Page - 159 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2038 data of the Polish power system,whereas in [22] the samemethod is used to predict residential energyconsumption. In [23], theauthorspropose threedifferentmethodsforensembleprobabilistic forecasting. The ensemblemethods are derived from seven individualmachine learningmodels, whichincluderandomforest,amongothers,andit is testedinthefieldofsolarpowerforecasts.Onthe otherhand, in [24] theauthorsestablishanovelensemblemodel that isbasedonvariationalmode decompositionandtheextremelearningmachine. Theproposedensemblemodel is illustratedwhile usingdata fromtheAustralianelectricitymarket. Themainobjectiveof thispaper is to illustrate theperformanceofdifferentensemblemethods forpredicting theelectricityconsumptionof someuniversitybuildings, analyzing theiraccuracies, relevantpredictors, computational times,andparameterselection. Besides,weapply theprediction results to thecontextofDirectMarketConsumers (DMC) in theSpanishElectricityMarket. InSpain,electricitypriceseemstobeaboveourEuropeanneighbors,mainlydueto theenergy productionmixandtheweakelectrical interconnectionswith theCentralEuropeanElectricitySystem andMarkets, but consumers cando little about that. Therefore, it is quite challenging forSpanish consumers toreduce thiscost. Currently,ahighvoltageconsumer (voltagesupplygreater than1kV), which is the caseof a small campusuniversity, canopt for two typesof supply: captive customer (price freelyagreedwitharetaileroraprovider)andDirectMarketConsumer(alsocalledqualified customer), takingadvantageof theoperationof thewholesalemarkets thatare involvedintheSpanish ElectricitySystem.The literatureconcerningthe topicofDMCisnearlynon-existentandit reduces to someofficialwebpages, suchas [25,26]. Inorder toparticipateasaDMCintheElectricityMarket, thecustomerneeds toevaluatehis load requirements,withroughly twodays inadvance.Anotherobjectiveof thispaper is toevaluate the savings that theuniversitywouldhaveparticipatingasaDMC, takingthe48-h-aheadpredictionsof oneof theensemblemethodsanalyzed. Themaindifferences among thepresentpaper and thepreviousonesdealingwith theusing ofensemblemethodsfor forecastingporpoises (forexample, ref. [27]employs thegradientboosting method formodeling the energy consumptionof commercial buildings) are the following: in the present paper, we propose the XGBoostmethod as a useful tool for amedium-size consumer to purchase theelectricitydirectly in thewholesalemarket. For that, adifferentpredictionhorizon(48h ahead) is considered and thenewpredictors are needed. Indeed,wehighlight the importance of calendarvariables (distinguishingdifferent typesof festivities) for thecaseofelectricityconsumption inuniversitybuildings. This approachallowsus toevaluate the savingsof thiskindof customers participatingasDirectMarketConsumers.Anothernoveltyrespect topreviouspapers is theusingof conditional forestasanensemblemethodtoget loadpredictions,aswellas theconditional importance measure toevaluate therelevanceofeachfeature. Firstly, inSection2, fourensemblemethodsbasedonregression treesaredescribed. Section3 depicts the customer in study (a small campusuniversity) and thedata, discusses the parameter selection for each ensemblemethod aswell as other relevant aspects and it shows theprediction results. Finally, inSection4, theeconomicsavingofasmall campusuniversity iscomputedwhenit participatesasaDirectMarketConsumer insteadofacquiringtheelectricpowerfromatraditional retailer.Note that it isnotanenergyefficiencystudy, theeconomicsaving isgiven justbythe typeof supply: retailorwholesalemarket. 2. EnsembleMethodsBasedonRegressionTrees Taking intoaccount the typeofdata in theanalysis (continuousdatacorrespondingtoelectricity consumption), in this section,wewill focusondescribing tree-basedmethods for regressionandsome relatedensemble techniques.However,decisiontreesandensemblemethodscanbeappliedtoboth regressionandclassificationproblems. Theprocess of building a regression tree canbe summarized in two steps: firstly,wedivide thepredictorspace intoanumberofnon-overlappingregions (forexample J regions),andsecondly, 159