Seite - 355 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1253 inaccurateestimation,consideringtherandomlyselecteddistributionparameters [5].Additionally, Ref. [6]carriedoutchargingloadpredictionofEVsbasedonthestatisticalanalysisofvehicledatafrom theperspectiveof timeandspace. Inorder tosimulate thedrivingpatternsofEVs,Ref. [7]outlined animprovedcharging loadcalculationmodel,wherechargingprobabilitywasproposedto illustrate theuncertaintyof chargingbehaviors andkerneldensity functions. Multidimensionalprobability distributionfunctionswereutilizedtoreplacedeterministicones,andarandomnumberwasgenerated topresent thecouplingcharacteristicsofdrivingdiscipline. Theviewofbigdatawas indicated in the literature [8],whichcalculatedthe loadofeveryEVsat thechargingstation,andsummedthemup; thus, loadforecastingresultswereobtained.Nevertheless, thesestatisticalapproachesarecriticizedby researchers for theirweaknessofuniversality,dueto thedifficultyofparameterdetermination. With the rapid development of artificial intelligence (AI) technology, intelligent algorithms, whichmainly include artificial neural networks (ANNs) and support vectormachines (SVM), are graduallyapplied to loadforecastingofEVchargingstationsbyscholars [9]. Ref. [10]employedback propagationneuralnetwork(BPNN)models topredict thedaily loadcurveofEVchargingstations, withconsiderationofvariousfactors.Here, fuzzyclusteringanalysisbasedontransferclosuremethods wasadoptedtoselect thehistorical loadsimilar to thepredictedoneas the trainingsamples, soas to improveforecastingaccuracy. ThedrawbacksofBPNNare theexistenceofmanyparameters toset, andtrappingintothe localminimumorover-fittingeasily. Toaddress theseproblems,Ref. [11]studied a short term load forecastingmodel for EVcharging stations on the basis of radial basis function neuralnetworks (RBFNN),andmodified itby theuseof fuzzycontrol theory. Theresults showedthat predictionaccuracywasfurther improved. In [12],particleswarmoptimizationandspikingneural networkswerecombinedtoforecast theshort termloadofEVchargingstations. Thefindingsrevealed that thepredictionaccuracyof theproposedmodelwassuperior toBPNN.AnSVMintegratedwith geneticalgorithmswasexploited inshort termloadforecasting forEVchargingstations in [13],which illustratedthat itwasdifficult forSVMstodealwith large-scale trainingsamplesandachieve ideal predictionaccuracy. Theaforementionedalgorithmsbelongtoshallowlearningwithweakability in processingcomplexfunctions,andcannotcompletelyreflect thecharacteristicsof informationbased onpriorknowledge. To thisend,deep learningalgorithmsprovidebetterways topresentdata feature byabstractingthebottomfeaturecombination intohigh-level [14]. Atpresent,deep learningalgorithmshavebeenwidelyapplied invariousfields, especially in the field of prediction. Ref. [15] executed an advertising click rate predictionmethodbased on a deepconvolutionalneuralnetwork (CNN).Thismodelaccomplishedfeature learning throughthe simulationofhuman thinking, andanalyzed the role ofdifferent features in forecasting. Ref. [16] successfully introduceddeepstructurenetworks intoultrashort termphotovoltaicpowerpredictions. AdeepbeliefnetworkwithrestrictedBoltzmannmachinewaspresentedtoextractdeepfeatures to finishtheunsupervised learning,andthesupervisedBPNNwastakenas theconventionalfitting layer toobtain the forecastingresults. Ref. [17]builtdeepCNNforbioactivitypredictionofsmallmolecules in drug discovery applications. These studies have demonstrated that deep learning algorithms have better prediction accuracy in comparison to shallow learning. CNNallows the existence of deformeddataandreducesparameters throughlocal connectionandweightsharing; thus, forecasting precisionandefficiencycanbegreatly improved[18].Asaresult,CNNisselectedas theprediction model in this paper. Notably, the fitting accuracy of CNN is influenced by its two parameters’ selection,namely:weightandthreshold.Consequently, it’svital toapplyanappropriate intelligent algorithmtodetermine thesesvalues. Several traditionaloptimizationalgorithmshavebeenusedto selectparameters forCNN,suchasgeneticalgorithms,particle swarmoptimizationsandantcolony algorithms.Althoughtheabovealgorithmshave theirownadvantages, theyalsohavecorresponding shortcomings. Forexample,geneticalgorithmcannotguarantee theconvergenceto thebest,andis easy to fall into the local optimum,which leads to adecrease inpredictionaccuracy [19]. Particle swarmoptimizationwill appear inpremature convergence indifferent situations [20]. Ant colony algorithmshave lowsearchingefficiencyandlongcalculationtimes,andlocal searchaccuracy isnot 355