Improving Market Clearing Price Prediction by Using a Committee Machine of Neural Networks

《Improving Market Clearing Price Prediction by Using a Committee Machine of Neural Networks》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、IEEETRANSACTIONSONPOWERSYSTEMS,VOL.19,NO.4,NOVEMBER20041867ImprovingMarketClearingPricePredictionbyUsingaCommitteeMachineofNeuralNetworksJau-JiaGuo,StudentMember,IEEE,andPeterB.Luh,Fellow,IEEEAbstract—Predictingmarketclearingpricesisanimportantbutdifficulttask,andneuralnetworkshavebeenw

2、idelyused.Asingleneuralnetwork,however,maymisrepresentpartoftheinput-outputdatamappingthatcouldhavebeencorrectlyrepre-sentedbydifferentnetworks.Theuseofa“committeemachine”composedofmultiplenetworkscaninprinciplealleviatesuchadifficulty.Amajorchallengeforusingacommitteemachineistoproperl

3、ycombinepredictionsfrommultiplenetworks,sincetheperformanceofindividualnetworksisinputdependentduetomappingmisrepresentation.ThispaperpresentsanewmethodinwhichweightingcoefficientsforcombiningnetworkpredictionsFig.1.Schematicofanensemble-averagingcommitteemachinewithnetworkaretheprobabi

4、litiesthatindividualnetworkscapturethetruepredictionsy^andweightingcoefficientsa.input-outputrelationshipatthatpredictioninstant.TestingoftheNewEnglandmarketcleaningpricesdemonstratesthatthenewspaceofmixturesofexpertsisdividedintoseveralregionsmethodperformsbetterthanindividualnetworks,

5、andbetterthancommitteemachinesusingcurrentensemble-averagingmethods.(subspaces)towhichdifferentneuralnetworksareassigned.Agatingnetworkdecideswhichnetworkpredictionwillbese-IndexTerms—Committeemachines,energypriceforecasting,lectedbasedoninputdata[6].Incontrast,thesecondapproachmultipl

6、emodelapproach,neuralnetworks.combinespredictionsofmultiplenetworks.Awell-knownmethodistheensemble-averagingmethodasdepictedinFig.1I.INTRODUCTIONwherepredictionsofneuralnetworksarelinearlycombinedEURALNETWORKShavebeenwidelyusedinmanybasedonastraightaverageorthestatisticsofhistoricalpre

7、dic-Nforecastingproblems,includingloadandmarketclearingtionerrors[7][9].price(MCP)predictionsforpowersystems[1][3].ThemainTheneuralnetworksinFig.1maybeofdifferentkinds,orofreasonfortheirsuccessisthattheyarecapableofinferringthesamekindbutwithdifferentconfigurations(e.g.,differenthidde