Seite - 258 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1678 the forecastsproducedfor4May.Onlytheheat loaduptothe timeof the forecastwasusedas input to produce the forecast. Realweather forecastswereusedasweather inputs for4May,asopposedto thehistoricalweatherdatausedfor training. It is clearhowthe three forecastmodelsproducesimilar, yetdistinct forecasts. On4May, theMLPmodelappears toproduce thebest forecast, especially in themorning. Figure4summarizes theperformanceof the threemodels in the threedifferentdatascenarios. The toppanel showsthe forecastperformance thatcouldbeachievedifweather forecastswere100% accurate, simulatedbyusinghistoricalweatherdata. Thebottompanel showstheperformanceusing realweather forecasts.Comparingthe threedatascenarios,wesee thebenefitof includingdifferent data types in themodeling. In thefirst scenario, only laggedheat loadandweatherdataareused as input. In thesecondscenario,genericcalendardata is includedaswell, andinthe thirdscenario, localobservances, nationalholidays, andschoolholidaysarealso includedas inputs to themodel. Includingcalendardatasignificantly improvesperformance, comparedtoonlyusingweatherdata. Extendingthe inputdatawithholidaydataaswell results inanadditional,butsmall improvement comparedtousinggenericcalendardataonly.Obtainingthe localholidaydatacanbe laboriousor impossible, so it ispositive toseegenericcalendardatayieldingcomparableresults. It ismucheasier toapply thesemodels toawiderangeofdistrictheatingsystemsaroundtheworld if it canbedone withoutcollecting localholidaydata. Figure3.Example forecasts for4May2016. The forecastswereproducedon3Mayat10:00andbased onrealweather forecasts, calendar,andholidaydata. Figure4allowsforcomparisonof theperformanceof the threemachine learningmodelsaswell. TheOLSmodelstandsoutbyperformingsignificantlyworsethantheothertwomodels inallscenarios. TheOLSmodel has a rootmean square error of 38.9MW, compared to 31.1MWand29.3MWfor theother twomodelswhenusingrealweather forecasts, calendar,andholidaydata (bottompanel). Thepoorperformanceof theOLSmodel canbeattributed to its linear structure. The relationship betweentheoutdoortemperatureandtheheat loadinatemperateclimate isnonlinear. Thiscauses the linearmodel toperformpoorlyduringsummerbyundershooting theheat loadandoverestimating its variance. The twononlinearmodels,MLPandSVR,performsimilarly in thesescenarios. TheSVR 258