第29卷 第4期2003年7月自 动 化 学 报
ACTAAUTOMATICASINICAVol129,No14July,2003
MultipleModelPredictiveControlforMIMOSystems1)
LINing LIShao2Yuan1 XIYu2Geng(InstituteofAutomation,ShanghaiJiaotongUniversity,Shanghai 200030)1(E2mail:syli@sjtu.edu.cn)
Abstract Amulti2model2basedpredictivecontrol(MMPC)strategydealingwithnonlinearmodel2basedpredictivecontrol(NMPC)forMIMOsystemsisdevelopedinthispaper.Firstlyamulti2modeli2dentificationmethodisgiven.Usingfuzzysatisfactoryclusteringalgorithmpresentedinthispaper,thecomplexnonlinearsystemcanbequicklydividedintomultiplefuzzyparts.Aglobalmodelcanbeob2tainedbysometransformationoftheobtainedmultiplelinearmodels.AnMMPCalgorithmisthereforedesignedfortheglobalMIMOsystemswithsystemperformanceanalysis.TakingapHneutralizationcontrolsystemassimulationexample,thesimulationresultsverifytheeffectivenessofMMPConcom2plexnonlinearsystems.
Keywords MIMOsystems,multi2model,model2basedpredictivecontrol(MPC),fuzzysatisfactoryclustering,pHneutralizationprocess
1)SupportedbyNationalNaturalScienceFoundationofP.R.China(69934020and60074004)ReceivedMay8,2002;inrevisedformAugust28,2002收稿日期 2002205208;收修改稿日期 20022082281 Introduction
RecentlyModelPredictiveControl(MPC)hasbecomeanattractiveresearchfieldinauto2
maticcontrolforitsadvantagesoverconventionaltechniquesandsuccessfulapplicationsinin2
dustry.MPCalgorithmswereoriginallydevelopedforlinearprocesses,butthebasicideacan
betransferredtononlinearsystems[1,2].Unfortunately,twomajorissueslimititspossibleapplica2
tiontononlinearsystems.Thefirstistheirassumptionofamodelthathastobequiteaccurate;
however,themodelingofindustrialsystemsoftenpresentsproblemsofnonlinearity,strong
coupling,uncertainty,andevenwideoperatingrange,asatisfiedmodelisalwaysdifficultto
obtain.Thesecondisthatanonlinearnon2convexoptimizationproblemmustbesolvedforeach
samplingperiodwithalgorithmswhichareusuallytooslowforreal2timecontrolduetoalarge
amountofcomputation.Thefactshaveforcedthecontrolcommunitytostudysimplificationsof
thisgeneralapproachinordertoremovethesedrawbacks.Usually,thenonlinearmodelislin2
earizediterativelyineachcontrolintervaltosolvetheaboveproblems.Thispaperwillpresenta
newsolutionbasedonmulti2modelapproach.
Multi2modelapproachesareverypropertocontrolindustrialprocesses,especiallychemical
processesfortheirinherentlynonlinearityandlargesetpointchangesorloaddisturbances.
Basedondivide2and2conquerstrategy,multi2modelapproachesdeveloplocallinearmodelsor
controllerscorrespondingtotypicaloperatingregimes,thenfittheglobalsystemthroughcer2
tainintegrationoflocalmodelsorcontrollers.Actually,applyingmulti2modelcontroltononlin2
earortime2varyingsystemshasalonghistory.However,multi2modelapproachforMIMOsys2
temsseldomappearsinliteratures.
Inthispaper,aMulti2ModelPredictiveControl(MMPC)ispresentedtodealwithNMPC
problemofMIMOsystems.Firstly,amulti2modelmodelingmethodusingT2Sstructuremodel
isintroduced.Usingfuzzysatisfactoryclusteringalgorithmgiveninthispaper,acomplexnon2
linearsystemcanbequicklydividedintolocalsystems,andtheglobalsystemcanbedescribed
byintegrationofthelocallinearmodels.Secondly,mergingtheobtainedmultiplelinearmodels
withMIMOGeneralizedPredictiveControl(GPC),anovelMMPCalgorithmisdesignedfor
theglobalsystem.Asamajorbenefitofthemulti2modelstrategy,linearpredictivecontrollerscanbeused.Then,thepapertriestouseMMPCtoregulateatypicalcomplexnonlinearpro2
cess:anMIMOpHneutralizationsystem.
