Page - 48 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2019,12, 164 informationof each layer isdependent on that of theothers. FeedbackANNsare appropriate for complexandtimevaryingproblems[33–35].Ontheotherhand,thelearningmodesofANNsfallunder threecategories: supervised [36],unsupervised [37],and re-enforced [38]. In thefirstcategory, theANN attempts tominimizeminimumsquareerror (MSE) forknowntargetvector (i.e., the input/output vectorsare specified). Foragiven input/output, error is calculatedbetweenoutputand the target values. ThiserrorisusedtoupdatetheweightsandbiasesoftheANNtominimizetheMSEtoacertain threshold. In the secondcategory, theANNdoesnotneedexplicit targetdata. Thesystemadjusts itsoutputbasedonself-learningfromdifferent inputpatterns. In the thirdcategory, theconnections betweenANsarereinforcedevery timetheseareactivated. Since this researchwork is limited inscope tosupervised learningonly,wediscusssomeof these latest/relevantworksas follows. In reference [27], authorspresentahybrid techniquesubject to short-termprice forecastingof SGs. Thishybrid techniquecomprises twosteps; feature selectionandprediction. In thefirst step, amutual information-basedtechnique is implementedtoremoveredundancyandirrelevancyfrom the input load-time series. In the secondstep,ANNalongwithevolutionaryalgorithm isused to forecast the time series of the future load. In this process, the authors assume sigmoid activation functionforartificialneurons (ANs) ,andLevenburg-Marquardtalgorithmfor training. Inaddition, theauthorsfine-tunesomeadjustableparametersduring thefirstandsecondstepsviaan iterative searchprocedurewhich ispartof theirwork. Subject to forecastaccuracy, this technique isefficient as it embeds various techniques; however, the cost paid is high execution time. In reference [28], theauthors investigatestochasticcharacteristicsofSG’s load.More importantly, theauthorspresent abi-levelDALF technique for SGs. In thefirst/lower level,ANNandevolutionaryalgorithmare implementedto forecast the future load/pricecurve. In thesecond/upper level, anEDEalgorithm is implementedtofurtherminimize thepredictionerrors. Effectivenessof thiswork is reflectedvia MATLABsimulationswhichdemonstrate that theproposed strategyperformsDALF inSGswith a reasonable accuracybypaying the cost of highexecution time. Thehybridmethodology in [39] completes theDALFtask infoursteps: (i)dataselection; (ii) transformation; (iii) forecast;and(iv)error correction. Instepone, somewell-knowntechniquesofdataselectionareusedtominimize thehigh dimensionalitycurseof input load-timeseriescharacteristics. Steptwodealswavelet transformation of the selectedcharacteristicsof input load-timeseries toenable redundancyand irrelevancyfilter implementation. Followedbystepthree,whichusesANNandatrainingalgorithmsubject toDALFin SGs.More importantly, theychoosesigmoidactivationfunctionforANsduenon-linearcapturability. Finally, errorcorrectingfunctionsareusedinstepfour to improvetheproposedDALFmethodology intermsofaccuracy. Insimulations, thismethodologyis testedagainstpracticalhousehold loadwhich demonstrates that thismethodology isverygoodfor improving theaccuracybypayingthecostof highcomplexity.Anothernovel strategy ispresented in [40] topredict theoccurrenceofpricespikes inSGs. Theproposedstrategyuseswavelet transformation for input featureselection. AnANNis thenusedtopredict futurepricespikesbasedonthe trainingof theselected inputs. (iv)Fuzzyneuralnetwork-basedmodels:Dovehetal. [21]present fuzzyANN-basedmodel for load forecasting. In theirwork, the inputvariablesareheterogeneous. Theyalsomodel theseasonaleffect via a fuzzy indicator. In reference [22], the authorspresent a self-adaptive load-forecastingmodel forSGs. Tocorrelatedemandprofile informationandtheoperationalconditions,aknowledge-based feedbackfuzzysystemisproposed. Foroptimizationoferror,amultilayeredperceptronANNstructure is usedwhere training is done via back propagationmethod. Someother hybrid strategies such as [23,24] focusonfuzzyANNaswell.Wang[23]presentselectricdemandforecastingmodelusing fuzzyANNmodel,whereas,Cheetal. [24]presentanadaptive fuzzycombinationmodel.Cheetal. iteratively combinedifferent subgroupswhile calculating fuzzy functions forall the subgroups. A fewmoreworkscombiningfuzzyANNwithotherschemesarepresented in [25,26]. Subject to fuzzy neuralnetworkcontrollerdesignfor improvingpredictionaccuracy,membership functions toexpress the inference rulesby linguistic termsneedproperdefinitions. As fuzzysystems lack such formal 48