• Aucun résultat trouvé

Experiments with the interior-point method for solving large scale optimal power flow problems

N/A
N/A
Protected

Academic year: 2021

Partager "Experiments with the interior-point method for solving large scale optimal power flow problems"

Copied!
9
0
0

Texte intégral

(1)

and education use, including for instruction at the authors institution

and sharing with colleagues.

Other uses, including reproduction and distribution, or selling or

licensing copies, or posting to personal, institutional or third party

websites are prohibited.

In most cases authors are permitted to post their version of the

article (e.g. in Word or Tex form) to their personal website or

institutional repository. Authors requiring further information

regarding Elsevier’s archiving and manuscript policies are

encouraged to visit:

http://www.elsevier.com/copyright

(2)

ContentslistsavailableatSciVerseScienceDirect

Electric Power Systems Research

j o ur na l ho me p ag e :w w w . e l s e v i e r . c o m / l o c a t e / e p s r

Experiments with the interior-point method for solving large scale Optimal

Power Flow problems

Florin Capitanescu

a,∗

, Louis Wehenkel

b

aUniversityofLuxembourg,InterdisciplinaryCentreforSecurity,ReliabilityandTrust(SnT),6,rueRichardCoudenhove-Kalergi,L-1359Luxembourg,Luxembourg

bUniversityofLiège,InstitutMontefiore,Sart-TilmanB28,B-4000Liège,Belgium

a rt i c l e i n f o

Articlehistory:

Received16March2012 Receivedinrevisedform 17September2012 Accepted1October2012

Keywords:

OptimalPowerFlow

Security-ConstrainedOptimalPowerFlow Interior-pointmethod

Nonlinearprogramming

a b s t r a c t

Thispaperreportsextensiveresultsobtainedwiththeinterior-pointmethod(IPM)fornonlinearpro- grammes(NLPs)stemmingfromlarge-scaleandseverelyconstrainedclassicalOptimalPowerFlow(OPF) andSecurity-ConstrainedOptimalPowerFlow(SCOPF)problems.Thepaperdiscussestransparentlythe problemsencounteredsuchasconvergencereliabilityandspeedissuesofthemethod.

© 2012 Elsevier B.V. All rights reserved.

1. Introduction

TheOptimalPowerFlow(OPF)[1]isastatic,nonlinear,non- convex,large-scale optimization problem withcontinuous and discretevariables.TheSecurity-ConstrainedOptimalPowerFlow (SCOPF)isageneralizationoftheOPFproblemthatensuresaddi- tionallythesystemsecuritywithrespecttoa setofpostulated contingencies[2].OPFisnowadaysanessentialtoolinpowersys- temsplanning,operationalplanningandreal-timeoperation.

Due to critical environmentalissues, nowadays mostpower systemshavetoaccommodatea significantlevel ofpenetration ofrenewableintermittentgeneration.Thisrequiressmarterways ofcontrolinreal-time accordingtotheprinciples:just-in-time, just-in-place,andjust-in-context[3].Sinceaclassicalday-ahead preventiveSCOPFapproachwouldbenotanymoresustainable,we foreseethattheneedforusinganautomaticadaptiveoptimalcon- trolscheme,andinparticulara(SC)OPFinreal-time,ismoreand morestringent.Thisraisesthesameoldconcernsregardingmostly thereliabilityandspeedofsuchanapproach[4,5].Furthermore, inEuropethereisatrendtocheckthesecurityandproposecoor- dinatedcontrolactionsforalargeinterconnection,composedby independentpowersystems,inabroaderway[6],whichcallsfor

∗ Correspondingauthor.Tel.:+3524666445345;fax:+3524666445669.

E-mailaddresses:florin.capitanescu@uni.lu(F.Capitanescu), l.wehenkel@ulg.ac.be(L.Wehenkel).

toolsabletodealwithverylargescaleoptimizationproblems,with uptohundredsofmillionsofvariablesandconstraints.

IPM[7,8]hasbeensuccessfullyappliedfornearlytwodecades tovariousOPFproblems[9–13].ThemainadvantagesoftheIPM are:(i)easeofhandlinginequalityconstraintsbylogarithmicbar- rierfunctions,(ii)speedofconvergence,and(iii)astrictlyfeasible initialpointisnotrequired.ThedrawbacksoftheIPMare:(i)the heuristictodecreasethebarrierparameter,(ii)therequiredpos- itivityofslackvariablesandtheircorrespondingdualvariablesat everyiteration(whichmaydrasticallyshortentheNewtonstep length),and(iii)itdoesnotwarmstartwell.

Confidentiality of large scale real-life power systems data deprivesthepowersystemscommunityofrealisticbenchmarks atleastinthefieldofOPFandpreventsresearchersfromrepor- tingOPFresultsobtainedinrealisticconditionsandtherebythe fairassessmentofexistingOPFmethodsonvariousdifferentprob- lems.Untilrecently,whenalarge2746-busmodelofthePolish powersystembecamefreelyavailable[14],thelargesttestbedfor theOPF/SCOPFprogramswasanIEEEsystemof300buses[15].

Thelarge-scale1NLPOPFproblemstackledbyIPMreportedin theliteratureinvolvepowersystemsof2098buses[16,17],2256 buses[18,19],2423buses[9],2746buses[20],2935buses[21], 3012buses[22],and3467buses[10].Otheralternativealgorithms havebeenalsousedforlarge-scaleOPF,e.g.Newtonmethodwas

1Wearbitrarilyconsiderapowersystemas“large”ifitcontainsmorethan2000 buses.

