Seite - 186 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2008 Algorithm1:TrainingrestrictedBoltzmannmachinesusingcontrastivedivergence 1 //PositivePhase 2 h0 =σ (Wv0 +c) 3 foreachhiddenunith0i: 4 ifh0i>rand(0,1)//rand(0,1) representsasampledrawnfromtheuniformdistribution 5 h0i=1 6 else 7 h0i=0 8 //NegativePhase 9 v1 =σ (WTh0 +b) 10 foreachvisibleunitsv1j: 11 ifv1j>rand(0,1) 12 v1j=1 13 else 14 v1j=0 15 //UpdatePhase 16 h1 =σ (Wv1 +c) 17 W=ε (h0v0T−h1v1T) 18 b=ε (h0−h1) 19 c=ε (v0−v1) Ascanbeseen inFigure4,a trainedRBMcloselyresemblesasingle layerofanANN.Westack RBMs to formanANN.First, RBM1 is trainedbasedonour inputdatausingAlgorithm1. Then, theentire inputset is fed into thevisible layerofanowfixedRBM1,andtheoutputsat thehidden layerarecollected. Theseoutputsareusedas the inputs to trainRBM2.Thisprocess is repeatedafter RBM2is fully trainedtogenerate the inputs forRBM3,andsoon,asshowninFigure5. This training isunsupervised,meaning thatno target outputs aregiven to themodel. It has informationabout the inputs andhow they are related to one another, but the network is not able to solve any real problemyet. Figure5.Graphical representationofhowRBMsare trainedandstackedto functionasanANN. Thenextstep is trainingaDNN.Backpropagation isusedto train theneuralnetworktosolvea particularproblem. Sinceourproblemisshort-termloadforecasting,naturalgas loadvaluesareused 186