Seite - 181 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2008 During thewinter, when outside temperatures are low, the heatload is high. When the outside temperature ishighduring thesummer, theheatload isapproximatelyzero. Otherusesofnatural gas, suchascooking,dryingclothes,andheatingwaterandotherhouseholdappliances,arecalled baseload. Baseload isgenerallynotaffectedbyweatherandtypically remainsconstant throughout the year.However,baseloadmayincreasewithagrowth in thecustomerpopulation. Naturalgasutilityoperationsgroupsdependonreliableshort-termnaturalgas loadforecasts to makepurchasingandoperatingdecisions. Inaccurateshort-termforecastsarecostly tonaturalgas utilitiesandcustomers.Under-forecastsmayrequireanaturalgasutility topurchasegasonthespot marketatamuchhigherprice.Over-forecastsmayrequireanaturalgasutility tostore theexcessgas orpayapenalty. In this paper, we apply deep neural network techniques to the problem of short term load forecastingofnatural gas. Weshowthat amoderately sizedneuralnetwork, trainedusingadeep neuralnetworktechnique,outperformsneuralnetworkstrainedwitholdertechniquesbyanaverageof 0.63 (9.83%)pointsofweightedmeanabsolutepercenterror (WMAPE).Additionally,a largernetwork architecture trained using the discussed deep neural network technique results in an additional improvementof0.20 (3.12%)pointsofWMAPE.Thispaper isanextensionofReference [2]. Therestof themanuscript isorganizedas follows. Section2providesanoverviewofnaturalgas forecasting, includingthevariablesused in typical forecastingmodels. Section3discussespriorwork. Section4providesanoverviewofANNandDNNarchitectureand trainingalgorithms. Section5 discusses thedataused invalidatingourmethod. Section6describes theproposedmethod. Section7 explains theexperimentsandtheir results. Section8providesconclusions. 2.OverviewofNaturalGasForecasting Thebaseloadofnaturalgasconsumption,whichdoesnotvarywith temperature foranoperating area, typicallychangesseasonallyandslowlyas thenumberofcustomers,or theirbehavior, changes. Given thenear steadynatureofbaseload,mostof theeffort in forecastingnaturalgas load focuses onpredictingtheheatload(loadwhichvarieswith temperature).Hence, themost important factor affectingthenaturalgas loadis theweather. Figure1showsthatnaturalgas loadhasaroughly linearrelationshipwith temperaturesabove 65◦F.Forthisreason, it is importanttoconsideravarietyoftemperature-relatedexogenousvariablesas potential inputs toshort-termloadforecastingmodels. Thissectiondiscussesa fewof theseexogenous variables,which includeheatingdegreeday (HDD),dewpoint (DPT), coolingdegreeday (CDD), dayof theweek(DOW),anddayof theyear (DOY). Figure 1. Weighted combination of several midwestern U.S. operating areas, including Illinois, Michigan,andWisconsin.Authorsobtaineddatadirectly fromlocaldistributioncompanies. Thedata is from1January2003 to19March2018. 181