0378-7796/$seefrontmatter © 2012 Elsevier B.V. All rights reserved.

http://dx.doi.org/10.1016/j.epsr.2012.10.001

(3)

appliedtoa2436-busgrid[23],anon-interiorpointcomplemen- taritymethodwasappliedtoa2098-bussystem[16,24],amodified barrierLagrangianfunctionwasappliedtoa2111-bussystem[25], a modifiedbarrier approach wasappliedtoa 2256-bus system [26],andatrust-regionbasedaugmentedLagrangianmethodwas appliedtonetworkmodelsof2935buses[21]and3012buses[22].

Furthermore,althoughseveralcommercialNLP SCOPFpackages areavailablefromvariousvendors[27–29],andareroutinelyused bymanysystemoperators,thescientific literaturereportingon experimentsusingSCOPFsolversonlarge-scalesystemsisquite limitedexceptofthefollowingworks:[30]usesanSLPapproach foramodelof12,965buses,[29]reliesonaninterior-pointsolver fora2351-busnetwork,[31]employsaconicprogrammingfora 2383-busPolishsystem,[32]reliesontheIPMfora2746-busPolish system,[22]appliesIPMtoa3012-busgrid,[33]usesaSLPforthe 3551busesUKsystem,and[34]usesamethodcombiningIPMand conjugategradientfora15,000busEuropeansystem.

Motivatedbythefactthattheseworkspresentgenerallyonly oneexampleofa successfulapplicationofa givenalgorithmto agivenOPF/SCOPFproblem,themaincontributionofthispaper liesin thetransparent report of extensiveresults with a non- commercialIPM-basedOPFprogram,developedbytheauthorsfor researchpurposes[35],onseverelyconstrainedlarge-scaleOPFand SCOPFproblems.Thechallengesofthesecomputationsarerelated to:(i)thesizeofour8387-bussystem,whichmodelsalargepartof Europe,isroughlythreetimeslargerthanpreviouslyreportedIPM- basedNLPOPFs,(ii)thecomplexityofdata(e.g.coexistenceofmany verylonglines,stemmingfromsomenetworkequivalents,andvery shortlines),(iii)thetoughnessoftheoptimizationproblemscon- sideredcomparedtotheliterature(e.g.oftenmorethanthousand constraintsarebindingattheoptimum),and(iv)thenumberand varietyofcontrolvariables(e.g.fewthousandscontrolvariables forthecontrolofbothactiveandreactivepowersareconsidered together).Thepaperalsodiscussesinatransparentwayreliability andspeedissuesoftheIPM.

Thepaperisorganizedasfollows.Section2introducestheOPF problemswhileSection3brieflydescribestheinterior-pointalgo- rithmusedforsolvingthem.Section4providesnumericalresults obtainedwiththeIPMforseveralOPFandSCOPFproblems.Section 5concludes.

2. FormulationoftheOptimalPowerFlowproblem

For the sake of facilitating the reader’ interpretation of our resultsaswellastomakethepaperself-containedwedescribe inthissectiontheformulationoftwoclassicalOPFproblems.

2.1. Notations

Letusdenoteby:n,g,c,b,l,t,o,a,andstherespectivenum- bersof:buses(n),generators(g),loads(c),branches(b),lines(l), alltransformers(t),transformerswithcontrollabletaps(o),phase shifters(a),andshuntelements(s),respectively.

We formulate the OPF problem with complex voltages expressedinrectangularcoordinates[12,35]:

Vi=ei+jfi, i=1,...,n,

whereeiandfiareitsrealandimaginarypart,respectively,the voltagemagnitudebeing:

Vi=



e2i +fi2.

G

sk

1 : r

k

B

sk

G

k

B

sk

j

i B

k

G

sk

Fig.1.Modelofagenericbranch.

Wemodelanybranchk asasymmetricalquadruplein in serieswithanidealtransformerwithcomplexratio1/rk(seeFig.1), where:

rk=rk1+jrk2, k=1,...,b,

rk2=rk12 +rk22 , k=1,...,b.

Atransformerwithtap-changerisaparticularcasewhererk2=0 whilealineisaparticularcasewhererk1=1andrk2=0.

2.2. Objectivefunctions

Inthispaperwedealwithtwoclassicalobjectivesnamelymin- imumgenerationdeviationwithrespecttothebasecase(MD):

MD=min



g i=1

(Pgi−Pgi0)2, (1)

andminimumactivepowerlosses(MPL):

MPL=min



b k=1

GskVi2+GskVj2 rk2

+Gk



Vi2+Vj2

r2k −2rk1(eiej+fifj)+rk2(eifj−fiej) rk2



, (2)

wherePgiandPgi0 istheactivepowerandbasecaseactivepowerof generatori,Gk(respectivelyGsk)istheconductance(respectively halfshuntconductance)ofthebranchklinkingbusesiandj.

2.3. Controlvariables

Weconsiderthefollowingcontrolvariables:generatoractive power,generatorreactivepower,controllabletransformerratio, shuntreactanceandphaseshifterangle.

