Page - 303 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1561 objectof the intervalprediction isa largerCPvaluealongwithanarrowwidth. Therefore, the IO shouldhaveanadvantage (3) IncorporationofCEEMDANinthehybridmodels is improvedtheperformancessignificantly because of the denoising preprocessing. In most cases, the CPs are larger than 80% and 90%,whichmeansmore than 80% target loadvalues are coveredby thepredicted intervals. Furthermore, in some cases, theCPs can reach 100%, such as the second and third quarters inNSW, and the second quarter inQLD. Such accuracy can ensure that the power supply meets the demand. Comparedwith the original LUBE and E–LUBE, the hybridmodelwe proposed(CEEMDAN–IO–E–LUBE)elicitedasignificant improvement in theelicitedresultsof intervalprediction. (4) Witha largerwidth coefficient, theCPsofourmodelswerealmost satisfactory. The smallest CPwasmore than70%,andthe largestCPwasable toreach100%,which isperfect for interval prediction inSTLF.However, thePINAWandPINRWwerealmostall larger than10,andeven reached thevalueof 20 in secondquarter inQLD.But theproposedmodel still outperforms othermodels. (5) Consideringtheaccumulatedwidthdeviation(AWD), fora largerwidthcoefficient, theproposed model (CEEMDAN-IO-E-LUBE)hasasmallerAWDcomparedwithotherbenchmarkmodels inmost cases. According to thedefinitionofAWD,asmallerAWDmeansmore targetvalues fall into thepredicted intervals. For theresults inwhichthe targetvaluesareover thebounds, thedeviations are relatively small. In this experiment, theAWDsof theproposedmodel are satisfactory inmostcase. Forsomecases, theAWDsisevenclosedto0,whichmeansalmostall target loadvalues fall into thepredicted intervals.Accordingto thesepredicted intervals, load dispatchwillbemorerational. Figure5.Performanceofdifferentsampleswith thewidthcoefficient0.05. 303