Seite - 140 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2080 Determiningthedaily,weeklyandseasonalpatternsofconsumers isat therootof the load-forecasting problem. Pattern recognition techniques stemfromthecontextof computervisionandfromthere, theyhaveevolvedtoapplications inallfieldsofengineering informsofdifferent typesofArtificial Intelligence. These techniques (AI)havegainedattentionover the last 20years. AIoffersavariety of techniques that generally require the selection of certain aspects of their topologybut they are able tomodelnon-linearbehavior fromobservingpast instances. The termrefers tomethods that employArtificialNeuralNetworks [12–16],FuzzyLogic [13,15,17–20],SupportVectorMachines [21] orEvolutionaryAlgorithms[15,17,22–24]. Hybridmodelsare those that combine theuseof twoor more techniques in the forecastingprocess. Thesearesomeexamples that includesomeof thealready mentioned[15,23,25–27].OtherapplicationofpatternrecognitionandAItechniques toSTLFinclude smaller scalesystems,whichpresent theirownspecificities [28,29]. Thepreviousparagraph focused solely on the forecasting engineused to calculate the actual forecast, as thispartusuallyreceives themostattention.However, it isnot theonlykeyaspectof the forecastingproblem. In [30], it isproposedastandardthat includes5stages thatneedtobeproperly addressed inorder toobtainaccurate forecasts: • DataPre-processing: Data normalizing, filtering of outliers anddecomposition of signals by transforms. This lastaspecthasreceivedsignificantattentionrecently [23,24,31,32]. • InputSelection:Analysisof theavailable informationandofhowtheforecastingenginesprocess this informationbest. In [33], anexampleofhowtodeterminewhichvariableshouldbe included isshown.The informationaboutspecialdays isalso includedinthisstage, relevantattempts to determinethebestwaytoconvert typeofdayinformationtovalid input to the forecastingengine are foundin[18–20,34–36]. • TimeFrameSelection:Refers todeterminingwhichperiodshouldbeusedfor training. In [16], a timeschemeincludingsimilardays isproposed. In thispaper, this issuewillbeaddressedby determininghowtheavailabilityofhistoricaldataaffects theaccuracyof forecastscarriedoutby different forecastingengines. • LoadForecasting:Refers to the forecastingengine. • DataPost-Processing:De-normalizing, re-composition,etc. To sum up, it is also relevant to mention examples of real world applications [37–39]. Thepublishingofmodels thatarevalidatedthroughactualusebythe industry insteadof throughlab conditions isespecially important for theadvancementof thefield[2]. The referred examples contain descriptions of particular forecastingmodels that are usually described by defining their input and the inner workings, topology, configuration and other characteristics of the forecasting engine. They also include the results of themodelwhen it has beentestedonaspecificdatabaseandforacertainperiodof time. Thismethodologyhasprovideda widevarietyofmodels for the industryandscientificcommunity tochoose fromforanyparticular application.However, ithasprovidedverylittle informationonhowtocompareeachmethodandhow todeterminethestrongandweaksuitsofeachtechnique. The lackofanalysisof thecharacteristics of thedatabase, and insomecases theuseof testingperiodsshorter thana fullyear,makes itvery difficult for thereader toapriorideterminewhichof theproposedmodelswouldsuitbest theirown personalcase. This issuehasbeen treated in [40,41], inwhich theauthorsproposea certainmethodology to adoptdifferent techniquesdependingontheforecastingproblem.Thesepapers includeananalysis of the loadprior to the actual forecastingprocess. However, they only test one technique for the forecastingengine. In [42] the issueofpredictabilityofdatabases isaddressedtoprovideabenchmark indicator thatcouldprovideafaircomparisonamongresultsofdifferentmodelsondifferentdatabases thatmayormaynot be similarly affectedby the same factors (temperature, social activities . . . ). This typeof informationalongwiththestandardizationproposedin[30]wouldbeuseful todetermine 140