2.4. Equalityconstraints

Equalityconstraintsmainlyinvolvenodalactiveandreactive powerbalanceequations,which,fortheithbus(i=1,...,n),take ontheform:

Pgi−Pci−(e2i+fi2)



k∈Bi

(Gsk+Gk)

+



k∈Bi

[(eiejk+fifjk)(rk1Gk+rk2Bk)−(fiejk−eifjk)(rk2Gk−rk1Bk)]=0, (3) Qgi−Qci+(ei2+fi2)[Bsi+



k∈Bi

(Bsk+Bk)]

+



k∈Bi

[(eiejk+fifjk)(rk2Gk−rk1Bk)+(fiejk−eifjk)(rk1Gk+rk2Bk)=0, (4)

(4)

wherePgiandQgiaretheactiveandreactivepowersofthegenerator connectedatbusi,PciandQciaretheactiveandreactivedemandsof theloadconnectedatbusi,Bsiistheshuntsusceptanceconnectedat busi,Bidenotesthesetofbranchesconnectedtobusi(and∀k∈Bi, jkdenotestheotherbustowhichbranchkisconnected),GkandBk (respectivelyGskandBsk)arethelongitudinal(respectivelyshunt) conductanceandsusceptanceofthekthbranchconnectedtobusi.

Additionalequalityconstraintsmayexist,asfor examplethe settingofaphaseshifterratiotoaspecifiedreferencerk0: rk12 +rk22 −rk02 =0, k=1,...,a, (5) orthesettingofgeneratorvoltagetoaspecifiedreference:

e2i+fi2−(Viref)2=0, i=1,...,g. (6) 2.5. Inequalityconstraints

Theoperationallimitson(longitudinal)branchescurrentand voltagesmagnitudetakeontheform:

 

 

(G2

k+B2k)



Vi2+Vj2

rk2 −2rk1(eiej+fifj)+rk2(eifj−fiej) r2k



≤Ikmax,

k=1,...,b (7)

(Vimin)2≤e2i +fi2≤(Vimax)2, i=1,...,n. (8) Physicallimitsofpowersystemdevicescanbeexpressedas:

Pgimin≤Pgi≤Pgimax, i=1,...,g, (9) Qgimin≤Qgi≤Qgimax, i=1,...,g, (10) r1imin≤r1i≤r1imax, i=1,...,o, (11) Bmini ≤Bsi≤Bmaxi , i=1,...,s, (12) tanimin≤r2i

r1i ≤tanmaxi , i=1,...,a, (13) wherefortheithgeneratorPgimin,Pgimax(respectivelyQgimin,Qgimax)are itsactive(respectivelyreactive)outputlimits,fortheithcontrol- labletransformerr1iminandr1imaxareboundsonitsratio,fortheith shuntBmini andBmaxi areboundsonitssusceptance,andfortheith phaseshiftermini andmaxi areboundonitsangle.

3. Multiplecentralitycorrectionsinterior-pointalgorithm WesolvetheaboveOPFproblemsusingthemultiplecentral- itycorrections(MCC)interior-pointalgorithm[17,35,36],thatthe authorsfoundafterextensiveexperimentsovertheyearsasthe mostreliable IPM algorithm. In order to make the paper self- containedwe brieflydescribethisalgorithminthis sectionand refer the interested reader to [35] for further implementation details.

3.1. OptimalityconditionsintheIPM

TheOPFformulationsoftheprevioussectioncanbecompactly writtenasageneralnonlinearprogrammingproblem:

min f(x), (14)

s.t. g(x)=0, (15)

h(x)0, (16)

wheref(x),g(x)and h(x)areassumedtobetwicecontinuously differentiable,xisa(dim(C)+2n)-dimensionalvectorthatencom- passes both control variables (vectorof size dim(C))and state variables(realandimaginarypartofvoltageatallbuses),gisa p-dimensionalvectoroffunctionsandhisaq-dimensionalvector offunctions.

TheLagrangianassociatedwiththisNLPproblemwithintheIP frameworkcanbedefinedas:

L(y)=f(x)−



q i=1

lnsi−␭Tg(x)−␲T[h(x)s],

wheres=[s1,...,sq]Tisthevectorofslackvariables,␭and␲are thevectorsofLagrangemultipliers(alsocalleddualvariables),is apositivescalarcalledbarrierparameter,andy=[s␲␭x]Tgroups allvariables.

TheperturbedKKTfirstordernecessaryoptimalityconditions areobtainedbysettingtozerothederivativesoftheLagrangian withrespecttoallunknowns[8]:

⎢ ⎢

⎢ ⎢

sL(y)

L(y)

L(y)

xL(y)

⎥ ⎥

⎥ ⎥

=

⎢ ⎢

−e+S

−h(x)+s

−g(x)

f(x)−Jg(x)T␭−Jh(x)T

⎥ ⎥

=0, (17)

whereSisadiagonalmatrixofslackvariables,e=[1,...,1]T,∇f(x) isthegradientof f,Jg(x)istheJacobianof g(x)andJh(x) isthe Jacobianofh(x).Notethatinordertofacilitatethepresentation wehaverewrittenthecomplementarityslacknessconstraintsas

e+S=0butaproperimplementationshouldrelyonthecon- straints/s+=0soastopreservethesymmetryoftheHessian.

Thepureprimal-dualIPalgorithmconsistsinsolvingiteratively thelinearizedKKTconditionsfortheNewtondirectionykwhile decreasingthebarrierparameterkgraduallytozeroasiterations progress:

H(yk)

⎢ ⎢

⎢ ⎢

sk

␲k

␭k

xk

⎥ ⎥

⎥ ⎥

=

⎢ ⎢

⎢ ⎢

keSkk h(xk)−sk

g(xk)

−∇f(xk)+Jg(xk)Tk+Jh(xk)Tk

⎥ ⎥

⎥ ⎥

(18)

where H(yk) is the Hessian matrix (of second derivatives) (∂2L(yk)/∂y2).Wedenotebydim(KKT)thesizeofthisHessian.

