Page - 75 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3433 Table1.Correlation indexcomparisonofEMDandVMD. Decomposition IMF 1 2 3 4 5 6 7 8 9 10 EMD 0.58 0.42 0.40 0.28 0.46 0.35 0.02 0.43 0.82 0.96 VMD 0.98 0.83 0.80 0.63 0.53 0.26 0.15 0.01 −0.02 −0.02 Inaddition,VMDcanremovetheinherentnoise.ActualAMIdatahavenoiseowingtotheinterference due toperipheral electronicdevices. VMDcan improve theaccuracyof the load forecasting through thedeep learning trainingandregularizationprocessbyreducing theweightofhigh frequencies that aresusceptibletonoise,suchasVMF-8,VMF-9,andVMF-10,whichhavelowcorrelationindicesof less than1%.TheAMIusedinthisstudyhasathree-timeshighersamplingthanconventionalAMIandcan reducethemodeluncertaintyasmoresamplesaremeasured.Theproposedmethodreducestheprediction uncertaintybytrainingthedecomposedsignalwiththehighsamplingAMI. 6.CaseStudies Thetimeseries forecastingmodelsweresimulatedonreal-worlddatasetsofbusinessbuildings. Weconductedthecasestudieswithdifferentpredictionmodelsandpredictiontimescales. Theweekly predictionresults foronehouraheadloadforecastingareshowninFigure7. 0 200 400 600 800 1000 1200 1400 1600 1800 2000 150 200 250 300 350 400 (a) 350 400 450 500 550 600 180 200 220 240 260 280 300 320 340 360 380 (b) Figure7.Actualanddifferent loadforecasting foraweek. (a)Weekly loadforecasting; (b)Monday loadforecasting. 6.1. ComparativeConventionalLoadForecastingModels To validate the efficacy of the proposed VMD-LSTM RNN, eight load forecasting models, includingARIMA, SVR, GPR,NARX,NARXwith EMD,NARXwith VMD, LSTM, LSTMwith EMD,andLSTMwithVMD,werecomparedunder thesamebenchmarks (RMSE,MAE,andMAPE). TheARIMAmodelhasbeenusedfor time-seriesprediction.However,with theriseofmachine learning, theGPRandSVRmodelsarebeingutilized. Toaccount forseasonality inanARIMAmodel, threehyperparameterswereused: autoregression, stationarity,andmovingaverage. TheGPRmodel usesstatisticalhyperparameters, includingvarianceandlength,whereas theSVRmodeldependson kernelparameters,apenaltyfactor,andinsensitivezonethickness. TheARIMA,GPR,andSVRmodels aretrainedthroughcross-validationandADAMoptimizationorparticleswarmoptimization[2,26–29]. Tocompare theperformanceof theRNNs,wecomparedtheresultsofapplyingtwodecomposition methods to theNARXandLSTMmodelsThepredictionresultsofallmodelsareshowninFigure7, 75