Seite - 47 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2019,12, 164 2.1. LinearModels Linearmodelsgivecontinuousresponsewhich isa functionor linearcombinationofoneormore predictionvariables. Thesemodelsdependonthesynthesisofall featuresofaproblemthat ismoreor lesssolvedbycomplexequations. Examplesof thesemodels includespectraldecomposition-based models,ordinary least square-basedmodels,ARMA,etc. Since thepredictionofdemandiscomplex duenon-linearities, the linear forecastmodelspredictwithhighrelativeerrorsdueto their inability to mapthecomplexrelationshipbetweeninputandoutput. Thus,developmentof linearmodels ishighly challenging. Furthermore,Haganetal. [10]highlightedtherelative limitation(s)of linearmodelsas comparedtonon-linearmodels. Therefore, this researchwork is focusedtowards thediscussionof non-linearmodelsonly. 2.2.Non-LinearModels Whentheobservationaldata ismodeledbynon-linearcombinationofoneormoreprediction variables, themodel is said tobenon-linear. Todescribe therelationbetweenresidualandperiodical components,BunnandFarmer [32] realize/concludetheabilityofnon-linearmodels toovercomethe limitation(s)of linearmodels. Inreference [9], thenon-linearmodelsare furthercategorized intofive classes: (i) supportvectormachine-basedmodels; (ii)Markovchain-basedmodels; (iii)ANN-based models; (iv) fuzzyneuralnetwork-basedmodels; and(v)stochasticdistribution-basedmodels. These modelsarediscussedas follows. (i)Supportvectormachine-basedmodels: Inreference[11],Niuetal. proposesupportvectormachine andantcolonyoptimization-based load-forecastingtechniqueforanSG.Theauthorsuseantcolony optimization techniqueforpreprocessingof the inputdata. In thispaper, systemminingtechnique is usedfor featureselection. Theselectedfeaturesare fed into the forecasterwhich isasupportvector machine-basedpredictor.Another importantworkhasbeenpresentedbyLietal. in [12]. Thisvaried versionof theauthors is least squares-basedsupportvectormachine. Similarly, reference [13]models thecyclicnatureofdemandbysupportvectormachine-based linear regression. Inconclusion, the supportvectormachine-basedworksarebetter in termsofaccuracy;however,developmentof these models ishighlychallengingduetohighcomplexity. (ii)Markov chain-basedmodels: Subject to robustnessofDALF forecast strategy, authors in [14] propose aMarkov chain-based strategy. This stochastic strategy aims to tackle load-time series fluctuationsassociatedwithenergyconsumptionofusersinaheterogeneousenvironment. TheMarkov chains are used to predict the future duty cycles of appliances. The technique is robust due to theirmemorylessnature (predictedpatternonlydependson the current states; past states arenot considered). In reference [15],Markov chainMonteCarlomethod isused tomodel the switching patternofhouseholdappliances. Insimulations, theyconsider100households foroneweak.However, thismodel limited inscopeas itapplies tosituations in theNetherlandsonly.Anotherworkin[16] proposesexplicitdurationhiddenMarkovmodelalongwithdifferentialobservation-basedmodel to predict individual loadofappliances. Theauthorscollect theaggregatedpowersignalsbyordinary smartmeters. ThememorylessnatureofMarkovchainsnotonlymakes theDALFstrategyrobustbut alsorelatively lesscomplex incomparisonto theaforementionedtechniques.However, thememory lessnatureofMarkovchainsalsohasadrawback; lessaccuracy. (iii)ANN-basedmodels: ANNs learn fromexperience/training to predict future valueswhile being fedwith relevant input information. The advantages of thesenetworks includebut arenot limited to self-organization, adaptive learning, fault tolerance, ease of integrationwith existing network/technology,andreal timeoperation. Theabilities togeneralizeandtocapturenon-linearity incomplexenvironmentsmakeANNsveryattractive inproblemsof loadforecasting. Thereare two basicarchitecturesofANN; feed forwardand feedback. The formeronecarries informationfrominput to outputviahiddenlayer in forwarddirectiononly, i.e., the informationofeach layer is independent fromthatof theothers. FeedforwardANNsarewidelyusedforpatternrecognitionandforecasting problems. The lateronecarries information inbothdirections, forwardandfeedback, suchthat the 47