3.2. TheMCCalgorithm

WebrieflyoutlinetheMCCalgorithmtosolvetheKKToptimal- ityconditions(17):

I. Initialization.

Settheiterationcountk=0.Chose0>0.Initializey0,taking carethatslackvariablesandtheircorrespondingdualvariables arestrictlypositive(s0,␲0)>0.

II.Thepredictorstep.

(a) Solve the system (18) for the affine-scaling direction, obtainedbyneglectinginitsright-handside:

H(yk)

⎢ ⎢

⎢ ⎢

skaf

␲kaf

␭kaf

xkaf

⎥ ⎥

⎥ ⎥

=

⎢ ⎢

⎢ ⎢

−Skk h(xk)−sk

g(xk)

−∇f(xk)+Jg(xk)Tk+Jh(xk)Tk

⎥ ⎥

⎥ ⎥

(5)

(b) Computetheaffinecomplementaritygapkaf:

kaf=(skkafskaf)T(␲kkaf␲kaf)

where˛kaf∈(0,1]isthesteplengthwhichwouldbetaken alongtheaffinescalingdirectionifthelatterwasused(19).

(c)Estimatethebarrierparameterforthenextiteration:

kaf=min

⎧ ⎨



kaf

k



2

,0.2

⎫ ⎬

afk q

wherek=(sk)Tkdenotesthecomplementaritygapatthe currentiterate.

III Thecorrectorstep.

(a) Compute a trial2 point ˜yk=yk+˛˜kykaf, where ˛˜k= min(˛kaf˛,1)withthedesiredimprovementonthestep lengthı˛=0.2.

(b) Computethecomplementarityproductsatthetrialpoint

˜vk= ˜Skk.

(c)Identifycomponentsof ˜vkthatdonotbelongtotheinterval [ˇminkafmaxkaf], calledoutliercomplementarity prod- ucts,whereˇmin=0.1andˇmax=10.

(d)Because the corrector step effort focuses on correcting theoutliersonlyinordertoimprovethecentralityofthe nextiterate,sometargetcomplementaryproducts(˜vk)tare defined:

(vki)t=

⎧ ⎪

⎪ ⎨

⎪ ⎪

ˇminkaf, if ˜vikminkaf ˇmaxkaf, ifv˜ikmaxkaf v˜ik, otherwise

(e)Thecorrectordirectionykcoisobtainedasthesolutionof thefollowinglinearsystem:

H(yk)

⎢ ⎢

⎢ ⎢

skco

␲kco

␭kco

xkco

⎥ ⎥

⎥ ⎥

=

⎢ ⎢

⎢ ⎣

(vk)t− ˜vk 0 0 0

⎥ ⎥

⎥ ⎦

wherethenonzerocomponentsoftheright-hand-sidecor- respondtotheoutliercomplementarityproductsonly.

(f)Update3thenewsearchdirection:

yk=ykaf+ykco (g)Choiceofthesteplength.

Determinethemaximumsteplength˛kp∈(0,1](respec- tively˛kd∈(0,1])intheprimal(respectivelydual)variables spacealongtheNewtondirectionyk suchthat sk+1>0 (respectively␲k+1>0):

˛kp=min



1,min

sk i<0

−ski

ski



˛kd=min



1, min

k i<0

−ik

ik



(19) where∈(0,1)isasafetyfactoraimingtoensurestrict positivenessofslackvariablesandtheircorrespondingdual variables.Atypicalvalueofthesafetyfactoris=0.99995.

(h)Updatesolution:

sk+1=skkpskk+1=␲kkd␲k xk+1=xkkpxkk+1=␭kkd␭k

2 Notethatatthispointsomeslackvariablesand/ortheircorrespondingdual variablesmayviolatethestrictpositivityconditions(sk,k)>0.

3 Thecorrectorstepcanbeappliedseveraltimes.Insuchacase,thecurrent directionykbecomesthepredictorforanewcorrector.

Table1

Testsystemsummary.

n g c b l t o a s

8387 1865 4669 14,561 12,474 2087 589 84 178

(i)Repeatthecorrectorstepapre-definednumberoftimesas longastheimprovementinthesteplengthissatisfactory butinferiorto1.

(IV) Checkconvergence.

A(locally)optimalsolutionisfoundandtheoptimization processterminateswhen:primalfeasibility,scaleddualfea- sibility,scaledcomplementarity gapandobjective function variationfromaniteration tothenextfallbelowsometol- erances[9,10,12]:

max{max

i {−hi(xk)},||g(xk)||}≤ 1 (20)

||∇f(xk)−Jg(xk)T␭−Jh(xk)Tk||

1+||xk||2+||␭k||2+||␲k||2

1 (21)

k

1+||xk||22 (22)

|f(xk)−f(xk−1)|

1+|f(xk)| ≤ 2 (23)

whereusually 1=10−4and 2=10−6.

Ifconvergencewasnotachievedsetk←k+1andgotostep II.

4. Numericalresults

4.1. Descriptionofthepowersystemmodel

Weusea8387-busmodifiedmodel4oftheinterconnectedEHV EuropeanpowersystemwhichspansfromPortugalandSpainto Ukraine, Russia and Greece. Notice that in this model the real parametersoftheindividualpowersystemcomponents(e.g.lines, transformers,etc.),thenetworktopology,aswellasthelimitson generators active/reactivepowers,transformers ratioandangle, voltages,andbranchcurrentshavebeenbiased.Nevertheless,this modelisrepresentativefortheEuropeaninterconnectioninterms ofsystemsizeandcomplexity.

