Seite - 122 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3283 adesigner’sexpertiseandexhibits lowversatility [18].Ontheotherhand, thispaperproposesdata post-processingtechniquecombinedapproaches toconstructahybridforecastingmodelbycombining randomforestandmultilayerperceptron(MLP). 3.HybridShort-TermLoadForecasting In this section,wedescribeourhybridelectrical load forecastingmodel. Theoverall steps for constructingthe forecastingmodelareshowninFigure1. First,wecollectdailypowerconsumption data, timeseries information,andweather information,whichwillbeusedas independentvariables forourhybridSTLFmodel.Aftersomepreprocessing,webuildahybridpredictionmodelbasedon randomforestandMLP.Lastly,weperformaseven-step-ahead(oneweekor145hahead) timeseries cross-validationfor theelectrical loaddata. Figure1.Ourframeworkforhybriddailyelectrical loadforecasting. 3.1.Dataset TobuildaneffectiveSTLFmodel forbuildingsorbuildingclusters, it is crucial tocollect their realpowerconsumptiondata thatshowthepowerusageof thebuildings in therealworld. For this purpose,we considered three clusters of buildingswithvariedpurposes andcollected their daily powerconsumptiondata fromauniversity inKorea. Thefirst cluster iscomposedof32buildingswith academicpurposes, suchas themainbuilding,amenities,departmentbuildings, central library,etc. Thesecondcluster is composedof20buildings,withscienceandengineeringpurposes.Comparedto otherclusters, thisclustershowedamuchhigherelectrical load,mainlyduetothediverseexperimental equipmentanddevicesusedin the laboratories. The thirdclustercomprised16dormitorybuildings, whosepowerconsumptionwasbasedontheresidencepattern. Inaddition,wegatheredotherdata, includingtheacademicschedule,weather,andeventcalendar. Theuniversityemploys the i-Smart systemtomonitor theelectrical loadinreal time. This isanenergyportalserviceoperatedbytheKorea ElectricPowerCorporation(KEPCO)togiveconsumerselectricity-relateddatasuchaselectricityusage andexpectedbill tomakethemuseelectricityefficiently. Throughthis i-Smartsystem,wecollected thedailypowerconsumptionofsixyears, from2012 to2017. Forweather information,weutilizedthe regional synopticmeteorologicaldataprovidedbytheKoreaMeteorologicalOffice (KMA).KMA’s mid-term forecast provides information including thedate,weather, temperature (maximumand minimum),andits reliability formore thansevendays. TobuildourhybridSTLFmodel,weconsideredninevariables;month,dayof themonth,day of theweek, holiday, academic year, temperature,week-ahead load, year-ahead load, andLSTM Networks. In particular, the day of theweek is a categorized variable andwepresent the seven daysusing integers1 to7accordingto theISO-8601standard[32].Accordingly,1 indicatesMonday and7indicatesSunday.Holiday,which includesSaturdays,Sundays,nationalholidays,andschool anniversary[33], indicateswhether thecampus isclosedornot.Adetaileddescriptionof the input variablescanbefoundin[12]. 122