Page - 386 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1138 TheperformancesaregoodwithLSTMnetworks indealingwith timesequencebut thereare moreparameters to traincomparedwithGRUneuralnetworks. In theproposednetwork, theGRU layershave285,300parameters to trainwhile theLSTMlayershave380,400parameterswith thesame architecture. Thecost timefor trainingwithLSTMnetwork isabout20%longer thantrainingwith GRUneuralnetworks in theexperiments in thispaper. TheMAPEsofnetworkwithLSTMandGRU layer in thesamearchitecturewith thesamesamples inCategory2during the trainingprocessare showninFigure14.WecanconcludethatGRUneuralnetworksdobetter inbothconvergencespeed andtrainingtime,whichdependsonthe improvedsinglestructureofGRUunits. (SRFK *58 QHXUDO QHWZRUN /670 QHWZRUNV Figure14. TheMAPEsofnetworkwithLSTMandGRUlayers in the samearchitecturewith same samplesduringthe trainingprocess. Wealsoperformedtheexperiments tocomparewithcurrentmethodssuchasback-propagation neuralnetworks (BPNNs) [7,8], stackedautoencoders (SAEs) [17],RNNs[24,25],andLSTM[29–31]. Theirparametersandstructuresaresetasdescribed inSection3.2. ThecomparedaverageMAPEs of thesemethods, trainedandtestedwithall samplesdescribedat thebeginningof this subsection, areshowninFigure15. ThespecificvaluesofaverageandmaximalMAPEsareshowninTable10. Moreover, theresultsofninecustomersareshowntovalidate thebetterperformanceof theproposed methods. TheMAPEs for 30November 2013 are shown inTable 11. It canbe concluded that the proposedmethodresults insmallererror inbothaverageandmaximalMAPEs. Theproposedmethod performsbettercomparedto theothercurrentmethods inmostcases forshort-termloadforecasting inWanjiangarea. %311V 6$(V 511V /670 3URSRVHG 0HWKRG Figure15.ComparedaverageMAPEstrainedandtestedwithall samples in threecategories. 386