Inordertoassesstherobustnessofourtoolwehavechosenvery tightoperationallimitsandphysicalboundsovercontrols.Asacon- sequencethebasecaseisquiteconstrained,e.g.manygenerators havenarrowphysicalactive/reactivepowerlimits,manyvoltage limitsareverytight,theanglerangeofseveralphaseshiftersisvery small,6linesareloadedatmorethan90%,etc.Thesedatacontain manyveryshortlines(e.g.98lineshavetheirreactancelowerthan 0.0002pu)andmanyartificiallines,stemmingfromsomenetwork equivalents,withverylargeimpedancevalues,whichleadtorel- ativelyill-conditionedproblems.Forinstanceforthisdatasetthe ratiobetweenthemaximumimpedanceandminimumimpedance overallbranchesisaround408,854,whilefortheIEEE118sys- tem[15](respectivelythePoland2746-busmodel[14])thisratio isaround51(respectively15,516).Inaddition,thisdatasetcon- tains743brancheswithalargerratiothanthemaximumratioof thePoland2746-busmodel.

Asummaryofthecharacteristicsofthistestsystemsisgivenin Table1.

4Morecomprehensivedatasetsofthissystemareavailableuponrequestfrom theprojectPEGASEweb-page:www.fp7-pegase.eu.

(6)

Table2

Problemsdefinition.

Problem Basecase Obj. Controlvariables Inequalityconstraints

Pg Pgs Vg r1 Bs a Pg Pgs Qg r1 Bs a V I

P1 A MPL × × × ×

P2 A MPL × × × × × × × × ×

P3 A MPL × × × × × × × × × ×

P4 A MD × × × ×

P5 A MD × × × × ×

P6 A MD × × × × × × × × ×

P7 A MD × × × × × × × × × × ×

P8 A MD × × × × × × × × × ×

P9 A MD × × × × × × × × × × × ×

P10 B MD × × × × × ×

P11 B MD × × × × × × × × × ×

P12 B MD × × × × × × × × × × × ×

4.2. Simulationassumptions

Weconsiderin oursimulationstwobasecasesdenotedasA andB.InthebasecaseAsomevoltagelimitsarenotmet.Thebase caseBisthesameasAexceptthattheflowlimitofalinehasbeen decreasedtocreateathermalcongestion.

Unlessotherwisespecified,weusethedefaultsettingsofthe MCCalgorithm[35]inallcomputationstoenableafairassessment oftheperformancesofthealgorithm.

AlltestshavebeenperformedonaPCof2.8-GHzand4-GBRAM, usingtheIPM-basedNLPOPFprogram,developedforresearchpur- posesbytheauthorsinC++withintheCygwinenvironment[35].

4.3. OPFproblemsdefinition

InordertotesttherobustnessandefficiencyoftheIPM,several OPFproblemshavebeensolvedinvolvingdifferentcombinations ofobjectives,controlsandconstraints,asshowninTable2.Inthis tablePg,Psg,Vg,Qg,r1,Bs,a,I,andVrefertogeneratoractivepower, slackgeneratoractivepower,generatorterminalvoltage,generator reactivepower,controllabletransformerratio,shuntsusceptance, phaseshifterangle,branchcurrent,andbusvoltagemagnitude, respectively,whilesymbol“×”meansthatthevariable/constraint isconsideredintheOPF.

In allproblems the equality constraintsare the power-flow equationsandsometimesphaseshifterangles(e.g.inP7,P9,and P12)andgeneratorsvoltages(e.g.inP4toP7).

4.4. ObservationsabouttheOPFresults

Table 3 provides for each problem the size dim(C) of the setofcontrol variables,thesize dim(KKT)ofthe fullsystemof KKToptimalityconstraints(17),andthenumberofiterationsto

Table3

Problemssizeandconvergencedetails.

Problem dim(C) dim(KKT) Iter. Time(s)

P1 1825 76,221 43 16.0

P2 2592 80,056 45 16.8

P3 2592 105,004 65 47.4

P4 1250 72,204 9 5.0

P5 1250 105,752 10 8.1

P6 2017 109,587 14 10.7

P7 2185 110,175 13 10.0

P8 4413 114,109 9 8.3

P9 4581 114,697 9 7.2

P10 3646 110,274 17 12.1

P11 4413 114,109 14 10.6

P12 4581 114,697 16 50.8

Table4

NumberandtypeofactiveconstraintsforproblemP3.

Activeconstraints Total

Qg I V r1 Bs

330 18 1064 83 63 1558

convergenceaswellasthecorrespondingCPUtimes5(inseconds).

Furthermore,ingeneralthelargerthenumberofactiveconstraints thelargerthenumberofiterationstoconvergenceandhencethe computationaltime.Noticethattheconvergenceisachievedwithin areasonableCPUtimeforallproblemsinspiteofthelargesizeof thesystemandallthecomputationalchallengesmentionedpre- viously.Theaveragecomputationaleffortperiterationgenerally increaseswiththeproblemsizewhiletheoverallcomputingtime seemstoberathermoreinfluencedbytheproblemsizethanbythe numberofcontrolvariables.

Table4showsthenumberandthetypeofactiveconstraintsat theoptimumofproblemP3.

Theverylargenumberof1558activeconstraintsconfirmsthe extremedifficultyoftheproblem.Thealgorithm’sabilitytoeffec- tivelycoordinate 18activelinecurrentconstraints(and3other linesloadedabove96%)togetherwith1064Vlimitsisremarkable.

