Seite - 264 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1678 the forecasts slightlyonaverage. The largest improvement is seenonholidayswhere theerrorcanbe reducedby5%,comparedtoonlyusinggenericcalendardata. Figure 7. Forecast performance of the SVRmodel on the year 2016 using realweather forecasts, calendar,andholidaydata. Thesecondandthirdgroupofbinsrefer toweekdaysandweekends that arenotalso includedinholidays.Holidaysrefer toalldays thatareobservances,nationalholidays, orschoolholidays (seeTable1). 4.Conclusions Wehavetestedheat loadforecastswithhorizons from15hto38h, relevant fordistrictheating productionplanning considering theday-aheadelectricitymarket. Theworkwasbasedonseven yearsofheat loadandweatherdata for the largedistrictheatingsystemofAarhus,Denmark. Inorder tomeasure the forecastperformance thatcanrealisticallybeexperienced inactualoperation,weused blindtestingonawholeyearwithrealweather forecasts. Threemachine learningmodelshavebeentested: anordinary least squaresmodel,amultilayer perceptron,andasupportvector regressionmodel. TheSVRmodelperformedbest,beatingtheOLS model by a largemargin and theMLPmodel by a smallmargin. All themodelswere trainedon laggedheat loaddataandweatherdata. The forecastperformancecouldbesignificantly improvedby includinggenericcalendardata, suchasmonth,weekday,andhourofday.Asmaller improvementof the forecastscouldbegainedbysupplyingthemodelswith localholidaydata includingobservances, national holidays, and school holidays. This improvementwasmost significant on holidays and weekends. Localholidaydatacanbedifficult andtime-consumingtoobtain,butmerely including laggedheat load,weather,andgenericcalendardatacanprovideagoodoverall forecastperformance. TheSVRmodelusingweather, calendar,andholidaydatahadthebestperformance. Theroot mean square errorwas 29.3MW, and themeanabsolutepercentage errorwas 6.4%. This forecast modelbeatallothermodels thatwehaveseenfor theAarhussystem.Thecommercial forecast system, currently inoperation in theAarhusdistrictheatingsystem,hadanRMSEof41.9MW,andaMAPEof 8.3%onthe testyear. Including localholidaydatashowedonlyminoroverall improvements in forecastperformance, andincludingnewdata types in forecastmodels requiresacarefulevaluationof the trade-offbetween 264