Seite - 252 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1678 Heat load forecastinghas been studied extensively in the scientific literature. The successful application of simple linear models in [3,4] has inspired us to use an ordinary least squares (OLS)model as a simple benchmarkingmodel. Statistical time-series models, such as SARIMA (seasonalautoregressive integratedmovingaverage)models [4,5]andgrey-boxmodelscombining physical insightwith statisticalmodeling [6], arenaturalwaysofhandling the temporalnatureof load forecasting. Thesemodels are usually linear and strugglewithmultiple seasonality. In [7], theauthors comparedanumberofmachine learningalgorithms, includinga simple feed forward neural network, support vector regression (SVR), andOLS. They concluded that the SVRmodel performsbest. ThestrongforecastingcapabilitiesofSVRmodelshavealsobeendemonstrated in[8], whereheatdemandwasforecastedbasedonnaturalgasconsumption.Neuralnetworkshavebeen widely applied in load forecasting. Several studies apply simple feed forwardnetworkswithone hidden layer suchas themultilayerperceptron (MLP) [7,9,10]. Arecurrentneuralnetwork isused in [11] tobetterhandlenon-stationarities in theheat load.Acomprehensivereviewof loadforecasting indistrictscanbefoundin[1]. In thepresentstudy,wechose tocompare threedifferentmachine learningmodels:OLS,MLP, andSVR,as theyhaveallproveneffective forheat loadforecasting. Somestudiesattempt to include thedifferentconsumerbehavioronweekdaysandweekends.Workingdaysandnon-workingdays aremodeledwithdistinctprofiles in [12], andin[4]mid-weekholidayswere treatedasSaturdaysor Sundays. In [13], thecorrelationbetweenelectric loadandweathervariableswasexploited to forecast theaggregated loadusingMLPmodels, andtheauthorsexploredthedifferentautocorrelationsof the loadonweekdaysandweekends. In thisstudy,weincludegenericcalendardatasuchas thedayof theweek,aswellas localholidaydata toaccount forobservances,nationalholidays,andcity-specific holidays, i.e., schoolholidays. School holidays are often planned locally, and some religious holidays, e.g., Easter, fall on differentdates eachyear. Therefore, generic calendardata is insufficient formodelingevents that dependonlocalholidays.Heatconsumersbehavedifferentlyonholidaysandchangethepatternof consumption, so including localholidaydata inheat loadforecastmodelshas thepotential to improve forecastaccuracy. Thenoveltyof thisworklies in theapplicationofnewdatasources, specifically localholidaydata, tocreateheat loadforecastingmodels thatmoreaccuratelycaptureconsumerbehavior. To thebestof ourknowledge, schoolholidaydatahasnotpreviouslybeenusedforheat loadforecasting.Weisolate theeffectofusinglocalholidaydatabyemployingmachine learningmodels thathaveproveneffective forheat loadforecasting in thepast.Moreover,webaseourmodelingonavery largeamountofdata. Sevenyearsofhourlyheat loadandweatherdatasupplementedwithdataaboutnationalholidays, observances,andschoolholidayshelp the forecastmodelscapturerare loadevents. Theremainderof thepaper is structuredas follows. TheMethodologysectiondescribes thedata foundation, themachine learningmodels, andthevalidationandtestingprocedure. In theResults section, the forecastingmodelsarebenchmarkedandcompared,andthepotentialofusingnewdata sources isevaluated. Thepaper iswrappedupintheConclusionsection. 2.Methodology In thissection,wedescribe thedata foundationandhowtheheat loadforecastingmodelswere built,validatedandtested. Thefocusof thispaper is tocreateheat loadforecasts thatare relevantonthe timehorizonof the day-aheadelectricitymarket. Therefore,a forecastmustbeproducedeachmorningat10:00 foreach hourofthefollowingday.Thistimeline, illustratedinFigure1,allowstimeforcommunicationbetween different actors inaproductionsystemandforplanningof the followingday’sheatproduction in accordancewith thebids in theday-aheadelectricitymarket. 252