NoticethatmostNLPOPFtestscenariosoftheliterature,andin particularforIPM-basedOPFs,rarelyhavemorethanafewactive line-currentconstraintsanda fewhundredofotheractivecon- straints(seehowever[30],aworkonSLP-basedSCOPF,where23 linecurrentconstraintsareactive).

4.5. OPFfailureonaverytoughproblem

WenowfocusonproblemP8butminimizetheoverallactive powergenerationandhencereplacethequadraticobjective(1)by thelinearone:

L=min



g i=1

Pgi. (24)

ForthisproblemtheMCCalgorithmgetsstuckonaninfeasible non-optimalpoint.Furthermore,theconvergenceisnotrestored even after trying few classical heuristic techniques (e.g. differ- entinitialvaluesfor thebarrierparameter,smallervalues of thesafetyparameter,anddifferentinitializationsofvariables) [16,40].However, by using anotherIPM algorithm, namelythe predictor–corrector(thatwegenerallyfoundslightlylessreliable thantheMCCalgorithm)alocallyoptimalsolutionisfound.Table5

5CPUtimeconcernstheoptimizationprocessonly.

(7)

Table5

NumberandtypeofactiveconstraintsforthemodifiedproblemP8.

Activeconstraints Total

Pg Qg I V r1 Bs

828 379 76 510 58 80 1931

Table6

SCOPFstatisticsforproblemP2.

Numbercontingencies dim(C) dim(KKT) Iter. Time(s)

0 2592 80,056 45 16.8

1 4417 158,101 53 78.5

2 6242 236,146 55 155.0

3 8067 314,191 75 389.0

Table7

SCOPFstatisticsforproblemP8.

Numbercontingencies dim(C) dim(KKT) Iter. Time(s)

0 4413 114,109 9 8.3

1 6238 217,100 10 30.2

2 8063 320,091 15 93.0

3 9888 423,082 11 140.9

4 11,713 526,673 10 204.3

providesthenumberandthetypeofactiveconstraintsatthisopti- mum.Thechoiceofthislinearobjectiveleads,asexpected,toa highnumberofactivepowergeneratorconstraintsactiveatthe optimum.Theverylargetotalnumberof1931activeconstraints, aswellasthelargenumberof76activebranchcurrentconstraints (plus11linesloadedabove95%attheoptimum),indicateagainthe toughnessofthistestscenarioascomparedwiththosefromthe literature.

4.6. SCOPFsolutions

InordertofurtherassesshowtheIPMscaleswithproblemsize weconsidertheproblems P2andP8solvedbya SCOPFinpre- ventiveonlymode[2].Thelatterconsistsinduplicatingbasecase constraints(3)–(13)foreachcontingencyincludedintheSCOPF andassumesthatcontrolvariableshavethesamevalueinboth basecaseandcontingencystates.ThisSCOPFisaugmentedwith onecontingencyatthetime.

Tables 6and 7provide for eachproblemthe sizeof theset ofcontrolvariables,thesizeofthefullsystemofKKToptimality constraints,thenumberofiterationstoconvergenceandtheCPU times.Onecanobservethatforbothproblemsthecomputational effortincreasesmuchmorethanlinearlywiththesizebut still remainsreasonablegiventheproblemsize.Furthermore,asinpre- viousexamples,weobservethatthesizeoftheproblemhasamore significantimpactonthecomputationaleffortthanthenumberof controlvariables.

4.7. Comparisonwithavailablesolvers

Inthis sectionweperforma comparison withthetwo most efficientalternative availablesolvers for largescalesystems on Matpower Version4.1runningunder Matlab7.13[22],namely theMatlabInteriorPointSolver(MIPS)andthePrimalDualInte- riorPointMethod(PDIPM).AlthoughMatpowerpossessesseveral interestinglessconventionalmodelfeatures,thecurrentversion doesnotallowmodellinganyamongourtwelvemoreclassicalOPF problemsintermsofobjectivefunctionandcontrolvariables.In ordertoenableanasfairaspossiblecomparison,weconsideronly theOPFobjectivefunctionofminimumgenerationcost/deviation, generators active/reactive powers as control variables, and

Table8

OPFcomparisonwithMatpower.

Basecase Obj. Oursolver it/time

MIPS it/time

PDIPM it/time

A L 97/71.9 66/31.8 76/37.8

A-PST L 346/252 Failed Failed

A Q 13/10.7 47/22.1 Failed

A-PST Q 20/15.7 32/17.8 42/20.6

constraintsonbranchcurrents,voltagemagnitudeand physical limitsongenerators’active/reactivepowers.Stillthereremaintwo slightmodellingdifferencesconcerningthetransformertransver- sal susceptanceand thehandlingof current constraints,asour solverimplementsonlyconstraintsonlongitudinalbranchcurrents (7),whereasMatpowersolversimplementsuchconstraintsatboth endsofeachbranch.Furthermore,PDIPMimplementsonlyMVA flowconstraints.However,asthemaincomputationalburdenin IPM isthefactorizationofa linearsystemofequations[12],we noticethatasMatpowersolversrelyonthe“reducedKKTsystem”

[10,21],thenumberofinequalityconstraintsdoesnotaffectthe sizeofthissystem,contrarytoourimplementationwhichusesthe fullKKTsystem(17).

Table8presentstheresultsofourexperimentsforfourfeasi- bleOPFproblems,where“L”denotesthelinearobjective6function (24),“Q”denotesthequadraticobjectivefunctionMD(1),andthe basecaseA-PSThasbeenobtainedfromthebasecaseAbysetting tozerotheangleofall84phaseshiftersandusingasstartingpoint theloadflowsolutionofcaseA.

