources)I Electric mobility (large flexible demand, spatio-temporal patterns).

41GW75%

Germany17 August 2014 wind

solarhydro

biomass

Distribution grid

solar

wind

hydro + biomass

Installed renewable generationGermany 2013

24 GW

15 GW

Transmission grid

6 GW

2015 2020

200k

400k

600k

800k

PHEV

BEV

SwitzerlandVISION 2020

Electricityconsumption

Buildings40.9%

Industry31.3%

Transportation27.8%

Energy consumptionby sector(2010)

73.9%

25.9%

Primary fuelconsumption

Electric VehicleFast charging

120KWTeslasupercharger

4KWDomestic

consumer



Distribution grid congestion

Operation of the grid close or above the physical limits, due tosimultaneous and uncoordinated power demand/generation.

→ lower e�ciency, blackouts→ curtailment of renewable generation

→ bottleneck to electric mobility

Fit-and-forget→ unsustainable grid reinforcement



Distribution Grid

Transmission Grid

x x xx

Distribution Grid

Transmission Grid

CONT

ROL

LAYE

R

Curtailment Reduced hosting capacity

Higher renewable generationLarger hosting capacity

overvoltage

renewable generation

controlled

undervoltage

power demand

uncontrolled

uncontrolled

controlled

distributiongrid

control

controlcontrol

control

I Virtual grid reinforcementI same infrastructureI more sensors and intelligenceI controlled grid = larger capacity

I Transparent control layerI invisible to the usersI modular design


OVERVIEW

1. A feedback control approach2. A tractable model for control design3. Control design example

I Reactive power control for voltage regulation4. Next step

I Optimization on the power flow manifold

5. Conclusions


A feedback control approach

A FEEDBACK CONTROL APPROACH



Distribution grid model

active power ph

reactive power qh

voltage magnitude vh

voltage angle θh h

0

microgenerator load

supply point

Grid equations

diag(u)Yu = s

whereI uh = vhejθh complex voltagesI sh = ph + jqh complex powers




active power ph

reactive power qh


voltage angle θh h

0

microgenerator load

supply point

Actuation

I Tap changer / voltage regulators – supply point voltage v0I Reactive power compensators – reactive power qh

I static compensatorsI power inverters of the microgenerators (when available)

I Active power management – active power phI smart building control, storage and deferrable loadsI generator curtailment and load shedding




active power ph

reactive power qh


voltage angle θh h

0

microgenerator load

supply point

Sensing

I Power meters – active power ph and reactive power qhI Voltage meters – nodal voltage vhI Phasor measurement units (PMU) – voltage magnitude vh and angle θh

(PQube @ UC Berkeley, GridBox in Zürich/Bern, Smart Grid Campus @ EPFL)I Line currents, transformer loading, . . .



A control framework

grid sensing

grid actuation

Power distributionnetwork

plant

state x

power demands

power generation

Control objective

Drive the system to a state x∗ =[v∗ θ∗ p∗ q∗

]subject to

I soft constraints x∗ = argminx J(x)

I hard constraints x ∈ XI chance constraints P [x 6∈ X ] < ε



Feedforward control

grid sensing

grid actuation


plant

state x

power demands

power generation

OPF

Conventional approach: Optimal Power Flow

I Similar to power transmission grid OPFI Motivated by encouraging results on OPF convexification

(Lavaei (2012), Farivar (2013), . . . )I Requires full disturbance knowledge - full communicationI Heavily model basedI Requires co-design of grid control and users’ behavior



Feedback control

grid sensing

grid actuation


plant

state x

power demands

power generation

FEEDBACK

input

disturbance

output

Control theory answer

I Robustness against parametric uncertainty/unmodeled disturbanceI Time varying demand/generation becomes disturbanceI Model-free designI Explored so far only for limited cases (e.g. purely local VAR control)I Allows modular design of grid control and users’ behavior



A similar scenario: frequency control

frequencyPowernetwork

plant

state x

power demands

power generation

FEEDBACK

input

disturbance

output

primary control

secondary control

In the transmission grid, feedback is used for frequency regulationI Frequency deviation as a implicit signal for power unbalanceI Purely local proportional control: primary droop controlI Integral control: secondary frequency regulation


