Page - 278 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1605 modelsconsideredinthisstudy. Inall themodels, theSMLE-basedBPNN-SVRmodeldidnotonly accomplish thehighestaccuracyat the levelestimation,whichwasmeasuredbytheMAPEcriteria, it additionallygot thehighesthit rate indirectionprediction,whichwasestimatedbytheDAcriterion. Thenagain, amongthemajorityof themodelsutilizedasapartof this investigation, thesingleLR modelperformedthepoorest inallprogressionaheadforecasts. LRmodelnotonlyhadthe lowest levelaccuracy,whichwasmeasuredbyMAPE,butalsoacquiredtheworst score indirectionaccuracy, whichwasmeasuredbytheDAcriteria. Themainreasonmightbe thatLRwasaclassof the typical linearmodelandit couldnotcapture thenonlinearpatternsandoccasionalcharacteristicsexisting in thedataseries.Apart fromtheSMLE-basedBPNN-SVRandLRmodels,whichperformedthebest andthepoorest, respectively.Allmodels listed in this studyproducesomeinterestinglyblendresults, theseoutcomeswereanalyzedbyusingfourestimationcriteria (i.e.,MAPE,DA,T-test, andCGR). Figure5. Illustrated10-aheadconsumptionpredictionandMAPE.(a)Singlemodels. (b)Errorofsingle models. (c)Classicensemblemodels. (d)Errorofclassicensemblemodels. (e)SMLEmodels. (f) error ofSMLEmodels. Firstly, in the case of level accuracy, the results of the MAPE measure demonstrated that the SMLE-based BPNN-SVR performed the best, followed by SMLE-based BPNN, SMLE-based SVRmodels, SVRandBPNN, and theweakestmodelwasLRas shown inFigure 5b,f.Moreover, 278