TheresultsshowthatbothMIPSandPDIPMsolvershavealower computationaleffortperiterationthanoursolver,whichiscer- tainlyduetotheuseofthereducedKKTsystemandpossiblythe libraryusedtosolvethissystem.However,oursolverconverges inreasonabletimeandisbothfasterandneedsalowernumber ofiterationsforthequadraticobjective.Theresultsalsoindicate that,similarlytothefailureofoursolverfortheproblemdescribed inSection4.5,otherexcellentsolversmayalsooccasionallyfail, especiallyonverytoughoptimizationproblemsasincaseA-PST.

Inthelattercasetheverylargenumberofiterationsofoursolverto convergenceisnotsatisfactoryandcouldbepracticallyconsidered asacaseoffailure,e.g.ifamaximumrunningtimewasrequired.

Wepresumethatbyrunningagainthesolverwithdifferentsetsof parameters[16]theconvergenceoffailedcasescouldberestored.

Acomprehensivecomparisonconcerningtherelativereliabilityof thesesolversisoutofthescopeofthepaper,asitshouldconsider manydifferentOPFproblemsundervariedoperatingconditions.

WeonlyunderlinetheneedtoimprovethereliabilityofIPMcodes forverytoughlargescaleOPFproblems.

Takingintoaccounttheslightmodelingdifferencesandthelarge numberof84branchcurrentbindingconstraints,weassessedthat inallcasesMIPSandoursolverconvergedpracticallytothesame solution.

4.8. SCOPFsolutionsona3012-bussystem

WeconsideramodelofthePolandpowersystemwhichOPF dataareavailableonMatpowerweb-site[22].Thesystemcom- prises3012buses,3371lines,201transformers,and298generators (obtained after aggregating coexisting generators at the same bus).

Weconsidertheproblemofminimizinggenerationcostsolved byaSCOPFincorrectivemode[37].Weuse292generatorsactive

6NoticethatforthisobjectivefunctionoursolverfailedforamorecomplexOPF problemintermsofcontrolsandconstraints(seeSection4.5).

(8)

Table9

SCOPFresultsforthePolandsystem.

Numbercontingencies dim(C) dim(KKT) Iter. Time(s)

0 480 33,640 35 6.7

1 960 68,446 38 24.4

2 1440 103,252 37 60.1

3 1920 138,058 42 122.3

4 2400 172,864 41 177.2

5 2880 207,670 41 68.7

6 3360 242,476 39 76.0

7 3840 277,282 40 88.2

8 4320 312,088 46 113.8

9 4800 346,894 47 129.0

10 5280 381,700 47 144.9

11 5760 416,506 47 158.1

12 6540 451,312 48 176.3

powerand188generatorsreactivepowerascontrolvariables,and constraintsonbranchcurrents,voltages,physicallimitsongenera- tors’active/reactivepowers,andactivepowerre-dispatchof292 generatorsascorrectiveactions.

Table9showstheresultsobtainedwithoursolverforincreasing numbersofcontingenciesincludedintheSCOPF.Exceptmaybefor thecaseswherethenumberofcontingenciesincludedintheSCOPF rangesfrom1to4,thealgorithmscalesverywell.

TheseresultsshowthattheSCOPFsolutionforthisalreadyquite largesystemisreasonablyfastandcanbeenvisagedevencloseto real-time.Furthermore,comparingSCOPFresultsofTables7and9 for problems of close sizes one can observe that, the average computationaleffortperiterationforthe3012-bussystemiscon- siderablyfasterthanforthe8387-bussystem,whichweexplain bythemuchworsenumericalconditioningofthelattersystem,as alreadydiscussedinSection4.1.

5. Discussion,conclusionsandperspectives

ThispaperhaspresentedextensivenumericalresultswithIPM forlargescaleOPF/SCOPFproblems,withthegoalofadvancingthe stateoftheknowledgeaboutthismethodintermsofrobustness andscalability.

Ourresultsshowthatnowadays,evenwithoutusingthemost powerfulcomputersavailableonthemarket,itiscertainlyfeasi- bletorunanNLPOPFforalargesysteminreal-timeprovidedthat areliablesolveris used.Furthermore,weprovethatourSCOPF codeis practicalona 3012-bus system.Ontheother hand,for thepoorlyconditioned8387-busgrid,computermemorylimita- tionspreventedusfromincludingmorethanfewcontingencies (usingthefullnetwork model)intotheSCOPF,andthereforeto assesshowthecomputationaleffortfurtherscaleswiththeprob- lemsize.However,even assumingsufficientcomputermemory, asonecouldforesee thatforsuchasystemfewtensof contin- genciesmight bebindingattheoptimum,oursolvercouldnot complywithclosetoreal-timerequirements,whileitcouldpos- siblystill beused in day-ahead operational planning.For such large-scalesystems,inordertorendertheSCOPFtractableinreal- timeapplicationsonewillhavetoadoptapproximatemodelsfor post-contingencystates[27,34]orresorttoBendersdecomposition [37].

AsthefactorizationoftheHessianmatrix isbyfarthemost expensivecomputationaltaskofaninterior-pointalgorithmitera- tionthechoiceofasuitablelibraryforthesolutionofasymmetric, verysparse,and veryill-conditionedsystemoflinearequations isparamount.Inourimplementationweusedthesparseobject- oriented library SPOOLES [38] which has several options, and wecouldexperimentthestrongdependenceofperformancevs.

robustnesstradeoffsontheparticularoptionschosen.

