Seite - 259 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1678 modelhas thesmallesterror,andthe focus in therestof thispaperwillbeontheSVRmodelusing weather, calendar,andholidaydata. Figure4.Rootmeansquareerrorof the three forecastmodelsOLS,MLP,andSVRontheyear2016. The toppanel (a) shows theerrorusinghistoricalweatherdata to simulate 100%accurateweather forecasts. Thebottompanel (b) showstheerrorusingrealweather forecasts. 3.1. TheValueof ImprovingWeatherForecasts Figure 4 has two panels. The top panel shows the forecast errors that could be achieved if weather forecastspredicted themeasuredweather completelyaccurately. Thishasbeensimulated byallowingthemodels touseactualmeasuredweatherdata, insteadofweather forecastsas input whenproducingthe loadforecast. Thetoppanel reflects thescenario in thewhichfutureweather is known.Thebottompanel showstheresults in thecasewherereal forecastdatahasbeenused instead. This is the actual forecast performance that canbeachieved in anoperational situation, given the currentqualityofweather forecasts. Havingaccess toweather forecastswithoutpredictionerrors could, inaperfectworld, reduce theerror from29.3MWto25.2MWin the forecasts fromthebest model.Whileanerror reductionof4.1MWisastart,perfecting theweather forecastonlyshaves14% off theerror. Theremainderof the loadforecasterrorhasothercauses thanweather forecasterrors,a result thatwasalso found in [24],whereensembleweatherpredictionswereused toquantifyheat loadforecastinguncertainty. TheOLSmodelusingonlyweatherdataandlaggedheat loaddoesnotperformnotablydifferent onhistoricalweatherdatacomparedtorealweather forecasts. ThiscanbeexplainedbytheOLSmodel attributinggreaterweight to the laggedheat loadcomparedto theweather,because therelationship betweentheheat loadandtheweather isnonlinear. 259