Page - 313 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1561 Ontheotherhand, inordertoobtainabetterperformanceandaccuracy, theproposedapproachis morecomplex. Thealgorithmwithhighercomplexityoften takes longer inpractice.AsTable7shows, comparedwithLSSVMandFITNET, theexecutiontimesof theproposedmodelare longer,which is themajordisadvantage.However,with thedevelopmentofhardware, theoperationalcapabilityof computercanbe improved,andtheexecution timecanbereduced. Furthermore,asakindofartificial intelligence technique, thefine-tuningof hyper-parameters in theproposedmodelwill take time, which isacommonsituation inacademicandindustrialfields. 5.4. FurtherResearchProspect This paper proposes a hybrid interval predictionmodel to predict the power load intervals. Comparedwithotherbasicmodels, thismodelhasachievedgoodresults intermsofcoverage, interval width,anddeviationerrorof theprediction interval. Themodelcanobtainrelativelyhighcoverage under the condition of relatively narrow intervalwidth, and the interval obtained can accurately reflect thechangesof futureshort-termpower loadandprovidemoreaccurateandreliablesupport for powerdispatch.Ontheotherhand, fordatasetswithmorecomplexchangesandnon-linear features, althoughtheperformanceofproposedmodel is improvedcomparedwiththe traditionalmodels, it is stillnot ideal insomecases. For theunfavorable results causedbythecharacteristicsofdatasets, wemayexplore the followingtwoaspects in future: (a) Findingandimprovingpredictionmethods thatcanbetter solve thenon-linearcharacteristicsof electrical loads,andimprovingtheperformanceofpredictivemodels incomplexsituations; (b) Fullyanalyzing therelevant characteristics in thepower loaddata, selectingdifferentmodels for different characteristics, and using ensemble learning to integrate and enhance the predictionresults. 6.Conclusions STLFis thebasicworkofpowersystemplanningandoperation.However, thepower loadhas regularity andcertain randomness at the same time,which increases thedifficulty ofdesiredand reliable STLF.Moreover, comparedwith thepredictionof specificpoints, interval predictionmay providemore informationfordecisionmaking inSTLF. In thisstudy,basedonLUBE,wedeveloped anovelhybridmodel includingdatapreprocessing,amulti-objectivesalpalgorithm,andE–LUBE. In theory, such a hybridmodel can reduce the influence of noise in a dataset and the parameter optimizationprocess ismoreeffectiveandefficient inE–LUBE. Inourproposedapproach,weusedamulti-objectiveoptimizationalgorithmtosearch for the parametersof theneuralnetworkandreconstructedthecost functionwithdouble intervalcriterions insteadofpointcriterions (suchasMSE) in the traditionalmethod.AsTables2–5show,bycomparing itwith traditionalmethods, theproposedapproachprovidesahigherCPandalower intervalwidthat thesametime,whichmakesupfor the lowerCPandhigher intervalwidthof traditionalmethods. Inordertoverifytheperformanceoftheproposedmodelandvalidatetheimpactoftheconstituent components inahybridmodel,wecollected16samples involvingfourstatesusingfourquarters in Australia,andset severalmodelcomparisons inexperiments Furthermore,accordingto thecomparisonandanalyses results, theconclusionsaresummarized as follows: (a) an efficient data preprocessing method was applied herein. Depending on the decomposition and reconstruction, thismethodcan significantly improve themodelperformance in STLF. (b)Compared to the traditional predictionmodels basedonneural networks, the newly developedE–LUBEmethodhas anadvantage in termsof comprehensiveperformance in interval prediction. It canbevalidatedthat thecontext layerwith the informationof the formerhiddenlayer can improvemodelperformance. (c)The introductionof thenovelmulti-objectivealgorithmMOSSA optimizedtheprocessofparametersearch. Thenewcost functionwasbasedonadouble-objective interval indexthatoutperformedthetraditionalsingle-objectivepointerror index(suchasMSE) in intervalprediction. (d)ForSTLFbasedontheE–LUBEmechanism,thewidthcoefficientisanimportant 313