Page - 368 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1253 Table3.Thedatadivisionofcase two. DataType DataRange SeasonType Trainingset 1 June2016–24August2016 Autumn 1September2016–23November2016 Winter 1December2016–21February2017 Spring 1March2017–24May2017 Summer Test set 25August2016–31August2016 Autumn 24November2016–30November2016 Winter 22February2017–28February2017 Spring 25May2017–31May2017 Summer Thefivemodels shownaboveare still used in this experiment,where theparameter settings ofNILA-CNN,LA-CNNandCNNare consistent. InSVM, the regularizationparameter is 2.0153, thekernelparameter is0.015,andthe lossparameter is0.013. Thestatisticalerrors includingRMSE, MAPEandAAEaredisplayedinFigure12. Figure12.RMSE,MAPEandAAEofpredictionmethods(II). ((a) is theerrorresultsof testset inSpring; (b) is theerror resultsof test set inSummer; (c) is theerror resultsof test set inAutumn; (d) is theerror resultsof test set inWinter). AsdemonstratedinFigure12, thevaluesofRMSE,MAPEandAAEofNILA-CNNinfourseasons areall the lowestamongthe forecastingtechniques,namely2.010,2.00%and1.97%inSpring,1.93%, 1.86%and1.80%inSummer,2.16%,2.14%and2.04%inAutumn,2.07%,2.00%and1.90%inWinter. Meanwhile, it canbenotedthat theoverallpredictionaccuracyofLA-CNNisbetter thanthatof the CNNmodel,andCNN-basedapproachesaresuperior toSVMandTS,whichproves theadvantages ofNI,LAandCNN.Therefore, theshort-termloadforecastingforEVchargingstationsbasedonthe NILA-CNNmodel is efficientenoughtocompetewithexistingapproaches inpredictionprecision. Asahybridalgorithm, theproposedmodel is able toprovideaccuratedata support for economic operationof thechargingstation. 368