Page - 263 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1678 withNovemberbeing theworstmonth. Septemberperforms thebestwithameanerrorofmerely 0.3MW. Varyingmonthly biases could be remediedby training separatemodels for eachmonth. However, the focus in thispaper is to investigate theeffectsof includingholidaydata tomodelhuman behavior, andtrainingmonthly forecastmodelswouldobscure theeffectsofusingholidaydata. There isalso thepossibility that theweather forecastsperformdifferentlyatdifferent timesofyear. Table2.Summaryof thehourly forecasterror foreachmonthfor theSVRmodelusingrealweather forecasts, calendar,andholidaydata.Histogramsof the forecasterrorappear inFigure6. Themonths with theworstperformanceare indicated inred, thebest ingreen. Thequantilesareevaluated inpairs, so thewidest symmetricquantile interval is consideredtheworst. RMSE ME ErrorQuantiles (MW) (MW) (MW) 10% 90% 1% 99% January 41.2 18.7 –24.0 64.1 –66.8 117.6 February 31.8 –2.2 –42.8 36.6 –91.6 61.3 March 36.9 2.0 –43.7 48.8 –80.4 89.3 April 34.2 –11.7 –52.9 27.1 –93.6 64.8 May 18.2 4.3 –14.7 25.6 –33.6 64.3 June 15.8 –3.3 –19.3 12.8 –45.3 34.0 July 10.6 5.0 -7.0 17.2 –16.0 25.8 August 14.2 6.9 –8.2 21.6 –20.1 41.6 September 14.3 0.3 –17.9 17.6 –35.0 33.3 October 25.6 4.8 –26.0 37.8 –43.5 70.0 November 38.1 20.5 –20.0 61.5 –59.1 98.1 December 43.1 12.5 –38.4 58.2 –96.7 115.0 Inconclusion, therearesignificantseasonalvariations in theperformanceof thebestheat load forecast. Theabsoluteerrorsare largest inwinterandsmallest insummer,withDecemberbeingthe hardestmonthto forecastandJulybeingtheeasiest. 3.3. TheValueofCalendarandHolidayData Thegoal of this analysis is togauge thepotential of including local holidaydata inheat load forecasts inorder tobetter capture the consumerbehavior. The reduction in theannual errorwas verysmallwhencomparingmodelswithonlygenericcalendardata tomodels including localholiday data. Thiswasclear fromFigure4b. It iswellknownamongdistrictheatingoperators thatheat load forecasts tendtoperformpoorlyonspecialoccasions, suchasChristmasorNewYear’sEve. These specialdaysarerare, so theperformanceonthosespecificdayshas little impactontheaverageannual performance(Figure4b). Improvedperformanceonspecialdays isvaluable toproductionplanners, andwhether including localholidaydatacan improveforecastperformanceonspecificdays isworth investigating inmoredetail. Figure7showstheperformanceof theSVRmodel in the threedatascenariosondifferentsetsof daysduringtheyear. “Holidays”refer toalldays thatareobservances,nationalholidays,orschool holidays. “Weekdays” includeallweekdays thatarenotalso inholidays, and“weekends” include allweekenddaysnot included inholidays. In 2016, therewere 201weekdays, 65weekenddays, and100holidays. There issignificantbenefit in includinggenericcalendardata in theforecastmodels forallday types.Onweekdays, there isnoperformance improvement togainbyincluding localholidaydata. The forecasterroronweekendscanbereducedby0.5MW.Notsurprisingly, thegreatestperformance increasecanbeobservedonholidays. Theholidayerrordecreasesby1.3MWwhenaugmentingthe modelingwith localholidaydata. Theholidayerror isgenerallysmaller thantheerror for theother daytypes. This isdueto theholidaysbeingdominatedbytheschools’ summerholidays,andtheerror isgenerally smallerduring thesummer. Summingup, including localholidaydataonly improves 263