Page - 364 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1253 wherexi is theactualvalue,xmin andxmax equals theminimumandmaximumvalues in thesamples, respectively,yi represents thenormalized load. 4.2.ModelPerformanceEvaluation Thispaperassesses the forecastingmodelbyusingthe followingappropriate indicators. (1) Relativeerror (RE): RE= xi− xˆi xi ×100% (12) (2) Rootmeansquareerror (RMSE): RMSE= √ 1 n n ∑ i=1 ( xi− xˆi xi ) 2 (13) (3) Meanabsolutepercentageerror (MAPE): MAPE= 1 n n ∑ i=1 |(xi− xˆi)/xi| ·100% (14) (4) Averageabsoluteerror (AAE): AAE= 1 n ( n ∑ i=1 |xi− xˆi|)/(1n n ∑ i=1 xi) (15) where x is the actual load of charging station and xˆ is the corresponding forecasted load, n represents thegroups in thedataset. Thesmaller theseevaluation indicatorsare, thehigher the predictionaccuracy. 4.3. ResultsAnalysis InNILA, set agemat = 3, κstrenth = 5, themaximumiterationnumber is 100, p= 0.5, and the specificiterationprocessisshowninFigure8.AscanbeseeninFigure8, theoptimalparameterofCNN isobtainedat the thirty-fifth iteration. Inorder tovalidate theperformanceof theproposedtechnique NILA-CNN, comparisons are made with the final forecasting results from different algorithms involvingLA-CNN, singleCNN, SVM, and time series (TS). The parameter settings in LA-CNN modelare consistentwith those inNILA-CNN.TheCNNmodel consistsofone featureextraction layerwhich includesaconvolutional layerwith12neurons,andasubsampling layerwith5neurons. Themaximumnumberoftrainingtimes,andthetrainingerror,are200and0.0001,respectively. InSVM, the regularizationparameter is 9.063, thekernelparameter equals 0.256, and the lossparameter is equal to3.185. InTable2, loadforecastingresultsarederivedfromfivedifferent techniques. 364