Page - 120 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3283 powerconsumptionreservemargin [5]. It includesdailyelectrical load,highestorpeakelectrical load, andveryshort-termloadforecasting(VSTLF).Generally,STLFisusedtoregulate theenergysystem from1htooneweek[7].Accordingly,daily loadforecasting isused in theenergyplanningfor the nextonedaytooneweek[8,9]. Ahigher-educationbuildingcomplex, suchasauniversitycampus, iscomposedofabuilding clusterwithahighelectrical load,andhence,hasbeena largeelectricpowerdistributionconsumer in Korea [10–12]. In termsofoperationalcostmanagement, forecastingcanhelp indeterminingwhere, ifany, savingscanbemade,aswellasuncoveringsysteminefficiencies [13]. In termsofscheduling, forecastingcanbehelpful for improving theoperational efficiency, especially inanenergystorage system(ESS)orrenewableenergy. Forecasting the electrical load of a university campus is difficult due to its irregular power consumptionpatterns. Suchpatternsaredeterminedbydiverse factors, suchas theacademicschedule, social events, andnatural condition. Evenon thesamecampus, theelectrical loadpatternsamong buildingsdiffer,dependingontheusageorpurpose. For instance, typical engineeringandscience buildingsshowahighpowerconsumption,whiledormitorybuildingsshowalowpowerconsumption. Thus, to accurately forecast the electrical loadofuniversity campus,wealsoneed to consider the buildingtypeandpowerusagepatterns. Byconsideringpowerconsumptionpatternsandvariousexternal factors together,manymachine learningalgorithmshaveshownareasonableperformanceinshort-termloadforecasting[3,4,6,8,14–17]. However, evenmachine learningalgorithmswithahigherperformancehavedifficulty inmaking accuratepredictionsatall times,becauseeachalgorithmadoptsadifferentweightingmethod[18]. Thus,wecansee that therewillalwaysberandomnessor inherentuncertainty ineveryprediction[19]. For instance,mostuniversitybuildings inKoreashowvariouselectrical loadpatternswhichdiffer, dependingontheacademiccalendar. Furthermore,Koreacelebratesseveralholidays,suchasBuddha’s birthdaysandKoreanThanksgivingdays, calledChuseok,duringthesemester,whicharecountedon the lunarcalendar. Since thecampususuallyremainsclosedontheholidays, thepowerconsumption of thecampusbecomesvery low. Insuchcases, it isdifficult forasingleexcellentalgorithmtomake accuratepredictions forallpatterns.However,otheralgorithmscanmakeaccuratepredictions inareas where thepreviousalgorithmhasbeenunable todoso. For thispurpose,agoodapproach is toapply twoormorealgorithmstoconstructahybridprobabilistic forecastingmodel [14].Manyrecentstudies haveaddressedahybridapproachforSTLF.Abdoosetal. [20]proposedahybridintelligentmethodfor theshort-termloadforecastingof Iran’spowersystemusingwavelet transform(WT),Gram-Schmidt (GS), andsupportvectormachine (SVM).Dongetal. [21]proposedahybriddata-drivenmodel to predict thedaily total loadbasedonanensembleartificialneuralnetwork. Inasimilarway,Leeand Hong[22]proposedahybridmodel for forecasting themonthlypower loadseveralmonthsahead basedonadynamic and fuzzy time seriesmodel. Recently, probabilistic forecastinghas arisenas an active topic and it couldprovide quantitativeuncertainty information,which canbeuseful to manage its randomness in thepower systemoperation [23]. Xiao et al. [18]proposednonegative constraint theory (NNCT)andartificial intelligence-basedcombinationmodels topredict futurewind speedseriesof theChengderegion. Juradoetal. [24]proposedhybridmethodologies forelectrical load forecasting inbuildingswithdifferentprofilesbasedonentropy-based feature selectionwith AImethodologies. Fengetal. [25]developedanensemblemodel toproducebothdeterministicand probabilisticwindforecasts thatconsistsofmultiplesinglemachine learningalgorithmsinthefirst layerandblendingalgorithmsinthesecondlayer. Inourpreviousstudy[12],webuiltadailyelectrical loadforecastmodelbasedonrandomforest. In this study, to improvethe forecastingperformanceof thatmodel,wefirstclassify theelectrical loaddatabypatternsimilarityusingadecisiontree. Then, weconstructahybridmodelbasedonrandomforestandmultilayerperceptronbyconsideringsimilar timeseriespatterns. Therestof thispaper isorganizedas follows. InSection2,we introduceseveralpreviousstudies onthemachine learning-basedshort-termloadforecastingmodel. InSection3,wepresentall thesteps 120