Seite - 427 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies 2018,11, 213 It isobvious that redcurve inFigure11,whichdenotes theDeepEnergyalgorithm, isbetter than othercurves inFigures6–10,whichfurtherverifies that theproposedDeepEnergyalgorithmhas the bestpredictionperformance. Therefore, it isproventhat theDeepEnergySTLFalgorithmproposed in thepaper ispracticalandeffective.AlthoughtheLSTMhasgoodperformance in timesequence problems, in this study, thereductionof training loss is stillnot fastenoughtohandle this forecasting problembecause thesizeof inputandoutputdata is too largefor the traditionalLSTMneuralnetwork. Therefore, the traditionalLSTMisnotsuitable for thiskindofprediction. Finally, theexperimental results showthatproposedDeepEnergynetworkprovides thebest results inenergy loadforecasting. 5.Discussion Thetraditionalmachine learningmethods, suchasSVM,randomforest, anddecisiontree,are widely used inmany applications. In this study, thesemethods also provide acceptable results. InaspectofSVM,thesupportingvectorsaremappedintoahigherdimensionalspacebythekernel function. Therefore, theselectionofkernel function isvery important. Inorder toachieve thegoal ofnonlinearenergy loadforecasting, theRBFischosenasaSVMkernel.Whencomparedwith the SVM,the learningconceptofdecisiontree ismuchsimpler.Namely, thedecisiontree isaflowchart structureeasy tounderstandandinterpret.However,onlyonedecisiontreedoesnothave theability tosolvecomplicatedproblems. Therefore, the randomforest,whichrepresents thecombinationof numerous decision trees, provides themodel ensemble solution. In this paper, the experimental resultsof randomforestarebetter thanthoseofdecisiontreeandSVM,whichproves that themodel ensemble solution is effective in theenergy load forecasting. Inaspectof theneuralnetworks, the MLPis thesimplestANNstructure.AlthoughtheMLPcanmodel thenonlinearenergyforecasting task, itsperformance inthisexperiment isnotoutstanding.Ontheotherhand, theLSTMconsiders datarelationships in timestepsduringthe training.Accordingto theresult, theLSTMcandealwith the timesequenceproblems,andthe forecastingtrendismarginallycorrect.However, theproposed CNNstructure, named theDeepEnergy, has the best results in the experiment. The experiments demonstrate that themost important featurecanbeextractedby thedesigned1Dconvolutionand pooling layers. Thisverificationalsoproves theCNNstructure iseffective in the forecasting,andthe proposedDeepEnergygives theoutstandingresults. Thispapernotonlyprovides thecomparisonof the traditionalmachine learninganddeeplearningmethods,butalsogivesanewresearchdirection in theenergy loadforecasting. 6.Conclusions Thispaperproposesapowerfuldeepconvolutionalneuralnetworkmodel(DeepEnergy)forenergy loadforecasting. Theproposednetworkisvalidatedbyexperimentwiththe loaddatafromthepast sevendays. In theexperiment, thedata fromcoastareaof theUSAwereusedandhistoricalelectricity demand fromconsumerswas considered. According to the experimental results, theDeepEnergy canpreciselypredict energy load in thenext threedays. In addition, theproposedalgorithmwas comparedwith fiveAI algorithms thatwere commonlyused in load forecasting. The comparison showed that performance of DeepEnergywas the best among all tested algorithms, namely the DeepEnergyhad the lowestvaluesofbothMAPEandCV-RMSE.According toall of theobtained results, theproposedmethodcanreducemonitoringexpenses, initial costofhardwarecomponents, and long-termmaintenance costs in the future smartgrids. Simultaneously, the resultsverify that proposedDeepEnergy STLFmethodhas stronggeneralization ability and robustness, thus it can achieveverygoodforecastingperformance. Acknowledgments:ThisworkwassupportedbytheMinistryofScienceandTechnology,Taiwan,RepublicofChina, underGrantsMOST106-2218-E-153-001-MY3. AuthorContributions: Ping-HuanKuowrote theprogramanddesigned theDNNmodel. Chiou-JyeHuang plannedthisstudyandcollectedtheenergy loaddataset. Ping-HuanKuoandChiou-JyeHuangcontributed in draftedandrevisedmanuscript. 427