Page - 139 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

energies Article EmpiricalComparisonofNeuralNetworkand Auto-RegressiveModels inShort-Term LoadForecasting MiguelLópez*,CarlosSans,SergioValeroandCarolinaSenabre DepartmentofMechanicEngineeringandEnergy,UniversidadMiguelHernández,03202Elx,Alacant,Spain; carsantr@gmail.com(C.S.); svalero@umh.es (S.V.); csenabre@umh.es (C.S.) * Correspondence:m.lopezg@umh.es;Tel.:+34-965-22-2407 Received: 26 June2018;Accepted: 1August2018;Published: 10August2018 Abstract:Artificial Intelligence (AI)hasbeenwidelyused inShort-TermLoadForecasting (STLF) in the last20yearsandithaspartlydisplacedolder time-seriesandstatisticalmethods toasecond row. However, theSTLFproblemisveryparticularandspecific toeachcaseand,while thereare manypapersaboutAIapplications, there is little researchdeterminingwhichfeaturesofanSTLF systemisbetter suitedforaspecificdataset. Inmanyoccasionsbothclassicalandmodernmethods coexist,providingcombinedforecasts thatoutperformthe individualones. Thispaperpresentsa thoroughempirical comparisonbetweenNeuralNetworks (NN)andAutoregressive (AR)modelsas forecastingengines. Theobjectiveof thispaper is todetermine thecircumstancesunderwhicheach model showsabetterperformance. It analyzesoneof themodels currently inuseat theNational TransportSystemOperator inSpain,RedEléctricadeEspaña(REE),whichcombinesbothtechniques. Theparameters that are tested are the availability of historical data, the treatment of exogenous variables, the trainingfrequencyandtheconfigurationof themodel. Theperformanceofeachmodel ismeasuredasRMSEoveraone-yearperiodandanalyzedunderseveral factors likespecialdays orextremetemperatures. TheARmodelhas0.13%lowererror thantheNNunder idealconditions. However, theNNmodelperformsmoreaccuratelyundercertainstresssituations. Keywords: short-termloadforecasting(STLF);neuralnetworks;artificial intelligence (AI) 1. Introduction ThedevelopmentofShort-TermLoadForecasting(STLF) toolshasbeenacommontopic in the lateyears [1–3]. STLFisdefinedas forecastingfrom1htoseveraldaysahead,andit isusuallydone hourlyorhalf-hourly. TheapplicationofSTLFincludetransportandsystemoperators thatneedto ensurereliabilityandefficiencyof thesystemandnetworksandproducers that require toestablish schedulesandutilizationof theirpower facilities. Inaddition,STLFis requiredfor theoptimizationof marketbiddingforbothbuyersandsellers inthemarket. Theabilitytoforeseetheelectricdemandwill reduce thecostsofdeviations fromthecommittedoffers. Theseaspectshavebeenespeciallyrelevant in the last decade inwhich thederegulationof theSpanishmarket followingEuropeandirectives hasbeenenforced. Inaddition, the increasingavailabilityof renewableenergysources,makes the balancingof thesystemmoreunstableas itaddsmoreuncertaintyontheproducingend.Allof these reasonsmakeSTLFacriticalaspect toensurereliabilityandefficiencyof thepowersystem. Forecastingmodelsuseseveral techniques thatcanbegroupedinStatistical,Artificial Intelligence andHybridtechniques. Statisticalmethodsrequireamathematicalmodel thatprovidetherelationship between load and other input factors. These methods were the first ones used and are still currently relevant. They includemultiple linear regressionmodels [4–6], time-series [7–10] and exponential smoothing techniques [11]. Pattern recognition is a key aspect of load forecasting. Energies2018,11, 2080;doi:10.3390/en11082080 www.mdpi.com/journal/energies139