Page - 330 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1449 4.4.Discussion In order to verify the performance of the forecasting approach, three basic techniques, includingWPA-LSSVM[37],LSSVM[38], andBPNN[39],were introduced tomakeacomparison. TheparametersettingsinWPA-LSSVMwereconsistentwiththoseintheestablishedmodel. InLSSVM, the regularization parameter γ and the kernel parameter σ2 were valued at 12.6915 and 12.0136, respectively. InBPNN, tansigwasutilizedas the transfer function in thehidden layer, andpurelin wasemployedas the transfer function in theoutput layer. Themaximumnumberofconvergencewas 200, theerrorwasequal to0.0001,andthe learningratewassetas0.1. Thedeterminationof the initial weightsandthresholdsdependontheirowntraining. Figure9 illustrates the loadforecastingresults ofFC-WPA-LSSVM,WPA-LSSVM,LSSVMandBPNN.Figure10presents thevaluesofREforeach predictionmethod. Figure9.Forecastingresults: (a) forecastingresultsofSpringtest; (b) forecastingresultsofSummer test; (c) forecastingresultsofAutumntest; (d) forecastingresultsofWinter test. 330