神经网络英文文献.docx

神经网络英文文献.docx

《神经网络英文文献.docx》由会员分享，可在线阅读，更多相关《神经网络英文文献.docx（10页珍藏版）》请在冰豆网上搜索。

神经网络英文文献

ARTIFICIALNEURALNETWORKFORLOADFORECASTING

INSMARTGRID

HAO-TIANZHANG,FANG-YUANXU,LONGZHOU

EnergySystemGroup,CityUniversityLondon,NorthamptonSquare,London,UK

E-MAIL:

,long.zhou.

Abstract:

Itisanirresistibletrendoftheelectricpowerimprovementfordevelopingthesmartgrid,whichappliesalargeamountofnewtechnologiesinpowergeneration,transmission,distributionandutilizationtoachieveoptimizationofthepowerconfigurationandenergysaving.Asoneofthekeylinkstomakeagridsmarter,loadforecastplaysasignificantroleinplanningandoperationinpowersystem.ManywayssuchasExpertSystems,GreySystemTheory,andArtificialNeuralNetwork（ANN）andsoonareemployedintoloadforecasttodothesimulation.ThispaperintendstoillustratetherepresentationoftheANNappliedinloadforecastbasedonpracticalsituationinOntarioProvince,Canada.

Keywords:

Loadforecast;ArtificialNeuronNetwork;backpropagationtraining;Matlab

1.Introduction

Loadforecastingisvitallybeneficialtothepowersystemindustriesinmanyaspects.Asanessentialpartinthesmartgrid,highaccuracyoftheloadforecastingisrequiredtogivetheexactinformationaboutthepowerpurchasingandgenerationinelectricitymarket,preventmoreenergyfromwastingandabusingandmakingtheelectricitypriceinareasonablerangeandsoon.Factorssuchasseasondifferences,climatechanges,weekendsandholidays,disastersandpoliticalreasons,operationscenariosofthepowerplantsandfaultsoccurringonthenetworkleadtochangesoftheloaddemandandgenerations.

Since1990,theartificialneuralnetwork（ANN）hasbeenresearchedtoapplyintoforecastingtheload.“ANNsaremassivelyparallelnetworksofsimpleprocessingelementsdesignedtoemulatethefunctionsandstructureofthebraintosolveverycomplexproblems”.Owingtothetranscendentcharacteristics,ANNsisoneofthemostcompetentmethodstodothepracticalworkslikeloadforecasting.Thispaperconcernsaboutthebehaviorsofartificialneuralnetworkinloadforecasting.AnalysisofthefactorsaffectingtheloaddemandinOntario,Canadaismadetogivean

effectivewayforloadforecastinOntario.

2.BackPropagationNetwork

2.1.Background

Becausetheoutstandingcharacteristicofthestatisticalandmodelingcapabilities,ANNcoulddealwithnon-linearandcomplexproblemsintermsofclassificationorforecasting.Astheproblemdefined,therelationshipbetweentheinputandtargetisnon-linearandverycomplicated.ANNisanappropriatemethodtoapplyintotheproblemtoforecasttheloadsituation.Forapplyingintotheloadforecast,anANNneedstoselectanetworktypesuchasFeed-forwardBackPropagation,LayerRecurrentandFeed-forwardtime-delayandsoon.Todate,Backpropagationiswidelyusedinneuralnetworks,whichisafeed-forwardnetworkwithcontinuouslyvaluedfunctionsandsupervisedlearning.Itcanmatchtheinputdataandcorrespondingoutputinanappropriatewaytoapproachacertainfunctionwhichisusedforachievinganexpectedgoalwithsomepreviousdatainthesamemanneroftheinput.

2.2.Architectureofbackpropagationalgorithm

Figure1showsasingleNeuronmodelofbackpropagationalgorithm.Generally,theoutputisafunctionofthesumofbiasandweightmultipliedbytheinput.Theactivation

functioncouldbeanykindsoffunctions.However,thegeneratedoutputisdifferent.

Owingtothefeed-forwardnetwork,ingeneral,atleastonehiddenlayerbeforetheoutputlayerisneeded.Three-layernetworkisselectedasthearchitecture,becausethiskindofarchitecturecanapproximateanyfunctionwithafewdiscontinuities.ThearchitecturewiththreelayersisshowninFigure2below:

Figure1.Neuronmodelofbackpropagationalgorithm

Figure2.Architectureofthree-layerfeed-forwardnetwork

Basically,therearethreeactivationfunctionsappliedintobackpropagationalgorithm,namely,Log-Sigmoid,Tan-Sigmoid,andLinearTransferFunction.TheoutputrangeineachfunctionisillustratedinFigure3below.

Figure.3.Activationfunctionsappliedinbackpropagation

（a）Log-sigmoid（b）Tan-sigmoid（c）linearfunction

2.3.Trainingfunctionselection

AlgorithmsoftrainingfunctionemployedbasedonbackpropagationapproachareusedandthefunctionwasintegratedintheMatlabNeuronnetworktoolbox.

TABLE.I.TRAININGFUNCTIONSINMATLAB’SNNTOOLBOX

3.TrainingProcedures

3.1.Backgroundanalysis

TheneuralnetworktrainingisbasedontheloaddemandandweatherconditionsinOntarioProvince,CanadawhichislocatedinthesouthofCanada.TheregioninOntariocanbedividedintothreepartswhicharesouthwest,centralandeast,andnorth,accordingtotheweatherconditions.Thepopulationisgatheredaroundsoutheasternpartoftheentireprovince,whichincludestwoofthelargestcitiesofCanada,TorontoandOttawa.

