基于统计特征的神经网络逐时能耗预测方法(SFBA)(IJITCS-V9-N5-4)

I.J.InformationTechnologyandComputerScience,2017,5,23-30PublishedOnlineMay2017inMECS()DOI:10.5815/ijitcs.2017.05.04Copyright©2017MECSI.J.InformationTechnologyandComputerScience,2017,5,23-30StatisticalFeaturesBasedApproach(SFBA)forHourlyEnergyConsumptionPredictionUsingNeuralNetworkFazliWahid,RozaidaGhazaliUniversitiTunHusseinOnn,MalaysiaE-mail:[e-mail],[e-mail]yofMalakand,PakistanE-mail:[e-mail]ZulfikarAliBhuttoInstituteofScienceandTechnology,PakistanE-mail:[e-mail]—Inthispaper,newstatisticalfeaturesbasedapproach(SFBA)forhourlyenergyconsumptionpredictionusingMulti-LayerPerceptronispresented.Themodelconsistsoffourstages:dataretrieval,datapre-processing,featureextractionandprediction.Inthedataretrievalstage,historicalhourlyconsumedenergydatahasbeenretrievedfromthedatabase.Duringdatapre-processing,filtershavebeenappliedtomakethedatamoresuitableforfurtherprocessing.Inthefeatureextractionstage,mean,variance,skewness,andkurtosisareextracted.Finally,Multi-LayerPerceptronhasbeenusedforprediction.ForexperimentationwithMulti-LayerPerceptronwithdifferenttrainingalgorithms,afinalmodelofthenetworkwasdesignedinwhichthescaledconjugategradient(trainscg)wasusedasanetworktrainingfunction,tangentsigmoid(Tansig)asahiddenlayertransferfunctionandlinearfunctionasanoutputlayertransferfunction.Forhourlyenergyconsumptionprediction,atotalofsixweeksdataoftenresidentialbuildingshasbeenused.Toevaluatetheperformanceoftheproposedapproach,MeanAbsoluteError(MAE),MeanSquaredError(MSE)andRootMeanSquaredError(RMSE),evaluationmeasurementswereapplied.IndexTerms—EnergyEfficiency,EnergyPrediction,EnergyManagement,Multi-LayerPerceptron,NeuralNetwork,ResidentialBuildings.I.INTRODUCTIONInordertofindthefuturedemandsofenergy,theenergyconsumptionpredictionatvariousscalesisveryimportant.Theaccurateenergyconsumptionpredictionalsoplaysakeyroleinefficientproduction,distribution,selling,operation,planningandmanagementintheenergymanagementsystem.Powergenerationsystemscanbemademorereliableifwiththeaccurateenergyconsumptionprediction.Further,ifefficientenergyconsumptioniscarriedout,theeconomy,fuelutilization,andothersectorsdependentonenergymanagementsystemscanbemanagedinabetterway[1].Theenergyconsumptionpredictioncanbedividedinto,shortterm,mediumterm,andlong-termcategoriesbasedonthetimeinwhichtheenergyhasbeenutilized.Differenttechniqueshavebeenproposedintheliteraturefortheenergyconsumptionpredictionwithvariousevaluationcriteria,parametersandthevaryingdegreeofaccuracies[2].Inthispaper,themostsuccessfulArtificialNeuralNetworkclassifierhasbeenusedforthepredictionofenergyconsumption.ArtificialNeuralNetworkhasbeenusedfordifferenttypesofmodelinginvariousdisciplinesincludingMedicine,Mathematics,Economics,Engineering,Meteorology,Psychology,HydrologyandNeurology[3-5].NeuralNetworkshavebecomepopularsincetheirfirstinceptionin1943[6].NeuralNetworkshavestrongcapabilityofsolvingpredictionproblemswithvariableshavingstochasticnature,unknownornon-linearvariations,orlesscontrolledenvironmentisrequiredfortheirdetermination[7].Further,duetotheflexibilityandlessassumption-dependency,thephysicalprocessingbetweeninputandoutputisnotrequiredinNeuralNetworks[8-9].Forthebettermanagementofenergyproductionandstoragesystem,thehourlyenergyconsumptionpredictionisveryimportant.TheobjectiveofthispaperishourlyenergyconsumptionpredictionusingMulti-LayerPerceptron(MLP).Similartopreviousworksinenergyconsumptionprediction,historicaldataofenergyconsumptionhasbeenusedtopredictfutureconsumption.Atotalofsixweeksdataoftenresidentialbuildingshasbeenusedforthehourlyenergyconsumptionprediction.Toevaluatetheperformanceoftheproposed24StatisticalFeaturesBasedApproach(SFBA)forHourlyEnergyConsumptionPredictionUsingNeuralNetworkCopyright©2017MECSI.J.InformationTechnologyandComputerScience,2017,5,23-30approach,MeanAbsoluteError(MAE),MeanSquaredError(MSE)andRootMeanSquaredError(RMSE),evaluationmeasurementswereapplied.Theremainingstructureofthepaperisorganizedas,thesectionIIpresentsLiteratureReview,sectionIIIcontainsProposedModel.SectionIVcontainsdetailsoftheMulti-layerPerceptron,sectionVcontainsExperimentalResultsandDiscussionandfinally,insectionVI,theConclusionisprovided.II.LITERATUREREVIEWTheresearchhasbeencarriedsincelastfourdecayinthefieldofenergypredictionforthebettermanagementofenergyresources.Carpinteiroetal.,in[10],usedNeuralNetworkbasedmodelfortheforecastingoftheshorttermload.Authorshaveintroducedanextensionoftheself-organizingmaptoSTLF.Theproposedmodelhastwoself-organizingmapsfortheloadforecasting.Euclidiandistancehasbeenusedforthecalculationofthedifferencebetweentwovectors.ThedatahasbeenextractedfromBrazilianutility.Thetrainingofthemodelhasbeencarriedoutintwophases:coarsemappingandfinetuning.Themodelhasprovidedsatisfactoryloadforecastinginshortandlongforecasting.Grossetal.,in[11],focusedontheshorttermloadforecasting.Theyhaveusedtimeseriesandregressionmodelsforloadforecasting.Ir