Page - 306 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1561 4.4. Experiment II:CaseswithSmallerWidthCoefficients In this experiment, we set the intervalwidth coefficient α = 0.025, whichmeanswe set the output tobe [0.925×X,X,1.025×X] forasinglesample in the trainingprocessof theneuralnetwork. Withanarrowwidthcoefficient, the lowerandupperboundswerecloser to the targetvalue in the trainingprocess,whichcanprovidemorevaluable information inpractice.However,anarrowbound might leadto the increaseofCP.Thus,asmallerwidthcoefficient requires themodels tohavebetter predictiveproperties. The results of this simulation are shown inTables 4 and5, and inFigure 6. Correspondingly, the followingconclusionscanbedrawn: (1) AsTable4andFigure6show, thedistinctionof themodels is similar toexperiment I.TheCPsof theoriginalLUBEandE–LUBEare thesmallestamongthemodels inoursimulation,andour proposedmodelCEEMDAN–IO–E–LUBEelicits thebestperformance (2) Forsomebenchmarkmodels in thisexperiment,withanarrowboundin the trainingprocess, theperformancewasnotadequatelysatisfactory.Asthecasesof the thirdquarter inNSWdenote andthesecondquarter inTAXshowtheCPsofLUBEandE–LUBEareclose to50%,which isnot conclusive inpractice.However,basedonthehybridmechanismweproposed, theperformances were improvedsignificantly. TheminimumCPvaluesofCEEMDAN–IO–E–LUBEcanreach70%, andthemaximumisclose to100%,suchas in the thirdquarter inQLD.Suchresults showthat thepredicted intervalscanbettercoveractualelectricitydemanddataandeconomizespinning reserve inpowergrid. (3) Withasmallerwidthcoefficient, theCPsdecreasedwhile thePINAWandPINRWarereduced. For thebenchmarkmodels, theresultsmostlydisplaysmallerCPsandlargerPINAWorPINRW. However, theproposedmodel isabletodemonstrate largerCPswithsmallerPINAWandPINRW values,which isequivalent toagoodperformance in intervalprediction. Insomecases, theCP valueswere larger than95%withPINAWandPINRWvalues less than10. Insuchcases, theCPs aresatisfactoryandthewidthsof thePIsaremostappropriate. (4) In terms of AWD in this experiment, the proposed model still showed a relatively small AWDcomparedwithotherbenchmarkmodels,whichmeanstheproposedmodelhasabetter performanceatpredictedaccuracy.Comparedwithexperiment I, theAWDsinthisexperiment arebigger. Forasmallerwidthcoefficient, thepredicted intervalwillbenarrower,whichmeans therewill bemore target points falling outside the intervals. In some situations, a narrower predicted interval isnecessary. Theproposedmodel isable toprovideabetterperformanceon theconditionof therequirementofanarrowerpredicted intervalofelectric load. 4.5. ComparisonsandAnalyses According to the comparisonof theabove twoexperimental results, thewidthcoefficienthas asignificant influenceonperformance,asshowninFigure7. Fromoneperspective, formostmodels, acoefficientwitha largerwidthmayleadtoa largerandmoresatisfactoryCPvalue,but the index about thewidthofPImaynotbedesired. Fromanotherperspective, formostmodels, anarrower widthcoefficientmayelicit thedesiredPINAWandPINRWvalues,but theCPisnotgoodenough. Considering such a situation, the proposedmodels alleviate the contradiction. Even though the CPvalueof theproposedmodelwilldeclinewhenthewidthcoefficientdecreases, comprehensive performance issatisfactory. Insomeexceptionalcases,owingto thecomplexityandinstabilityof the datasets, theperformanceof theproposedmodels isnotadequate,as thedescriptioninFigure3shows. 306