Page - 158 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

energies Article LoadForecastingforaCampusUniversityUsing EnsembleMethodsBasedonRegressionTrees MaríadelCarmenRuiz-Abellón1,AntonioGabaldón2,* ID andAntonioGuillamón1 1 DepartmentofAppliedMathematicsandStatistics,UniversidadPolitécnicadeCartagena, 30202Cartagena,Spain;maricarmen.ruiz@upct.es (M.d.C.R.-A.); antonio.guillamon@upct.es (A.G.) 2 DepartmentofElectricalEngineering,UniversidadPolitécnicadeCartagena,30202Cartagena,Spain * Correspondence: antonio.gabaldon@upct.es;Tel.:+34-968-338944 Received: 6 July2018;Accepted: 1August2018;Published: 6August2018 Abstract:Loadforecastingmodelsareofgreat importance inElectricityMarketsandawiderange of techniqueshavebeendevelopedaccordingto theobjectivebeingpursued. The increaseofsmart meters indifferentsectors (residential, commercial,universities, etc.) allowsaccessing theelectricity consumption nearly in real time andprovides those customerswith large datasets that contain valuable information. In thiscontext, supervisedmachine learningmethodsplayanessential role. Thepurposeof thepresentstudyis toevaluate theeffectivenessofusingensemblemethodsbasedon regression trees inshort-termloadforecasting. To illustrate this task, fourmethods (bagging, random forest, conditional forest, andboosting)areappliedtohistorical loaddataofacampusuniversity in Cartagena(Spain). Inaddition to temperature, calendarvariablesaswellasdifferent typesof special daysareconsideredaspredictors to improvethepredictions. Finally,arealapplicationtotheSpanish ElectricityMarket isdeveloped: 48-h-aheadpredictionsareusedtoevaluate theeconomical savings that theconsumer(thecampusuniversity)canobtain throughtheparticipationasadirectmarket consumer insteadofpurchasingtheelectricity fromaretailer. Keywords:ElectricityMarkets; loadforecastingmodels; regressiontrees; ensemblemethods;direct marketconsumers 1. Introduction Loadforecastinghasbeenatopicof interest formanydecadesandthe literature isplentywith awide variety of techniques. Forecastingmethods can bedivided into three different categories: time-seriesapproaches, regressionbased,andartificial intelligencemethods (see [1]). Among the classical time-series approaches, the ARIMAmodel is one of themost utilized (see [2–5]). Regression approaches, see [2,6], are alsowidely used in the field of short-term and medium-termloadforecasting, includingnon-linearregression[7]andnonparametric regression[8] methods. Recently, in [9] the authorsuse linearmultiple regression topredict thedaily electricity consumption of administrative and academic buildings located at a campus of London South BankUniversity. Severalmachine learning or computational intelligence techniques have been applied in the fieldofShortTermLoadForecasting. Forexample,decisiontrees [10],FuzzyLogicsystems[11,12], andNeuralNetworks [13–20]. In thispaper,wepropose theusingofaparticular setof supervised machine learningtechniques (calledensemblemethodsbasedondecisiontrees) topredict thehourly electricityconsumptionofuniversitybuildings. Ingeneral,anensemblemethodcombinesasetofweak learners toobtainastrong learner thatprovidesbetterperformance thanasingleone. Fourparticular casesofensemblemethodsarebagging, randomforest, conditional forest, andboosting,whichare described inSection 2. There some recent literature regarding randomforest and short-term load forecasting: forexample, in [21] theauthorsuse randomforest topredict thehourlyelectrical load Energies2018,11, 2038;doi:10.3390/en11082038 www.mdpi.com/journal/energies158