Page - 366 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1253 Figure8.The iterativeprocessofNILA. Figure9displays thepredictionresultsofTable2, shownformore intuitiveanalysis. Thevalues ofREobtained fromthe forecastingmodelsare illustrated inFigure10. Under thecircumstanceof electricitymarket, theerrorrangebetweenshort-termloadforecastingandtheactualvalueshould be [−3%, +3%]. It can be seen that theprediction error range ofNILA-CNN is controlledwithin [0.23%,2.86%]while thepredictionerrorrangesofLA-CNNandCNNare[0.62%,3.47%]and[−4%, 2.28%], respectively.Amongthem,6errorpointsofNILA-CNNarecontrolled in [−1%,1%],while thecorrespondingnumberofLA-CNNandCNNare3and0. TheerrorsofSVMmodelmostlyrange from[−6%,−4%]or [4%,6%],andadditionally, theerrorsofTSpresenta largefluctuationranging, from[−8%,−5%]and[5%,8%]. Thus, thepredictionprecision fromthesuperior to the inferiorcanbe rankedas follows:NILA-CNN,LA-CNN,CNN,SVM,TS.Thisdemonstrates thatNIcaneffectively improvetheperformanceofLA.Further,NILAisconducive tohighforecastingaccuracy,dueto the optimalparametersettingintheCNNmodel.AlthoughthepredictionresultsofNILA-CNNmodelare greater thanother fourmethods insomepoints, suchasat10: 30, theoverall errorsperformthebest. Figure9.Predictionresults. 366