模式識別神經(jīng)網(wǎng)絡(luò)課件

Stateofcharge(SOC)estimationofhighpowerNI-MHrechargeablebatterywithartificialneuralnetwork(ANN)Speaker:FuWeili080305008automatizationABSTRACTThispaperpresentsathree-layerfeed-forwardback-propagation(BP)artificialneuralnetwork(ANN),whoseoutputisbatterystate-of-charge(SOC),toestimateandpredictSOCofhighpowerNi-MHrechargeablebattery.Especially,theANNcansatisfyinglyestimateSOCofbatterywhosestartingSOCisnotoriginallyknownafterabouttenminutesconstantloaddischarging(CLD),andmostofabsolutevaluesofabsoluteerrorsarenotmorethan5%.1.IntroductionofSOCSOCisaveryimportantcomponentofbatterymanagementsystem(BMS);SOCinthispaperisdefinedasEq.(1):

Ca:availabledischargingcapacityofbatterywhichisoriginallyfullycharged,isafunctionofdischargingcurrent.i:dischargingcurrent.t:time.(1)2.ConceptsofhighpowerNI-MH

rechargeablebatteryThebatteryisabletodischargeandchargeathighratecurrent;Thirteendifferentconstantcurrentdischarging(CCD)datesandtwo(0.15Ω/celland0.675Ω/cell)CLDdatesareobtained;SixCCDdatasetsareselectedtotrainANN.DischargeRateAvailableCapacity(Ah)CellPlatformVoltage(A)0.2C/3.6A18.81.271C18.21.222C17.91.203C17.21.1510C13.50.8915C-0.8Table1A18Ahbattery‘sAvailableCapacityandCellPlatformVoltageatCCDat20℃3.ANNarchitectureThefirstlayerisinputlayer;Thesecondishiddenlayer,anditsactivationfunctionsis"logsig-moid"functionswhichisdefinedasEq.(2):LS(X)=1/(1+ex)(2)TheANNoutputSOCis:SOC=W2*LS(W1*X+B1)+B2whereXistheinputvector,B1andB2arethebiasvectorsofANNinthehiddenlayerandoutputlayer.W1、W2areweightmatrices.4.SelectingofANNinputsTemperaturefactorisneglected;BasedontheexperienceandknowledgeofbatterythefollowingvariablesareinitiallyselectedascandidateinputsoftheANN:Dischargingcurrenti;AccumulatedamperehoursAh=;BatteryterminalvoltageV;Time-averagevoltagetav(t)=

;Twicetime-averagevoltagettav(t)=

;...ThemethodtodeterminateinputvariablesTherearetwobasalparameters:i,v.Otherparameterscanbederivedfromthetwoandsamplingtimet.Whenj=3infigure2,satisfyingerrorsareobtained.Figure1

flowchartoftheproceduretodeterminateANNinputs5.Testingresults

TotesttheANNmodel,otherCCDdatasetsandtwoCLDdatasetsaresimulatedbythetrainedANN.maxofabsolutevaluesofabsoluteerror(abs.err.)ofdifferentCCDdatasetsareshowninTable2.ComparisonresultsoftwoCLDdatasetsaresummarizedinTable3.CCDcurrent(A)MAX︱err︱(%)51.69151.19206.21300.59350.40401.55502.09Table2max︱abs.err︱atCCD0.15Ω/cell0.675Ω/cellFirsttimewhen︱abs.err︱<5%10min6min︱abs.err︱<5%time/totaltime62.09%50.47%Table3ComparisonresultsofSOCpredictionatCLD6.Conclusion

Comparisonsbetweensimulationandmeasurementshowthat:

①maxofabsolutevaluesofabsoluteerrorisnomorethan2.1%atCCD;②ANNcanaccuratelypredictSOCofbatterywhosestartingSOC0isnotknownoriginallyafterabouttenminutesatCLD,andmostof︱abs.err︱<5%;③50ACCDisoutofthecurrentrangeoftrainingdatasetsandcurrentsof0.