Page - 419 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies 2018,11, 213 frequencies. Therefore, thewaveletdenosingalgorithmprovidesgoodelectricity loaddata forneural networktrainingandimproves loadforecastingaccuracy. In this study, anew load forecastingmodel basedonadeep learning algorithm ispresented. Theforecastingaccuracyofproposedmodeliswithintherequestedrange,andmodelhasadvantagesof simplicityandhighforecastingperformance. Themajorcontributionsofthispaperare: (1) introduction ofaprecisedeepneuralnetworkmodel forenergy loadforecasting; (2) comparisonofperformancesof several forecastingmethods;and, (3)creationofanovelresearchdirectionintimesequenceforecasting basedonconvolutionalneuralnetworks. 2.MethodologyofArtificialNeuralNetworks Artificial neural networks (ANNs) are computing systems inspired by the biological neural networks. ThegeneralstructureofANNscontainsneurons,weights,andbias. Basedontheirpowerful moldingability,ANNsarestillverypopular in themachine learningfield.However, therearemany ANNstructuresusedinthemachine learningproblems,but theMultilayerPerceptron(MLP) [32] is themostcommonlyusedANNtype. TheMLPisa fullyconnectedstructureartificialneuralnetwork. ThestructureofMLPisshowninFigure1. Ingeneral, theMLPconsistsofone input layer,oneormore hidden layers, andoneoutput layer. However, theMLPnetworkpresented inFigure1 is themost commonMLPstructure,whichhasonlyonehiddenlayer. In theMLP,all theneuronsof theprevious layerare fullyconnectedto theneuronsof thenext layer. InFigure1,x1,x2,x3, . . . ,x6 are theneurons of the input layer,h1,h2,h3,h4 are theneuronsof thehidden layer,andy1,y2,y3,y4 are theneuronsof theoutput layer. Inthecaseofenergyloadforecasting, theinput is thepastenergyload,andtheoutput is the futureenergyload.Although, theMLPstructure isverysimple, itprovidesgoodresults inmany applications. ThemostcommonlyusedalgorithmforMLPtraining is thebackpropagationalgorithm. Figure1.TheMultilayerPerceptron(MLP)structure. AlthoughMLPsareverygood inmodellingandpatter recognition, the convolutional neural networks (CNNs) provide better accuracy in highly non-linear problems, such as energy load forecasting. TheCNNuses the concept ofweight sharing. Theone-dimensional convolutionand pooling layerarepresentedinFigure2. The lines in thesamecolordenote thesamesharingweight, andsetsof thesharingweightscanbe treatedaskernels.After theconvolutionprocess, the inputsx1, x2,x3, . . . ,x6 are transformedto the featuremaps c1, c2, c3, c4. Thenext step inFigure2 ispooling, wherein the featuremapofconvolution layer is sampledanditsdimension is reduced. For instance, in Figure2dimensionof the featuremapis4,andafterpoolingprocess thatdimension is reducedto2. Theprocessofpooling isan importantprocedure toextract the importantconvolutionfeatures. 419