Neural-network-based-approach-for-short-term-load-

1Abstract—Short-termloadforecastisanessentialpartofelectricpowersystemplanningandoperation.Forecastedvaluesofsystemloadaffectthedecisionsmadeforunitcommitmentandsecurityassessment,whichhaveadirectimpactonoperationalcostsandsystemsecurity.Conventionalregressionmethodsareusedbymostpowercompaniesforloadforecasting.However,duetothenonlinearrelationshipbetweenloadandfactorsaffectingit,conventionalmethodsarenotsufficientenoughtoprovideaccurateloadforecastortoconsidertheseasonalvariationsofload.ConventionalANN-basedloadforecastingmethodsdealwith24-hour-aheadloadforecastingbyusingforecastedtemperature,whichcanleadtohighforecastingerrorsincaseofrapidtemperaturechanges.Thispaperpresentsanewneuralnetworkbasedapproachforshort-termloadforecastingthatusesthemostcorrelatedweatherdatafortraining,validatingandtestingtheneuralnetwork.Correlationanalysisofweatherdatadeterminestheinputparametersoftheneuralnetworks.ThesuitabilityoftheproposedapproachisillustratedthroughanapplicationtotheactualloaddataoftheEgyptianUnifiedSystem.IndexTerms—Loadforecasting,neuralnetwork,short-term,correlationanalysis.I.INTRODUCTIONRELIABILITY,securedserviceandeconomicsareessentialrequirementsforelectricitysupply.Utilitiesarealwaysfacedwiththechallengeofmeetingthegrowingloaddemandswhilemaximizingtheirshort-termandlong-termoperationalefficiency.Toachievethisgoal,severaloperationaltaskssuchasfuelallocation,unitcommitment,unitdispatch,reserveallocationandmaintenanceschedulingmustbecarriedoutefficientlytominimizetheoveralloperatingcost.Tomeetthefluctuatingload,utilitiesneedtoensuresufficienttransmissionandgenerationresourcestomeetforecasteddemand.Resourceprocurementsshouldbedoneeconomicallytooptimizetheoperationalcost.Sincetheeconomyoftheoperationandcontrolofpowersystemsaregreatlyaffectedbyloaddemand,significantsavingscanbeattainedbyincreasingtheaccuracyofloadforecast.Highforecasterrorcanleadtoeitheroverconservativeorunreliableoperation.Forexample,higherestimationofloadcancauseunnecessarystartupofadditionalgenerationunitsorexcessiveenergypurchase.Ontheotherside,lowestimationresultsininsufficientspinningreserve,whichiscontrarytosafeoperationrequirements.Lackofsupplyduetoerrorsinloadforecastcanalsoleadtonearrealtimeprocurement,whichisgenerallyatahigherpricecomparedtoscheduledtransactionscost.Itisevidentthatimprovementinloadforecasthasadirectpositiveimpactonsystemsecurityandalsocostofoperation[1].Currently,powerutilitiesareusingvariousloadforecastingtechniquesworldwide.Mostofthedevelopedmethodscanbebroadlycategorizedintothreegroups,namelyparametric,nonparametric,andartificialintelligencebasedmethods[2].Intheparametricmethods,amathematicalorstatisticalrelationshipisdevelopedbetweentheloadandthefactorsaffectingit.Someexamplesofthesemodelsaretimefunctions,polynomialfunctions,linearregressions,FourierseriesandAutoRegressiveMovingAverage(ARMA)models[3]-[6].Intime-seriesmethods,theloadistreatedasatimeseriessignalwithknownperiodicitysuchasseasonal,weekly,ordaily.Suchrepetitivecyclegivesaroughpredictionoftheloadatthegivenseason,dayoftheweek,andtimeoftheday.Thedifferencebetweentheestimatedandactualloadcanbeconsideredasastochasticprocess,whichcanbethenanalyzedusingKalmanfiltermethods[7].Nonparametricmethodsforecasttheloaddirectlyfromhistoricaldata.Forexample,usingnonparametricregression,theloadcanbeforecastedbycalculatinganaverageofhistoricalloadsandthenassignweightstodifferentloadsusingamultivariateproductkernel[8].Recently,significantinterestsandeffortshavebeendirectedtowardstheapplicationofartificialintelligencetechniquestoloadforecasting.Thisincludestheapplicationofexpertsystemstoloadforecasting[9],[10],andcomparingitsperformancetotraditionalmethods[1].Italsoincludestheuseoffuzzyinference[11]andfuzzy-neuralmodels[12],[13].However,themodelsthathavereceivedahighshareofeffortsandfocusaretheartificialneuralnetworks(ANNs).ThemainadvantageofANNsistheiroutstandingperformanceindataclassificationsandfunctionapproximation.ANNisalsocapableofdetectingNeuralNetworkBasedApproachforShort-TermLoadForecastingZainabH.Osman,MohamedL.AwadandTawfikK.Mahmoud978-1-4244-3811-2/09/$25.00©2009IEEE2dependenciesfromhistoricaldatawithouttheneedtodevelopaspecificregressionmodel.FirstpublicationsonANNapplicationtotheloadforecastingproblemweremadeinlate1980’sandearly1990’s[14].SincethenANNhavebeenwellacceptedinpractice,andareusedbymanyutilities[15]-[16].MostoftheconventionalANN-basedloadforecastingmethodsdealswith24-hour-aheadloadforecastingornextdaypeakloadforecastingbyusingforecastedtemperature.Thedrawbackofthismethodisthatwhenrapidchangesintemperatureoftheforecasteddayoccur,loadpowerchangessignificantly,whichleadstohighforecasterror[17].Inaddition,conventionalneuralnetworksuseallsimilarday’sdatathroughoutthetrainingprocess.However,trainingoftheneuralnetworksusingallsimilarday’sdataisacomplextaskanditdoesnotsuitlearningofneuralnetwork.Toovercometheproblemsmentionedabove,thispaperproposesaone-hour-aheadloadforecastingmethodusingthemostsignificantweatherdata.Intheproposedforecastingmethod,wea