Page - 144 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2080 The lackofavailabilityofall temperatureseriesaffects theaccuracyof thesystem.Bothmodels have been tested by including only one series of data and then adding the rest one at a time. This experimentallows todeterminewhichmodel canperformbestunder scarce informationand whichcanbenefit themost fromaricherdataset. 2.2.3. TemperatureTreatment As itwasaforementioned, temperaturehasanon-linear relationwithelectricityconsumption, asbothhighandlowtemperaturecausesanincrease indemand.Toillustrate this,Figure3showsa plotof theaverage loadonregulardaysat18hagainst theaverage temperatureof theday. Therefore, inorder for the forecastingengine tocapturesuchbehavior, itmayrequireapre-processingof thedata. Figure3.Scatterplotofnational loadat18hagainst thedailyaverage temperature inMadrid. One commonapproach to this is using a technique calledHeatingDegreeDays (HDD) and CoolingDegreeDays (CDD). This technique linearizes the temperature load relation bydefining thresholdforhighandlowtemperaturesandsplitting theseries intoonethataccounts forcolddays andanother thatdoes forhotdays. TheCDDandHDDseriesaredescribedinEquations (3)and(4) andtheyare furtherdiscussed in [34]. CDDd= { Tmed,d−THhot, if Tmed>THhot 0, otherwise (3) HDDd= { THcold−Tmed,d, if Tmed<THcold 0, otherwise (4) whereTmed,d is theaverage temperatureofdayd,THhot andTHcold are the thresholds forhotandcold daysandCDDdandHDDd are thevaluesofeachseries fordayd. This techniquerequires the thresholds tobeproperly tunedtoeach location’seffectonthe load. This optimization process is described in [39] and the optimal threshold for each zone has been calculated.However, therobustnessofeachmodelagainst thevariationof thesevalueshasbeentested byintroducingvariationsofupto12degreesoneachthreshold. 2.2.4.NeuralNetworkSize,RedundancyandComputationalBurden Accordingto theselectedtopologyshowninFigure1,partof theconfigurationof thenetwork is theselectionof thenumberofneurons in thehidden layer. Thecomplexityof thenetwork is related to 144