搜索
1MultivariableandClosed-LoopIdentificationforModelPredictiveControlYucaiZhu*FacultyofElectricalEngineering,EindhovenUniversityofTechnologyP.O.Box513,5600MBEindhoven,TheNetherlandsPhone+31.40.2473246,y.zhu@tue.nlFirminButoyiDowBeneluxN.V.P.O.Box48,4530AATerneuzen,TheNetherlands*Alsowith:Tai-JiControlHageheldlaan62,NL-5641GPEindhoven,TheNetherlandsPhone+31.40.2817192,fax+31.40.2813197,y.zhu@tip.nlAbstract:Inthisworkwewillstudymultivariableandclosed-loopidentificationoflargescaleindustrialprocessesforuseinmodelpredictivecontrol(MPC).Theadvantagesofclosed-loopidentificationwillbediscussedandrelatedproblemsofidentificationareoutlined.Then,asymptoticmethod(ASYM)ofidentificationisintroduced.Thefourproblems,testsignaldesignforcontrol,modelorder/structureselection,parameterestimationandmodelvalidation,aresolvedinasystematicmanner.Themethodprovidesaccurateinput/outputmodelandunmeasureddisturbancemodel,modelerrorsarequantifiedbyanuppererrorboundmatrixthatcanbeusedformodelvalidationandtestredesign.Todemonstratetheuseofthemethod,theidentificationofadeethanizerforuseinMPCwillbepresented.1IntroductionModelpredictivecontrol(MPC)technologyhasbeenwidelyappliedinrefineryandpetrochemicalindustriesandisbeginningtoattractinterestfromotherprocessindustries.DynamicmodelsplayacentralroleinMPCtechnology.Generallyidentifiedblack-boxmodelsareusedforMPCcontrollers.Industrialprojectexperiencehasshownthatthemostdifficultandtime-consumingworkinanMPCprojectismodellingandidentification.Typicalidentificationtestwilltakeseveralweeks.Theselongtestsmakeproductionplanningdifficult.Thetestsaredonemanually,whichdictatesextremelyhighcommitmentoftheengineersandoperatorsandthequalityofcollecteddatadependsheavilyonthetechnicalcompetenceandexperienceofthecontrolengineerandtheoperator.Thedatacollectedfromtheselongtestsmaynotbegoodenoughformodelidentificationduetonotenoughexcitationandprocessnonlinearity.Afterthetest,duetothelackofsystematicidentificationapproach,itcantakeanotherfewweekstoanalysethedataandtoidentifythemodels.Somepeopleevenbelievethatthelongtimeneededfortestsanddataanalysisisthepriceonemustpayforagoodmodel.Thehighcostofidentificationofcurrentindustrialapproachiscausedbyseveralfactors.Firstsinglevariabletestsmakethetesttimeunnecessarilylong.Secondly,open-looptestsareused2inindustrialMPCprojects.Whentheprocessissensitiveand/ornon-linear,suchasahighpuritydistillationcolumn,itisverydifficulttocarryoutopen-looptestsontheprocess.Upsetsoftenoccuranditisdifficulttofitalinearmodelwhencontrolledvariablevariationsaretoolarge.Finally,mostindustrialidentificationpackagesarebasedonFIRmodelthatisanonparametricmodelwithalargenumberofparameters.Incontrolcommunity,system(process)identificationhasbeenoneofthemostactivebranchesinthelastthreedecades.Anastonishingfactisthatmostoftheidentificationresultsdevelopedinthelast30yearsarenotusedbyindustrialcontrolengineers,althoughthereisanurgentneedforefficientandeffectiveidentificationmethodsinprocesscontrolindustry.Inthelastdecade,thereisarenewedinterestinclosed-loopidentificationandcontrol-relevantidentification;seevandenHof(1997)forarecentsummaryofthework.Manyidentificationschemeshavebeenproposed.Unfortunately,mostresearchersinvolvedinclosed-loopidentificationassumethattheexistingcontrollerislinearandtheprocessissinglevariable.ThismakestheirresultnotrelevantforuseintheframeworkofMPC,becauseaMPCcontrollerisnonlinearandtheprocessundercontrolismultivariable.Althoughcalled“conventional”bymanyresearchers,thepredictionerrormethodofLjung(1987)isstillamorepowerfulmethodologythanmanynewlyproposedschemesforuseinclosed-loopidentificationforMPC.Recently,Zhu(1998)hasdevelopedaso-calledASYMmethodthatusesautomatedmultivariabletestsandparametricmodels.Bettermodelscanbeobtainedwithmuchshortertestandover60%timesavingisreported.InthisworkwewillcontinuethedevelopmentofZhu(1998)andstudyclosed-loopidentification.InSection2thecharacteristicsofrefinery/petrochemicalprocessesarediscussedandclosed-loopidentificationismotivated.InSection3theASYMmethodisintroducedwhereclosed-looptestdesignandmodelvalidationwillbeemphasised.InSection4,theASYMisusedtoidentifyadeethanizerforMPCcontrol.Section5containstheconclusionsandperspectives.2Closed-LoopIdentification,WhyandHow?Closed-loopidentificationmeansthatprocessmodelisidentifiedusingdatacollectedfromaclosed-looptestwheretheunderlyingprocessisfullyorpartlyunderfeedbackcontrol.Inthissectionwewillmotivatetheuseofclosed-loopidentificationfrombothprocessoperationpointofviewandcontroltheoreticalpointofviewanddiscussvariousissuesaroundclosed-loopidentification.Wewillfocusourdiscussiononhydrocarbonprocessindustry(HPI)processeswhereMPCapplicationismostchallengingandalsomostbeneficial.Thisclassofprocessescanbecharacterisedasfollows:1)Largescaleandcomplex.AlargesizeMPCwillhave10to20MVsand20to40CVs.SomeCVs,suchasproductqualities,areveryslow(withdominanttimeconstantrangefrom30minutestoseveralhours),andotherCVsareveryfast,suchasvalvepositions(withtimeconstantinfewminutes).Thereexistinverseresponses,oscillationsa