外文翻譯-伊朗國家電網(wǎng)利用人工神經(jīng)網(wǎng)絡(luò)的短期負荷預(yù)測

1、英文原文ShtTermLoadFecastingofIranNationalPowerSystemUsingArtificialNeuralwkGenerationTw.BarzaminiM.B.MenhajSh.KamalvA.TajbakhshAbstract—Thispaperpresentsaneurobasedshttermloadfecasting(STLF)methodfIrannationalpowersystem(IN

2、PS)itsregions.Thisisanimprovedversionoftheonegivenin[1].ThearchitectureoftheproposedwkisathreelayerfeedfwardneuralwkwhoseparametersaretunedbyLevenbergMarquardtBP(LMBP)augmentedbyanEarlyStopping(ES)methodtriedoutfincreasi

3、ngthespeedofconvergence.Insteadofseasonaltraininganinputasamonthindicatisaddedtotheinputvects.TheshttermloadfecastingsimulatdevelopedsofarpresentssatisfactybetterresultsfonehouruptoaweekpredictionofINPSloadsregionofINPSB

4、akhtarRegionElectricCo(BREC).I.INTRODUCTIONLoadfecastinghasalwaysbeentheessentialpartofanefficientpowersystemplanningoperation.Generallytherearetwogroupsoffecastingmodelstraditionalmodels(modelbasedtechniques)moderntechn

5、ique(knownasmodelfreetechniques).Traditionalloadfecastingmodelsaretimeseriesregressionanalysis.Inrecentyearscomputationalintelligencemethodsaremecommonlyusedfloadfecasting[210].Multilayerfeedfwardneuralwksasuniversalappr

6、oximatesareverysuitablefloadfecastingbecausetheyhaveremarkableabilitytoapproximatenonlinearfunctionswithanydesiredaccuracy.ionoftheinputoutputtrainingdatainputvectoftheneuralwkplaysacrucialrole.Essentiallyinourcase(loadf

7、ecastingproblem)theMLPbasedwksaregreatlyaffectedbyionofinputs.DaytypeMonthtypehisticalloaddataweatherinfmation.Howtochoosethehourlyloadinputsfeachweeklygroupplaysanimptantroleinimprovingwksperfmance(sectionII).ThesecondN

8、irooResearchInstitute(NRI)STLF(NSTLFII)programisbasedonathreelayerfeedfwardneuralwkbuildingblock.FthetrainingofthisMLPinsteadofconventionalbackpropagation(BP)methodstheLevenbergMarquardtBP(LMBP)EarlyStopping(ES)methodswa

9、semployedindertoreachtheoptimumwk’sparametersfasteralsoinsteadofseasonaltrainingthemonthinputwasaddedtotheinputvects(sectionIII).SomeexamplesoftheNSTLFIIperfmancearepresentedusingINPSactualloadtemperaturedataoftheyear200

10、0regionofINPSBakhtarregionelectricalCo(BREC)actualloadtemperaturedataoftheyear2002(sectionIV).FuturewkscanaddressthefuzzysystemapplicationfspecialconditionsreshapingfwardMLPathreelayerfeedfwardMLPbuildingblockhasbeenused

11、.Generallyneuralwkswithahiddenlayerhavetheremarkableabilitytoapproximatemostnonlinearfunctionswithadesiredaccuracyifthereareenoughhiddenneurons.TherefethemodelshowninFig.1iscomposedofthreelayerseachlayerhasafeedfwardconn

12、ection.InthismodelinputsfeachweeklygroupareseparatelytrainedbyanMLP.TheinputlayerfhourlyloadfecastofeachweeklygroupSaturdayswkdaysThursdaysFridayshasrespectively13191619neurons(consistingofedloadlags3representatives’feca

13、stedtemperatures1nodetoindicatemonth).Throughadeepinvestigationwefoundthatahiddenlayerwith5neuronswksquitewell.Ofcoursethewkhasoneoutputneuron(Thefecastedload).TovalidatethequalityofthedevelopedMLPwerunitwithyear2000’loa

14、ddataofINPSyear2002’loaddataofBRECindertobeabletojudgethemeritofthemethod.Neuronsinthehiddenoutputlayershavenonlineartransferfunctionknownasthe“tangentsigmoid“(tansig)function:Fig.1.TheMLParchitecturefeachweeklygroupThew

15、eightedinputsreceivedbyatansignodearesummedpassedthroughthisnonlinearfunctiontoproduceanoutput.Thetansigfunctiongeneratesoutputsbetween–11itsinputsshouldbeinthesamerange.AsaresultitisnecessarytolimittheMLPinputstargetout

16、puts.Meanstarddeviationminimum(min)Maximum(max)nmalizationmethodshavebeentestedminmaxMethodhasbeened:Thisnmalizationmethodhasalsotheadvantageofmappingthetargetoutputtothenonsaturatedsectoftansigfunction.Thisprocesshelpsi

17、nimprovingtheaccuracyofboththelearningfecastingmodes.TheMLPscanbetrainedfeachweeklygroupofayeartherelatedweightsbiaseswillbegainedusedffecasting.AuserfriendlyinterfacehasalsobeendesignedfNSTLFIIwhichgivestheuserstheabili