A tractable model for control design

A TRACTABLE MODEL FOR CONTROL DESIGN



Power flow manifoldI Grid state x =

[v θ p q

]I Set of all states that satisfy the grid equations diag(u)Yu = s

→ power flow manifold M := {x | F(x) = 0}I Regular submanifold of dimension 2n (6n if three-phase)

10.5

p2

0-0.5

0

0.5

q2

1

1.2

1

0.8

0.6

0.4

v2

node 2node 1

v1 = 1, θ1 = 0

y = 0.4− 0.8j

v2, θ2p2, q2p1, q1

v21g− v1v2 cos(θ1 − θ2)g− v1v2 sin(θ1 − θ2)b = p1

−v21b + v1v2 cos(θ1 − θ2)b− v1v2 sin(θ1 − θ2)g = q1

v22g− v1v2 cos(θ2 − θ1)g− v1v2 sin(θ2 − θ1)b = p2

−v22b + v1v2 cos(θ2 − θ1)b− v1v2 sin(θ2 − θ1)g = q2



Power flow manifold approximation

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

Best linear approximant

Ax∗(x − x∗) = 0

Ax∗ :=∂F(x)

∂x

∣∣∣∣x=x∗

Tangent plane at a nominalpower flow solution x∗ ∈M

Example x∗: no-load solution

I Implicit – No input/outputs (not a disadvantage)I Sparse – The matrix Ax∗ has the sparsity pattern of the grid graphI Structure preserving – Elements of Ax∗ depend on local parameters

→ Bolognani & Dörfler, Allerton (2015)→ Source code on github

http://control.ee.ethz.ch/~bsaverio/papers/BolognaniDorfler_Allerton2015.pdf

https://github.com/saveriob/1ACPF



Power flow manifold approximation

2

1

0

�2

-1

-21.4

1.2

v2

1

0.8

0.6

0.5

-1

-0.5

0

1

1.5

p2

Standard modelsAdding assumption one obtains

I linear coupled power flowI DC power flowI rectangular DC flow

1.510.5

q2

0-0.5-121

p2

0

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4

-1

v2

Nonlinear coordinate transf.

xh = xh(xh),∂xh∂xh

= 1

Di�erent manifold curvature!I vh → v2h : LinDistFlow


Control design examples

Reactive power control for voltage regulation

CONTROL DESIGN EXAMPLES

REACTIVE POWER CONTROL FOR VOLTAGE REGULATION




Problem statement

active power ph

reactive power qh


voltage angle θh h

0

microgenerator load

supply point

I Inputs: reactive power qh of microgeneratorsI Outputs: voltage measurement vh at the microgeneratorsI Control objective:

I Soft constraints: minimize J(x) = vTLv (voltage drops)I Hard constraints: guarantee Vmin ≤ vh ≤ Vmax at all sensors




Control design for soft constraint

power flow manifold

linear approximant

1. Modeling assumption

2. Control specs3. Proj on linear manifold4. Derive feedback law

Modeling assumption: constant R/X ratio ρ.

Ax∗(x − x∗) = 0 becomes (around the no-load state)

[ρL −L−L −ρL

∣∣∣∣ −I 00 −I

]vθpq

= 0





power flow manifold

linear approximant

1. Modeling assumption2. Control specs

3. Proj on linear manifold4. Derive feedback law

Control specification: Distributed and asynchronous.

Minimal update δq{qh ← qh + δqk ← qk − δ

→ Communication graph Gcomm describing all possible updates (pairs h, k).





power flow manifold

linear approximant

search direction

x

1. Modeling assumption2. Control specs3. Proj on linear manifold

4. Derive feedback law

Search directions: By projecting each possible direction δq on the linearmanifold ker Ax∗ , we obtain feasible search directions in the state space.

δx =

− 1

1+ρ2 L†δq− ρ

1+ρ2 L†δq0δq





power flow manifold

linear approximant

search direction

x

Gradient of cost function

x + tδx1. Modeling assumption2. Control specs3. Proj on linear manifold4. Derive feedback law

Optimal step length: Given a search direction δx, we determine the steplength that minimizes the cost function J(x) = vTLv.