3.2.DataAcquisition

Therequiredtrainingdatacanbedividedintotwoparts:

inputvectorsandoutputtargets.Forloadforecasting,inputvectorsfortrainingincludealltheinformationoffactors

affectingtheloaddemandchange,suchasweatherinformation,holidaysorworkingdays,faultoccurringinthenetworkandsoon.Outputtargetsaretherealtimeload

scenarios,whichmeanthedemandpresentedatthesametimeasinputvectorschanging.

Owingtotheconditionalrestriction,thisstudyonlyconsiderstheweatherinformationandlogicaladjustmentofweekdaysandweekendsasthefactorsaffectingtheload

status.Inthispaper,factorsaffectingtheloadchangingarelistedbelow:

（1）.Temperature（℃）

（2）.DewPointTemperature（℃）

（3）.RelativeHumidity（%）

（4）.Windspeed（km/h）

（5）.WindDirection（10）

（6）.Visibility（km）

（7）.Atmosphericpressure（kPa）

（8）.Logicaladjustmentofweekdayorweekend

Accordingtotheinformationgatheredabove,theweatherinformationinTorontotakenplaceofthewholeOntarioprovinceischosentoprovidedataacquisition.Thedatawasgatheredhourlyaccordingtothehistoricalweatherconditionsremainedintheweatherstations.Loaddemanddataalsoneedstobegatheredhourlyandcorrespondingly.Inthispaper,2yearsweatherdataandloaddataiscollectedtotrainandtestthecreatednetwork.

3.3.DataNormalization

Owingtopreventthesimulatedneuronsfrombeingdriventoofarintosaturation,allofthegathereddataneedstobenormalizedafteracquisition.Likeperunitsystem,eachinputandtargetdataarerequiredtobedividedbythemaximumabsolutevalueincorrespondingfactor.Eachvalueofthenormalizeddataiswithintherangebetween-1and+1sothattheANNcouldrecognizethedataeasily.Besides,weekdaysarerepresentedas1,andweekendarerepresentedas0.

3.4.Neuralnetworkcreating

ToolboxinMatlabisusedfortrainingandsimulatingtheneuronnetwork.Thelayoutoftheneuralnetworkconsistsofnumberofneuronsandlayers,connectivityoflayers,activationfunctions,anderrorgoalandsoon.Itdependsonthepracticalsituationtosettheframeworkandparametersofthenetwork.ThearchitectureoftheANNcouldbeselectedtoachievetheoptimizedresult.Matlabisoneofthebestsimulation

toolstoprovidevisiblewindows.Three-layerarchitecturehasbeenchosentogivethesimulationasshowninFigure2above.Itisadequatetoapproximatearbitraryfunction,ifthenodesofthehiddenlayeraresufficient.

Duetothepracticalinputvalueisfrom-1to+1,thetransferfunctionofthefirstlayerissettobetansigmiod,whichisahyperbolictangentsigmoidtransferfunction.Thetransferfunctionoftheoutputlayerissettobelinearfunction,whichisalinearfunctiontocalculatealayer’soutputfromitsnetinput.Thereisoneadvantageforthelinearoutputtransferfunction:

becausethelinearoutputneuronsleadtotheoutputtakeonanyvalue,thereisnodifficultytofindoutthedifferencesbetweenoutputandtarget.

Thenextstepistheneuronsandtrainingfunctionsselection.Generally,TrainbrandTrainlmarethebestchoicesaroundallofthetrainingfunctionsinMatlabtoolbox

Trainlm（Levenberg-Marquardtalgorithm）isthefastesttrainingalgorithmfornetworkswithmoderatesize.However,thebigproblemappearsthatitneedsthestorageofsomematriceswhichissometimeslargefortheproblems.Whenthetrainingsetislarge,trainlmalgorithmwillreducethememoryandalwayscomputetheapproximateHessianmatrixwithn×ndimensions.Anotherdrawbackofthetrainlmisthattheover-fittingwilloccurwhenthenumberoftheneuronsistoolarge.Basically,thenumberofneuronsisnottoolargewhenthetrainlmalgorithmisemployedintothenetwork.Trainbr（Bayesianregularization）isamodifiedalgorithmoftheLevenberg-Marquardttrainingmethodtocreatenetworkswhichgeneralizewellsothattheoptimalnetworkarchitecturecanbeeasilydetermined.Impactsfromeffectivelyusedweightsandbiasesofthenetworkcanbeseenclearlybythisalgorithm.Andthenumberoftheeffectiveweightsandbiaseswillnotchangetoomuchwhenthedimensionofthenetworkisgettinglarge.Thetrianbralgorithmhasthebestperformanceafterthenetworkinputandoutputnormalizedintotherangefrom-1to+1.Animportantthingwhenusingtrainbrshouldbementionedisthatthealgorithmshouldnotstopuntiltheeffectivenumberofparametershasconverged.MoredetailsareavailableinMatlabneuralnetworktoolbox.

Numberofneuronsinthefirstlayeralsocanbeselectedtooptimizethenetworksothatanexpectedresultcanbemade.Generallyspeaking,themorecomplicatedarchitectureofthenetworkis,themoreaccuratetheoutputresultwillbe,however,thehigherchanceswillthealgorithmsuchastrainlmwithover-fitting.Inthispaper,thenumberofneuronsis8intrainlmalgorithm,and30intrainbralgorithm.

3.5.Neuralnetworktraining

Be