Seite - 421 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies 2018,11, 213 inDeepEnergyisperformedbythreeconvolutionlayers (Conv1,Conv2,andConv3)andthreepooling layers (Pooling1,Pooling2,andPooling3). TheConv1–Conv3areone-dimensional (1D)convolutions, andthe featuremapsareall activatedbytheRectifiedLinearUnit (ReLU) function. Besides, thekernel sizesofConv1,Conv2,andConv3are9,5,5, respectively,andthedepthsof thefeaturemapsare16,32, 64, respectively. ThepoolingmethodofPooling1 toPooling3 is themaxpooling,andthepoolingsize isequal to2. Therefore,after thepoolingprocess, thedimensionof the featuremapwillbedividedby 2toextract the important featuresof thedeeper layers. In the forecasting, thefirst step is toflat thePooling3 layer intoonedimensionandconstruct afullyconnectedstructurebetweenFlattenlayerandOutput layer. Inorder tofit thevaluespreviously normalized in therange [0,1], thesigmoidal function ischosenasanactivationfunctionof theoutput layer. Furthermore, in order to overcome the overfittingproblem, thedropout technology [34] is adopted in the fullyconnected layer.Namely, thedropout isanefficientwaytopreventoverfitting in artificialneuralnetwork. During the trainingprocess,neuronsare randomly“dead”. Asshownin Figure4, theoutputvaluesofchosenneurons (thegraycircles)areequal tozero incertain training iteration. Thechosenneuronsarerandomlychangedduringtrainingprocess. Furthermore, theflowchartofproposedDeepEnergyis represented inFigure5. Firstly, theraw energy loaddataare loadedinto thememory. Then, thedatapreprocessing isexecutedanddataare normalized in therange[0,1] inorder tofit thecharacteristicof themachine learningmodel. For the purposeofvalidationofDeepEnergygeneralizationperformance, thedataaresplit into trainingdata andtestingdata. The trainingdataareusedfor trainingofproposedmodel.After the trainingprocess, theproposedDeepEnergynetwork is createdand initialized. Before the training, the trainingdata arerandomlyshuffledto force theproposedmodel to learncomplicatedrelationshipsbetween input andoutputdata. Thetrainingdataaresplit intoseveralbatches.Accordingto theorderofshuffled data, themodel is trainedonallof thebatches.Duringthetrainingprocess, if thedesiredMeanSquare Error (MSE) isnot reachedin thecurrentepoch, the trainingwill continueuntil themaximalnumber ofepochsordesiredMSEis reached.Onthecontrary, if themaximalnumberofepochs is reached, then the trainingprocesswill stop regardless theMSEvalue. Finalperformances are evaluated to demonstrate feasibilityandpracticabilityof theproposedmethod. Input Conv1 Pooling1 Conv2 Pooling2 Conv3 Pooling3 Flatten Output Feature extraction Forecasting Figure4.TheDeepEnergystructure. 421