∇J(x) =

2Lv000

∇J(x + tδx)Tδx = 0 ⇒ t = (1 + ρ2)vTδq

δqTL†δq





power flow manifold

linear approximant

search direction

x


x + tδx1. Modeling assumption2. Control specs3. Proj on linear manifold4. Derive feedback law

Because the model is sparse and structure preserving. . .

t = (1 + ρ2)vTδq

δqTL†δq= (1 + ρ2)

vh − vkXhk

Gossip-like feedback law

{qh ← qh + (1 + ρ2) vh−vk

Xhk

qk ← qk − (1 + ρ2) vh−vkXhk




Convergence and performance analysis

microgenerator voltage vh

microgenerator reactive power qh


plant

state x

power demands

power generation

FEEDBACK

input

disturbance

output

I Asynchronous distributedfeedback control

I no demand or generation measurementI limited model knowledgeI no power flow solverI alternation of sensing and actuation.

{qh ← qh + (1 + ρ2) vh−vk

Xhk

qk ← qk − (1 + ρ2) vh−vkXhk




Convergence and performance analysis

0 10 20 30 40 50 60 70iteration

E [

J(x)

- J o

pt ]

1

0.1

0.01 56

58

60

62

100 150 200 250 300 350 400iteration

loss

es

[kW

]

→ Bolognani & Zampieri, IEEE TAC (2013)

I Extension to J(x) = uTLu (power losses), if θ can be measured (PMUs).I Proof of mean square convergence (with randomized async updates).I Explicit bound on the exponential rate of convergence.I Analysis of the dynamic performance (disturbance rejection).I Optimal communication graph: Gcomm ≈ Ggrid.

http://dx.doi.org/10.1109/TAC.2013.2270317




Communication co-design

0

microgenerator load

supply point Gcomm

Ggrid

GgridSparsity of thepower system

GcommSparsity of the

communication graph

Fundamental design problem: implications of the communicationarchitecture on the control performance.

Joint design vs. separation results




Communication co-design

Powerline communication10.0.0.1

18.0.1.2

18.0.1.3

... ...

...General purpose network

Wireless Multi-area




Control design with hard constraints




plant

state x

power demands

power generation

FEEDBACK

input

disturbance

output

v ≤ vh ≤ vqh≤ qh ≤ qh

I Power losses minimizationI Hard constraints on inputs and outputs.I Construct Lagrangian→ Saddle point algorithm

L(q, λ, η) = J(q) + λT(v − v) + ηT(qh − q)




Control design with hard constraints




plant

state x

power demands

power generation

FEEDBACK

input

disturbance

output

v ≤ vh ≤ vqh≤ qh ≤ qh

λh ← [λh + α(vh − v)]≥0

ηh ←[ηh + β(qh − qh)

]≥0

q← q− γ∇J(q)− λ−Lη

−Lη Di�usion term that requires nearest-neighbor communication.

→ Bolognani, Carli, Cavraro & Zampieri, IEEE TAC (2015)

http://dx.doi.org/10.1109/TAC.2014.2363931




Simulations and comparison

0 10 20 30 40 50

Volt

ag

e [

p.u

.]

1

1.05

1.1

React

ive p

ow

er

[p.u

.]

0

1v1

v2

|q1| |q2|

Modified IEEE 123 Distribution Test Feeder → githubLight load + 2 microgenerators→ overvoltage

2 sets of constraints:

{voltage limits vh ≤ vmax reactive power qh ≤ qh

https://github.com/saveriob/approx-pf





0 10 20 30 40 50

Volt

ag

e [

p.u

.]

1

1.05

1.1

React

ive p

ow

er

[p.u

.]

0

1

steady state error

saturation

vi

qi

qmaxi

−qmaxi

vmin vmax

qh(t) = −f (vh(t))

Fully decentralized, proportional controller.

Latest grid code drafts, Vovos (2007), Turitsyn(2011), Aliprantis (2013), ...





0 10 20 30 40 50

Volt

ag

e [

p.u

.]

1

1.02

1.04

1.06

1.08

React

ive p

ow

er

[p.u

.]

0

0.5

1

steady state error

saturation

vi

δqi

vmin

vmax

qh(t + 1) = qh(t)− f (vh(t))

Fully decentralized, integral controller.

Zero steady error without saturation limits. Li (2014)





0 10 20 30 40 50

Volt

ag

e [

p.u

.]

0.98

1

1.02

1.04

1.06

React

ive p

ow

er

[p.u

.]

