Seite - 395 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies 2018,11, 242 Q Q Q Q Q Q NQ NQ NQ NQ Q 5%0 5%0 5%0N '%1 ,QSXW OD\HU +LGGHQ OD\HU +LGGHQ OD\HU +LGGHQ OD\HUN ^ `T : E ^ `T : E ^ `N N NT : E 9LVLEOH /D\HU +LGGHQ /D\HU 9LVLEOH /D\HU 9LVLEOH /D\HU +LGGHQ /D\HU +LGGHQ /D\HU Figure2.Thearchitectureof theDBNwithkhiddenlayers. 3.TheProposedHybridModel In this section, thestructureof thehybridmodelwillbeproposedfirst. Then, theextractionof the energy-consumingpatternandthegenerationof theresidualdatawillbegiven. Finally, themodified DBN(MDBN)andits trainingalgorithmwillbepresented. Tobegin,we assume thatwehave collected the samplingdata forM consecutivedays, and, ineachday,wecollectedTdatapoints. Then, sampledtimeseriesofenergyconsumptiondatacanbe writtenasaseriesof1Dvectorsas Y={Y1,Y2, . . . ,YM} , (4) where Y1=[y1(1),y1(2), . . . ,y1(T)], ... YM=[yM(1),yM(2), . . . ,yM(T)], (5) andT is thesamplingnumberperday. 3.1. Structureof theHybridModel Thehybridmodelcombines themodifiedDBN(MDBN)modelwith theperiodicityknowledge of thebuildingenergyconsumptiontoobtainbetterpredictionaccuracy. Thedesignprocedureof the proposedmodel isdepicted inFigure3andisalsogivenas follows: Step1: Extract theenergy-consumingpatternas theperiodicityknowledgefromthetrainingdata. Step2: Removetheenergy-consumingpattern fromthetrainingdata togenerate theresidualdata. Step3: Utilize theresidualdata to train theMDBNmodel. Step4: Combinetheoutputs fromtheMDBNmodelwith theperiodicityknowledgetoobtain the finalpredictionresultsof thehybridmodel. 395