Seite - 253 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1678 Figure1.Timeline for theheat loadforecast that is relevant for the tradingdecisions in theday-ahead electricitymarket. Everydayat10:00a forecast isproducedforeachhourof the followingday. Theanalysis in thispaper isbasedonsevenyearsofdata for the totalhourlyheat loadofAarhus, Denmark. Theyears2009,2010,2012,2013,2014,2015,and2016wereused.Unfortunately,heat load data from2011hasnotbeenavailable tous.Wedenote theheat loadinhour tbyPt. 2.1.WeatherData Theheat demanddepends strongly on theweather. Hourly outdoor temperature,wind speed, andsolar irradiationfor thesevenyearswereobtainedfrom[14].Weatherdata fromthegeographical pointN56◦2′42.24′′,E9◦59′59.95′′ in thesouthernpartofAarhuswasused.Weather forecastsof the outdoor temperature,windspeed,andsolar irradiationwereprovidedbytheDanishMeteorological Institute (DMI) and used to test the performance of the heat load forecasts as realistically as possible. Theseweather forecastswerebasedontheHIRLAM(HighResolutionLimitedAreaModel), a numerical weather prediction system developed by a consortium of Europeanmeteorological instituteswith the purpose of providing state-of-the-art short-rangeweather predictions [15], for numericalweatherprediction,hadaforecasthorizonofupto54h,andweredisseminatedfour times aday[15].Wedenote theoutdoor temperature,windspeed,andsolar irradiancebyToutt ,v wind t and Isunt , respectively. 2.2. CalendarData Theheatdemandhasastrongsocial component thatdependsonhumanbehavior. Thesocial component ispart of the reason for thedaily andweeklypatterns in theheat load. Different load profilesonweekdaysandweekendscanalsobeexplainedbyconsumerbehavior. Inorder toallowthe forecastmodels toaccount for loadvariations thatare tied tospecificdays, seasons,andtimesofday, certaincalendardatawere includedas inputvariables. Specifically, thehourof theday, thedayof theweek, theweekend,andthemonthof theyearwereusedas input.Howthecalendardatawas encodedandincludedin themodels isdescribed inSection2.4.1. 2.3.HolidayData Inaddition togenericcalendardata,wealsousedmorespecific localdataaboutspecialdays that mayinfluence theheatconsumptionpattern.ThedistrictheatingsystemofAarhus,Denmark, served asourcase study. Therefore,weuseddataaboutDanishnationalholidays, observances, and local schoolholidays.Nationalholidaysandobservancesweresourcedfrom[16].Nationalholidays include NewYear’sDay,ChristmasDay,EasterDay,etc. andconstitute11daysperyear.Observances include, e.g.,ChristmasEveandConstitutionDayandamount to sixdaysperyear. Informationabout the municipal schoolholidayswascollected fromlocal schools in theAarhusareaandamounts to96days peryearonaverage.Note thatallnationalholidaysarealsoschoolholidays. It is clear that thiskind of information ishighly localandthatgatheringsuchdata, compared to thegeneric calendardata, ismoredifficult. The following analysiswill illuminatewhether including this data significantly improvesheat loadforecasts,or ifmoreeasilyavailabledata typesaresufficient. 253