Seite - 141 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2080 thecharacteristicsofaspecificproblemandthefeaturesofeachmodelavailable thatbestaddresses thesubjectathand. Consequently, there is consensus thatageneral solutiondoesnotexistandthat theSTLFproblem does not have a “one-size-fits-all” fix. Nevertheless, the objective of this paper is to provide a comparisonbetweentwoof themostcommonforecastingengines: theautoregressivemodel (AR)and theNeuralNetwork(NN).Thegoal is todeterminehowagivensetofconditionsandconfiguration parametersaffect theaccuracyofeachtechnique(ARandNN)anduse this information todefinetheir strongandweakpoints. Themethodology aims to determine the circumstances underwhich each of the forecasting enginesperformsmoreaccurately. Theconditionsof the forecasts: historicaldataavailable, sourcesof temperature information, computationalburden,maintenanceneeded .. . aremodifiedtodetermine howeachof themaffectseach technique. Inaddition, theperformanceresultsareanalyzed in termsof typeofdays (cold,hot, specialdays) inorder tobetterassesswhetheroneof the forecastingengines performsbetteronacertain typeofday. Thispaperprovidesresults fromarealapplicationusingtwodifferent techniquesunder thesame setofconditions. Theseresultsareclassifiedbythe typeofdayto facilitate theanalysis. Theobtained resultsprovideproof thatNNmodelsaremore reliablewhenmeteorological information is scarce (only few locations are available) orwhen it is notproperlypre-processed. Nevertheless, theNN requiresa largerhistoricaldatabase tomatchtheaccuracyof theARmodel. Theoverall results show that each technique is better suited for specific typesofdays, butmore importantly, that there are conditionsunderwhichonetechniqueclearlyoutperformstheother. Section2containsthedescriptionoftheforecastingenginesthatarecompared, theparametersand conditionsunderwhichthe forecastingenginesare testedandthecategorizationof typeofdaysused tocompare theresults.OnSection3, thecharacteristicsof thedatausedareexplained: characteristics of the load,meteorologicalvariablesandtheir treatmentandinformation todetermine the typeofday. Section4 includes theresultsobtainedonthe tests: a revisionofeachparameterandhowitsvariation affects theperformance tobothforecastingengines. Finally,Section5 includesabrief conclusionthat summarizes themost relevantaspectsof theresults. 2.Methodology Thissectionprovidesadetaileddescriptionof theanalyzedforecastingtechniques, theconditions underwhichtheyare testedandtheclassificationof theresultsusedtodrawconclusions. 2.1. ForecastingModels BothforecastingmodelsunderanalysisareextractedfromtheSTLFsystemcurrentlyworking at Red Eléctrica de España (REE), the Transport System Operator in Spain. They have been thorughlydescribed in [39], andhave been running on theREEheadquarters for over twoyears now.Bothforecastingenginesuse thesamedatafilteringsystemtodiscardoutliers,usuallycaused bymalfunctioningof thedataacquisitionsystems. Theforecastingschemeprovidesa forecastevery hour thatcontains the forecastedhourlyprofile for thecurrentdayandthenextninedays. Internally, each hour is forecasted separately by different sub-models. Therefore, each fullmodel includes 24sub-models to forecast the loadprofileofa fullday,anddifferentsubmodelsareuseddependingon howdistant in the future the forecasteddayis. Tosimplify thecomparison, themetric thatwillbeusedasreference is theerrorof the forecast madeat9a.m. for the full24hof thenextday. This forecast is themost relevant forREEas it is the onethatservesasabase foroperationandplanning. The input foranyof thesubmodels isavector thatcontains the latest load informationavailable, temperature forecasts, annual trendsandcalendar information. Thisdatawillbe furtherdiscussedon thenextsection,but it is thesameforboth techniquesARandNNthatarenowexplained. 141