Seite - 397 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies 2018,11, 242 are, respectively, thedatasetsofweekdaysandweekends,andM1+M2=M. Then, to generate the residual time seriesYRes for the building energy consumptiondata set, weuse the followingrules: I f Yz∈P, then Yz,Res=Yz−Y¯Weekday, (12) I f Yz∈Q, then Yz,Res=Yz−Y¯Weekend, (13) wherez=1,2, . . . ,M. Subsequently,YRes canbewrittenas YRes={Y1,Res,Y2,Res, . . . ,YM,Res} . (14) 3.3.ModifiedDBNandItsTrainingAlgorithm In this subsection, thestructureof theMDBNwillbeshownfirstly. Then, thepre-trainingprocess of theDBNpartwill bedescribed indetail. At last, the least squaresmethodwill be employed to determine theweightingvectorof theregressionpart. 3.3.1. Structureof theMDBN Intheparameteroptimizationof the traditionalDBNs, theCDalgorithmisadoptedtopre-train theparametersofmultipleRBMs,andtheBPalgorithmisusedtofinely tunetheparametersof the whole network. In this paper,we addan extra layer as the regressionpart to theDBN to realize theprediction function. Thus,we call it themodifiedDBN(MDBN). The structure of theMDBN is demonstrated in Figure 4. In addition,wepropose a training algorithm that combines theCD algorithmwith the least squaresmethodfor the learningof theMDBNmodel. Q Q Q QN /HDVW 6TXDUHV 0HWKRG '%1 O O O N [ [ [ Q[ Q[ Q[ E Ö\ &' &' ^ `: E ^ `: E 5HJUHVVLRQ Figure4.Thestructureof themodifiedDBN. Wedivide the trainingprocessof theMDBNinto twosteps. Thefirst stepadopts thecontrastive divergencealgorithmtooptimize thehiddenparameters in apre-trainway,while the secondone determines theoutputweightingvectorbythe least squaresmethod. Thedetaileddescriptionwillbe givenasbelow. 397