0

0.5

1

no steady state error

reactive power sharing

λh ← [λh + α(vh − v)]≥0

ηh ←[ηh + β(qh − qh)

]≥0

q← q− γ∇J(q)− λ −Lη

Networked feedback control (neighbor-to-neighbor async communication)

→ Cavraro, Bolognani, Carli & Zampieri, IEEE CDC (2016)

http://people.ee.ethz.ch/~bsaverio/papers/Bolognani_CDC2016.pdf




Chance constraints on the state

Chance-constrained decision

mininput δ

J(δ)

subject to Prob [x /∈ Xc] < ε

I Xc can encodeI under/over voltage limitsI power injection limitsI voltage stability region→ Bolognani & Zampieri, IEEE TPS(2015)

I A stochastic model for thedisturbance is available

I Via linear approximant→deterministic polytope constraints → Bolognani & Dörfler, PSCC (2016)

http://dx.doi.org/10.1109/TPWRS.2015.2395452

http://dx.doi.org/10.1109/TPWRS.2015.2395452

http://control.ee.ethz.ch/~bsaverio/papers/BolognaniDorfler_PSCC2016.pdf


Next step

NEXT STEP

OPTIMIZATION ON THE POWER FLOW MANIFOLD


Next step

Optimization on the power flow manifold

power flow manifold

linear approximant

x(t)


Projected gradient

x

Continuous time trajectory on the manifold:1. ∇J(x): gradient of the cost function (soft constraints) in ambient space2. Πx∇J(x): projection of the gradient on the linear approximant in x3. Evolve according to x = −γΠx∇J(x)


Next step

Staying on the power flow manifold

x =

[xexoxendo

]Exogenous variablesInputs/disturbances that areimposed on the model.Reactive power injection qi

Endogenous variablesDetermined by the physics ofthe grid.Voltage vi

Iterative algorithm: at each step1. Compute Πx∇J(x) (sparse Ax(t) ⇒ distributed algorithm)2. Actuate system based on δx = −γΠx∇J (exogeneous variables / inputs)3. Retraction step x(t + 1) = Rx(t)(δx) ⇒ x(t + 1) ∈M.

From iterative optimization algorithm to feedback control on manifolds.


Next step

Hard constraints on exogenous variables

Feasible input region

I Can be enforced via saturation of the corresponding coordinatesI Primal feasibility at all timesI The resulting feasible input region is invariant with respect to the

retraction.I We can saturate δx = −γΠx∇J(x) because

x + δ(x) ∈ F ⇒ x(t + 1) = Rx(t)(δx) ∈ F

→ Geometric Projected Dynamical Systems

I Extension of results on existence and uniqueness of executions forhybrid automata to manifolds

I Guarantees of no Zeno execution

Ongoing work with Adrian Hauswirth, Gabriela Hug, Florian Dörfler.


Next step

Hard constraints on endogeneous variables

Operational constraints

I Barrier functions not suitable:I Backtracking line search is not possible in closed loopI Primal feasibility cannot be guaranteed during tracking

I Time-varying penalty functions not suitable:I Persistent feedback control for tracking

I Can be tackled via dualization / Lagrangian approach.I The corresponding operational constraints are satisfied at steady

state, despite model uncertainty.→ Saddle/primal-dual algorithm on manifolds

Ongoing work with Adrian Hauswirth, Gabriela Hug, Florian Dörfler.


Next step

Optimization on the power flow manifold

0 100 200 300 400

0

2

4

6

8a

ctive

po

we

r

0 100 200 300 4000.9

0.95

1

1.05

1.1

vo

lta

ge

0 100 200 300 400-4

-3

-2

-1

0

1

rea

ctive

po

we

r

p1

p2

v1

v2

q1

q2


Conclusions

CONCLUSIONS


Conclusions

Conclusions

A power system problem for control theory tools!

I A tractable modelI implicit linearI sparseI structure preserving

I Output feedbackin power systems

I model-freeI robustI limited measurement

I Networked controlI co-design?

I Feedback control on thepower flow manifold

I exploit the physics ofthe system in the loop

DistributionGrid

FeedbackControl

Real-timemeasurementsmicro-PMUvoltage meas.line currents

Controlsignals

reactive powertap changers

voltage regulators

Power demandPower generation

Operatinggrid state

operationalconstraints

feasible power region(uncontrolled grid)

feasible power region(virtual reinforcement)


Conclusions

Thanks!

Saverio [email protected]

This work is licensed under the Creative Commons Attribution 4.0 Intl License.

mailto://[email protected]

http://creativecommons.org/licenses/by/4.0/

an optimization-on-manifold approach to the design of...

Documents