Furthermore,inourimplementation,forthesakeoffacilitat- ing theprogramming effort,wesolve ateach iteration the full systemofKKToptimalityconditionswhereas mostauthorsrec- ommendtheuseofthe“reducedsystem”only[10],whichmay break-downthesizeofthesystemuptoaround30%byappropriate eliminationofthecomplementarityslacknessandinequalitycon- straints.Althoughthecomplementarityslacknessconstraintsare trivialequationsandthereforeshouldnotposeadditionalprob- lemstoasmartlibrary,a furtherdecreaseofthecomputational timeshouldbeexpectedifonereliesonthe“reducedsystem”only.

Our study also shows that the IPM algorithms proposedin thepowersystemsliteraturemayleadtoconvergenceproblems onveryhard optimizationproblems.Themaincauseofconver- genceproblems oftheIPMis thattheiterationsbecomesstuck atanon-optimalpointifoneapproachestooearlythefeasibility boundaryandthereforetheslackvariablesprematurelygoto0, aphenomenoncalledjamming[39].Fortunately,therearesome well-knownremediestorestoretheconvergenceofanIPM(e.g.

changingtheinitialvalueofsomeparametersandespeciallythe barrier parameter, using a less aggressive decrease of the bar- rierparameter,usinganalternativeinterior-pointalgorithm,etc.) [16,40],buttheytakeadditionalcomputationaleffortwithoutguar- anteeingsolutionrestoration.

WeenvisagetoimprovethereliabilityofourIPMimplementa- tionbyfocusingonthreeaspects:theuseofanti-jammingremedies [39],regularization(ortoensurethesmoothnessofthesolution byaugmentingtheproblemwithpenaltyfunctionstobetterdeal withanill-posedHessianmatrix)[41],anduseofmeritfunctionsto ensurethatjointprogressismadebothtowardsalocalminimizer andtowardsfeasibility(thisprogressisachievedbyshorteningthe step-lengthalongtheNewtonsearchdirection)[42].

Asensiblesolutiontomitigatereliabilityissues,beyondusing appropriate librariesfor thesolutionof linearsystemsof equa- tions,wouldbetoembedarobustandfastgenericNLPsolverin theOPF[34].Nowadaysthereexistindeedseveralmaturegeneric NLPsolvers(e.g.KNITRO,IPOPT,LOQO,CONOPT,SNOPT,MINOS, etc.)thatprovedexcellentperformancesongenericproblemsas wellasforsomepowersystemproblemsmodeledinAMPL[43]or GAMS[44].

UsingOPF/SCOPFinreal-timeinvolvesamongothersthesuc- cessivesolutionsofcloselyrelatedNLPproblems.TheIPMdoesnot naturallywarmstartwellandhencecannotdirectlytakeadvan- tageofthesolutionsgottenattheprevioustime-steps.Although thecurrentspeedofIPMissatisfactorysince,aswithanyNewton- basedmethodthenumberofiterationsgenerallyscalesacceptably withtheproblemsize,anyimprovementofitswarmstart abil- ity(or,moregenerally,ofitslearningability)iswelcome.Welook forwardtoassessonSCOPFproblemstheimprovementsreported forgenericNLPs(seee.g.[41]).Furthermore,recentresults[45]

showthatmethodssupposedtowarmstartwell(e.g.SQP)mayfail onlargescaleproblemswithaverylargenumberofbindingcon- straintsattheoptimum:inparticular,itwasfoundthatforNLPOPFs withequilibriumconstraints(inwhichequilibriumconstraintsare modeledbyRNSfunctions)thewarmstartofKNITRO(withSQP option)failedwhiletheSQPsolverSNOPTdidnotperformfaster thanthecoldstartofKNITRO(withIPMoption);also,theSQPsolver SNOPTfailedoncoldstartonmostproblems,whileKNITRO(with IPMoption)wasfoundthemostreliableamongthesolverstested [45].

Finally,thepresentpaperfocusedonNLPaspectsoftheOPF problem,buttheimpactontheoverallreliabilityandspeedofthe OPFprocessofotherkeyfeatures(e.g.handlingofdiscretevari- ables,useoflimitednumbersofcontrolactions,etc.)remainalso tobecarefullyassessed[4,5,46,47].Inaddition,improvedwaysto analyzeandvisualizetheresultsoflarge-scaleOPF,e.g.incaseof gridcongestion,shouldbedevised[30].

Références

Documents relatifs

III Interaction between a front and an idealised mountain massive Inertio Gravity Waves Re-emission (Lott 2003). 2D explicit simulations of Gravity Waves breaking with

In the rest of this paper we will use the acronym ISCOPF-NC (Iterative Security-Constrained Optimal Power Flow with Network Compression) to denote the proposed

Figure 3: CPU time required for solving the elastic eigenvalue problem using standard method and double projection method as a function of the dofs number defined in Table 1.. Figure

For this reason, the literature is replete with metaheuristic algorithms, cooperative search methods and hybrid approaches, which stand between heuristics and exact

 ADMM based distributed optimization technique for AC-OPF by considering fuel cost as objective functions; active and reactive power balance as equality constraints; and boundary

Keywords: large berth allocation problem, long term horizon, algorithm portfolio, quality run-time trade-off, real traffic

In the following we present a general method development technique which entails iso- lating the RSN from the original model, analyzing the isolated subsystem and replacing

In this paper we present an approach, called Dual Approximate Dynamic Pro- gramming (DADP), to solve large-scale stochastic optimal control problems in a price decomposition