Seite - 239 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1893 Figure6. Vectorofweights (sortedchronologically)obtainedfor thepredictionofadayduringSpring. Figure7 isalsoof interest tounderstandhowthepredictionworks. There, theploton the left containsall thedaysof thedatasetagainstwhichthesimilaritywascomputedwithrespect tothecurve inblue.Atransparencyscalewhichmakesvisibleonly thosecurveswitharelativelyhighsimilarity index. Theplotontherightcontains the futuresof thepastdaysonthe left. Thesearealsoplotonthe transparentscalewith thecurve inorangewhich is thepredictiongivenbytheweightedaverage. Figure 7. Past and future segments involved in the construction of the prediction by KWF. Oneachpanel, all thedaysare representedwitha transparent colourmakingvisibleonly themost relevantdays for theconstructionof thepredictor. 5.ClusteringElectricalLoadCurves Individual load prediction is a difficult task as individual signals have a high volatility. Thevariabilityofeach individualdemandissuchthat theratiosignal tonoisedecreasesdramatically whenpassingfromaggregateto individualdata.Withalmostnohopeofpredictingindividualdata,an alternativestrategy is touse thesedata to improve thepredictionof theaggregatesignal. For this,one mayrelyonclusteringstrategieswherecustomersofsimilarconsumptionstructurewillbeput into classes inorder to formgroupsofheterogeneousclients. If theclientsaresimilarenough, thesignalof theaggregatewillgain inregularityandthus inpredictability. Manyclusteringmethodsexist in thespecialized literature.Weadopt thepointofviewof [33] where twostrategies forclusteringfunctionaldatausingwaveletsarepresented.While thefirstone allows to rapidly creategroupsusingadimension reductionapproach, the secondonepermits to betterexploit the time-frequency informationat thepriceofsomecomputationalburden. 239