Page - 262 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1678 Figure6.Histogramsfortheforecasterrorof theSVRmodelontheyear2016usingrealweatherforecasts, calendar, andholidaydata. Thedistributionof the forecast error isdepicted for eachmonth in the yearalongwith the10%and90%quantiles. ThenumberofbinswaschosenusingScott’s rule [25] withineachmonth.Apositiveerror indicates that the forecastwas toohigh,anegativeerror that it was too low. Fromthehistograms inFigure6, it is also clear that theerrordistributionsarenot completely symmetricaround0. In January, for instance, thedistribution isshiftedslightly to thepositive,and inApril it is shifted to thenegative side. The forecast appears tobebiaseddifferently indifferent months. Themeanerror foreachmonth(ME) is showninTable2. Thebiascanbeas largeas20.5MW, 262