Seite - 422 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies 2018,11, 213 Network converged? Load the raw energy load data Start Data preprocessing Split the training and testing data Shuffle the order of training data Initialize of the neural network Train the model on batchs Performance evaluation on testing data End Yes No Figure5.TheDeepEnergyflowchart. 4. ExperimentalResults In theexperiment, theUSADistrictpublicconsumptiondatasetandelectric loaddataset from 2016providedbytheElectricReliabilityCouncilofTexaswereused. Since then, thesupportvector machine (SVM)[35] isapopularmachine learningtechnology, inexperiment; theradialbasis function (RBF) kernels of SVMwere chosen to demonstrate the SVMperformance. Besides, the random forest (RF) [36],decision tree (DT) [37],MLP,LSTM,andproposedDeepEnergynetworkwerealso implementedandtested. Theresultsof loadforecastingbyallofthemethodsareshowninFigures6–11. Intheexperiment, thetrainingdataweretwo-monthdata,andtestdatawereone-monthdata. Inorder toevaluate theperformancesofall listedmethods, thedatasetwasdividedinto10partitions. In the firstpartition, trainingdataconsistedofenergy loaddatacollected in JanuaryandFebruary2016,and testdataconsistedofdatacollected inMarch2016. In thesecondpartition, trainingdataweredata collected inFebruaryandMarch2016,andtestdataweredatacollected inApril2016. Thefollowing partitionscanbededucedbythesameanalogy. In Figures 6–11, red curves denote the forecasting results of the correspondingmodels, and bluecurvesrepresent thegroundtruth. Theverticalaxesrepresent theenergyload(MWh),andthe horizontal axesdenote the time (hour). The energy load from thepast (24× 7) hwasusedas an inputof the forecastingmodel, andpredictedenergy loadin thenext (24×3)hwasanoutputof the forecastingmodel.After themodels receivedthepast (24×7)hdata, they forecastedthenext (24×3) henergy load, redcurves inFigures6–11. Besides, thecorrect information is illustratedbybluecurves. Thedifferencesbetweenredandbluecurvesdenote theperformancesof thecorrespondingmodels. For thesakeofcomparisonfairness, testingdatawerenotusedduringthe trainingprocessofmodels. Accordingto theresultspresented inFigures6–11, theproposedDeepEnergynetworkhas thebest predictionperformanceamongallof themodels. 422