Seite - 76 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3433 and thepredictionaccuracybydayof theweek is showninFigure8. Table2also summarizes the performanceatdifferent timescales. 6.2.WeeklyLoadForecasting Figure7 illustrates theSTLFforbuilding loadwithonehourago(12stepsahead). Tocheckthe performanceof theproposedmethodbasedonVMDandLSTM, thepredictionresultsofdifferent methodswerecompared.Acloser lookat thepredictionresults reveals theMondayloadforecasting in Figure7b. Theproposedmodelshowedrobustperformanceunderabrupt loadincreasesanddecreases in400samplesand500samples, respectively.Conventionalmodelsexhibitedconservativechanges to suddenloadchanges,andEMD-LSTMexhibitedexcessiveweightchanges. Figure8showstheaveragepredictiveerrorof thedifferentmethods. Theresultof loadforecasting withone-monthAMIdata is shown inFigure 8a, andFigure8b is theprediction resultwith three monthsofAMIdata. Therearedistinct loadcharacteristics for eachdayof theweek. EMD-LSTM hadlargeerrorswithanRMSEof32.68kWh,MAEof28.61kWh,andMAPEof12.24%onSundayin Figure8a.However, if thesizeof thedataset is sufficiently largeor thepredictiontimescale is long enough, the initialerrorcanbecorrected.Whenthedataare insufficientwithashort timescale, the inputof thereference loadprofile (which ismeasurementdataat themaximumobservable timebefore loadforecasting)canbeadominant featureofmachine learning,whichcausesa largeerror. Figure8 showsthat, if theLSTMcorrectlydecomposedperiodic features, ithadhighaccuracyevenwithsmall amountsofdata,but if therewasanerror in the feature, thepredictionerroralso increasedbecauseof thememorycell structureofLSTM. VMDcanreflectmoredominantpatterns thanEMDwithdistinctperiodicity. Theperformance difference of decomposition between EMD andVMD is shown in Figures 4 and 5. The RNNs usingVMDshowedperformance improvements. However, therewasadifference inperformance improvement between NARX and LSTM because the vanishing gradient problem was solved differently,whereNARXusedthedelayfactorandLSTMhadthememorycell structure.AsLSTM preservedcharacteristicsofdominantfeaturesthroughthememorycell,LSTMshowedhigheraccuracy thanNARXinSTLF. Sun Mon Tue Wed Thu Fri Sat 0 10 20 30 40 Sun Mon Tue Wed Thu Fri Sat 0 5 10 15 20 25 30 Sun Mon Tue Wed Thu Fri Sat 0 2 4 6 8 10 12 14 (a) Sun Mon Tue Wed Thu Fri Sat 0 10 20 30 40 Sun Mon Tue Wed Thu Fri Sat 0 5 10 15 20 25 30 Sun Mon Tue Wed Thu Fri Sat 0 2 4 6 8 10 12 14 (b) Figure8.Benchmarksofdifferentmodels. (a)One-monthAMIdata; (b)Three-monthAMIdata. 76