Page - 319 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1449 theory. Ashtarietal. [7] installedGPS equipment on 76 representative plug-in electric vehicles in Winnipeg,Canada,andcollecteddrivingdata for thewholeyear. The loadforecastingwasconducted in consideration of the actual charging sate of battery, stopping time, parking type and vehicle power system. In [8], fourvariables, including thenumberofvehiclesneedingbatterychangeper hour, thestartingtimeofcharging,drivingdistanceandchargingduration,were takenintoaccount under anuncontrolled swapping and charging scenario. On this basis, theMonteCarlomethod integratednon-parameterestimationapproachwasemployedin loadforecastingforelectricvehicle chargingstations.Aswecansee, traditionalpredictionmethodshavetheadvantagesofmature theory, perfectverificationapproachesandsimplecalculation,but theweaknessesofasingleapplicableobject andunideal predictionprecision are also notable. Accordingly, it is of great importance to apply intelligent forecastingtechniques to loadforecastingofchargingstationswith therapiddevelopment ofartificial intelligence technology. Intelligent algorithms for load forecastingchieflyconsist of artificialneuralnetworks (ANNs) and support vector machine (SVM) [9]. Back propagation neural network (BPNN), treated as representative of ANNs, is suitable for load prediction of quick-change e-bus charging stations. For example, [10] analyzed the load characteristics and influential factors, aswell as executing a BPNNmodel to predict the short-term load based on themeasured data of quick-change e-bus charging stations at theBeijingOlympicGames. The approach in this studyproved to beuseful. Additionally, some scholars have adoptedANNs for short-term load forecasting of other power systems. Xiaoetal [11] combined single spectrum analysis (SSA)withmodifiedwavelet neural network(WNN)toestablishareliableshort-termforecastingapproach in thefieldof load,windspeed andelectricityprice. Reference [12] proposedanovel ensemblepredictionmethod for short-term loadforecastingonthe foundationof theextremelearningmachine (ELM),where four improvements weremadeto theELM.Theresults showedthat thepredictionaccuracyof theproposedtechnique wassuperior to thestandardANNs.However, thedrawbacksofANNs includeslowconvergence andeasytrapping into the localminimum,whichgreatly limit the forecastingprecisionandstability. SVMmodelcaneffectivelyaddress theseproblems[13]; thus, thisapproachhasbeenwidelyusedin theresearchof loadforecasting. Reference [14]designedan incremental learningmodelonthebasisof SVMto implement loadpredictionunderbatcharrivalwitha largesample. Inreference [15], anSVM modelbasedontheselectionofsimilardays fordaily loadforecastingofelectricvehicleswascomeup with.Correlationanalysiswaspresentedtoextract the influential factorsandgreycorrelationtheory wasutilized toobtain a small sample set of similardays. ComparedwithANNs, theSVMmodel achievedbetter results for loadforecasting.Nevertheless, the transformationof thekernel functionto convert theproblemintoquadraticprogrammingreduces theefficiencyandaccuracyof traditional SVM[16]. Leastsquaressupportvectormachine(LSSVM)isamodifiedformforSVMwheretheleastsquares linearsystemservesas the loss functiontoavoidquadraticprogramming,andthekernel function is employedto transformpredictionproblems intoequations,aswellas toconvert inequalityconstraints into equality ones, which can improve the forecasting accuracy and speed [17]. Reference [18] introducedLSSVMtopredict theannual loadinChinawith therollingmechanism.Thegoodresults verifiedtheapplicabilityofLSSVMinloadforecasting.Reference [19]built ahybridmodel integrated LSSVMwithcuckoosearchalgorithm(CS) forshort-termloadforecasting. Thefindings indicatedthat thisproposedtechniquecouldobtaingoodpredictionresults. Remarkably,LSSVMhasnotyetbeen appliedto loadforecastingforquick-changee-buschargingstation. The learningandgeneralization abilityofLSSVMmodelhingesontheselectionof twoparameters,namely, regularizationparameterγ andkernelparameterσ2.Asaresult, it isnecessary toutilizeanappropriate intelligentalgorithmto determinethesevalues. Thecommonlyemployedoptimizationmethods includegeneticalgorithm (GA) [20], particle swarmoptimization (PSO) [21], CS [22] andbat algorithm (BA) [23]. However, GAhas thedisadvantagesofprematureconvergence, complexcomputation, smallprocessingscale, poorstabilityanddifficulty incopingwithnonlinearconstraints. Thepooraccuracyof local search ofPSOcannot fully satisfy theneedofparameteroptimization inLSSVM.The shortcomingofCS 319