2 Multi2modelidentificationbasedfuzzysatisfactoryclustering
Itiswellknownthatclusteringalgorithmsaimtodivideadatasetintoseveralsub2sets.
Therefore,theycanbenaturallyusedforsystemdivisioninmulti2modelapproach.Itis
thoughtthatclusternumbercintheclusteringalgorithmcorrespondstothenumberoflocal
modelsinthemulti2modelapproach.Therefore,todivideaglobalsysteminasatisfactoryway
equalstolookforaproperclusternumber.However,formanykindsofclusteringmethods,in2
cludingGKalgorithm[3],clusteringnumbercisalwaysneededinadvance,whichhampers
clusteringalgorithmstobeused.Thispaperaimstosolvecomplexsystemcontrolproblem.It
isknownthat,forcontrolproblems,themodelingprecisionisthecontraryofmodelnumbers.
Togiveattentiontothem,herewepresentasatisfactoryclusteringalgorithmbasedonGK.
Simplyspeaking,lettheclusteringmethodstartwithc=2(c∈[2,c3]),wherec3isthe
satisfactoryclusternumber.Thendeterminewhetheranewclustercentershouldbeincreased
ornot.Iftheclusteringresultisnotsatisfiedyet,fromthegivendataset,findoutasample
mostdifferentfromtheexistingclustercentersv1~vcasnewcentervc+1.Startwithv1~
vc+1asinitialclustercenters,andcomputethenewNOT2randompartitionmatrixU.Then
repeatGKalgorithmtodividethesetintoc+1parts.Dotheabovestepsagainuntiltheresultis
satisfactory.
ConsideraMISOsystem,whosedatasetZiscomposedofsysteminput2outputdata.De2
fineadatapairaszj=[φj,yj]T∈Rd+1,j=1,…,N,whereφjiscalledasregressionvector
orgeneralizedinputvector,yjissystemoutput.SupposeZisdividedintocclusters.Thatis,
thesystemcanbecomposedofclocalmodels.Theglobalmodelisconstructedbyfuzzyinter2
polationoftheselocalmodels.Multi2modelidentificationbasedonsatisfactoryclustering(Algo2
rithmⅠ)canbedescribedasfollows:
Step1.Setinitialclusternumberc=2.
Step2.UsingGKalgorithm,byinitialpartitionmatrixU0,divideZintocparts{Z1,
Z2,…,Zc}andobtainfuzzypartitionmatrixU=[μi,j]c×N.
Step3.Foreachsubset,identifyitsconsequentparametersusingstable2stateKalmanfil2
termethod[4].Thelocalmodelisthendescribedas
Riifφj,yj∈Zi then yi=pi0+pi1φj1+…+pidφjd i=1,…,c(1)
Step4.Computethesystemoutputy^correspondingtoinputzj
y^=∑c
i=1μijyi/∑c
i=1μij(2)
Topredicttheoutputy~ofanewinputφ~,returnGKalgorithmandusethefollowingequation
tocalculateμ~icorrespondingtoithrule[5],
μ~i(φ~)=1∑c
j=1DAxi(φ~,vxi)/DAxi(φ~,vxj)2/(m-1)(3)wherevxidenotestheprojectionoftheithclustercenterviontothegeneralizedinputspace;
DAxiφ~,vximeasuresthedistanceofthenewinputvectorfromtheprojectionofthecluster
centervxi;m>1isaparameterthatcontrolsfuzzinessofclusters.Thenthepredictedoutputy~
canbecalculatedby(2).
Step5.UseS=RMSEtoevaluatemodelingresults.IfS≤STHissatisfied,whereSTH
isgiventhreshold,modelingisover.Otherwise,gotoStep6.715No.4LINingetal.:MultipleModelPredictiveControlforMIMOSystems