Seite - 155 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2080 the accuracy of each technique differently and, therefore, these conditions need to be taken into considerationwhendesigningaforecastingsystem.Asageneral result, theARmodelappears tobe slightlymoreaccuratebutrequiresafiner tuningwhentreatingthe temperaturedataandrequiresa largeramountof temperaturedatasources. 5.Conclusions Many different short-term load forecastingmodels have been proposed in the recent years. However, it isdifficult tocompare theaccuracyor thegeneralperformanceofeachmodelwheneach one is testedunderdifferent conditions, testingperiodsanddatabases. Thegoalof thispaper is to provideaseriesofcomparisonsbetweentwoof themostusedforecastingengines: auto-regressive modelsandneuralnetworks. Thestartingpoint isa forecastingsystemcurrently inusebyREEthat includesboth techniques. Several tests havebeen run inorder todetermine the conditionsunder whicheachmodelperformsbest. The results show that bothmodels obtain very similar accuracy and, therefore both of them shouldremain inuse. TheARmodelobtainedabetteroverall resultunder thebestpossiblecondition but theNNmodelwassuperiorwhenfewer temperature locationsareavailable, the treatmentof the temperaturedata isnotproperlyadjustedor the feedback is limitedto less than7 laggeddays. TheAR showedhigheraccuracywhenhistoricaldata is limited to less than7years. Bothmodelshave the sameneeds in termsof trainingfrequency: aone-yearperiod inbetweentrainings is sufficient. Regardingcomputationalburden, theARmodel is less computationally intense than theNN. However, the optimum configuration found at 4 neurons in the hidden layer and 10 redundant networksonlycosts twiceasmuchas theARmodel. Therefore,neithermodelhasadefiniteadvantage onthis front. To sumup, this paper enables the researcher to establish a set of rules to guide them in the processof selectingordesigningaforecastingsystem.Theresultsof this researchofferverypractical information that responds toactualempirical implementationsof thesystemrather thanto theoretical experiments. Further research in this area should include theanalysis ofdifferentdatabases from othersystems. Theuseof informationfromothersystemswouldhelpdetermine if theconclusions drawnaregeneralordatabasespecific, inwhichcase, studyingthespecificitiesofeachdatabaseand determiningwhytheybehavedifferentlywouldalsobeofvalue to thefield. Author Contributions: M.L. conceived and designed the experiments; C.S. (Carlos Sans) performed the experiments;M.L.andC.S. (CarlosSans)analyzedthedata;M.L.;S.V.andC.S. (CarolinaSenabre)wrote thepaper. Acknowledgments:This research isabyproductofacollaborationprojectbetweenREEandUniversidadMiguel Hernández.Openaccesscostswillbe fundedbythisproject. Conflictsof Interest:Theauthorsdeclarenoconflictof interest. Thefundingsponsorshadnorole in thedesign of the study; in the collection, analyses, or interpretationofdata; in thewritingof themanuscript, and in the decisiontopublish theresults. References 1. Hippert,H.S.;Pedreira,C.E.;Souza,R.C.Neuralnetworks forshort-termloadforecasting:Areviewand evaluation. IEEETrans. PowerSyst. 2001,16, 44–55. [CrossRef] 2. Hong,T.;Fan,S.Probabilisticelectric loadforecasting:Atutorial review. Int. J.Forecast. 2016,32, 914–938. [CrossRef] 3. Kuster, C.; Rezgui, Y.; Mourshed,M. Electrical load forecastingmodels: A critical systematic review. Sustain. CitiesSoc.2017,35, 257–270. [CrossRef] 4. Papalexopoulos,A.D.;Hesterberg,T.C.Aregression-basedapproachtoshort-termsystemloadforecasting. IEEETrans. PowerSyst. 1990,5, 1535–1547. [CrossRef] 5. Charlton,N.;Singleton,C.Arefinedparametricmodel forshort termloadforecasting. Int. J.Forecast. 2014, 30, 364–368. [CrossRef] 155