Seite - 251 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

energies Article ImprovingShort-TermHeatLoadForecastswith CalendarandHolidayData MagnusDahl1,2,* ID ,AdamBrun2,OliverS.Kirsebom3 ID andGormB.Andresen1 ID 1 DepartmentofEngineering,AarhusUniversity, IngeLehmannsGade10,8000Aarhus,Denmark; gba@eng.au.dk 2 AffaldVarmeAarhus,MunicipalityofAarhus,Bautavej1,8210Aarhus,Denmark;adbr@aarhus.dk 3 DepartmentofPhysicsandAstronomy,AarhusUniversity,NyMunkegade120,8000Aarhus,Denmark; oliskir@phys.au.dk * Correspondence:magnus.dahl42@gmail.com Received: 28May2018;Accepted: 25 June2018;Published: 27 June2018 Abstract:Theheat loadindistrictheatingsystems isaffectedbytheweatherandbyhumanbehavior, andspecial consumptionpatternsareobservedaroundholidays. Thisstudyemploysa top-down approachtoheat loadforecastingusingmeteorologicaldataandnewuntraditionaldata typessuch asschoolholidays. Threedifferentmachine learningmodelsarebenchmarkedfor forecasting the aggregatedheat loadof the largedistrictheatingsystemofAarhus,Denmark. Themodelsare trained onsixyearsofmeasuredhourlyheat loaddataandablindyearof testdata iswithhelduntil the final testingof the forecastingcapabilitiesof themodels. In thisfinal test,weather forecasts from theDanishMeteorological Instituteareusedtomeasure theperformanceof theheat loadforecasts underrealisticoperationalconditions.Wedemonstratemodelswithforecastingperformancethat canmatchstate-of-the-art commercial softwareandexplore thebenefitof including localholiday data to improveforecastingaccuracy. Thebest forecastingperformance isachievedwithasupport vector regressiononweather, calendar,andholidaydata,yieldingameanabsolutepercentageerror of6.4%onthe15–38hhorizon.Onaverage, the forecastscouldbe improvedslightlybyincluding localholidaydata.Onholidays, thisperformance improvementwasmoresignificant. Keywords:districtheating; loadforecasting;machine learning;weatherdata; consumerbehavior; neuralnetworks; supportvectormachines 1. Introduction Energysystemsarechangingthroughouttheworld,andheat loadforecastingisgainingimportance inmoderndistrictheatingsystems[1]. Thegrowingpenetrationof renewableenergysourcesmakes energyproductionfluctuatebeyondhumancontrolandincreases thevolatility inelectricitymarkets. Stronger coupling between the heating and electricity sectorsmeans that production planners in systemswith combinedheat andpower generation need accurate heat load forecasts in order to optimize theproduction. It is not trivial to forecastdistrict heatingdemandon timescales that are relevant for trading ontheday-aheadelectricitymarket. Thetotalheat loadinadistrictheatingsystemis influencedby several factors—most importantly, theweather, thebuildingmassof thecity,andthebehaviorof the heatconsumers.Coldandwindyweather increases theheatdemand,andwarmandsunnyweather decreases it. Theconstitutionof thebuildingmass influenceshowtheheat loadresponds tochanges in theweather [2].Humanbehavior isanoftenoverlookedfactor, and,especially insummer, theheat demand is dominated byhotwater consumption rather than space heating. Consumer behavior canvaryconsiderably fromdaytoday,andtheheat loadonspecialoccasions,e.g.,NewYear’sEve, isnotoriouslydifficult to forecastaccurately. Energies2018,11, 1678;doi:10.3390/en11071678 www.mdpi.com/journal/energies251