Page - 77 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3433 TheMAPEofVMD-LSTMwasaround2%. In theweeklycomparison, the leasterroroccurredon Tuesday:RMSEof6.49kWh,MAEof3.98kWh,andMAPEof1.48%.Thiswasbecause thecorrelation betweendaysof theweekwasthehighestonTuesday.Ontheotherhand, therewasa largeerroron WednesdayandFridaybecause thecorrelationwasrelatively lower thanonotherdaysof theweek. TheproposedVMD-LSTMreflecting themixedperiodic pattern of the loadprofile basedon multi-decompositionwithdeeplearninghadthe lowesterror. 6.3. Benchmark forDifferentPredictionTimeScales Finally, in this section,weanalyze theaccuracyof the loadforecastingmethods for thecasestudy consideringdifferentprediction timescales (5m,1h,3h,24h,48h,72h). Theaccuracyresultsare summarized inTable2. Thebestaccuracieswereobtainedfor theshortestpredictiontimescale (5m) forallmodels. Theproposedmodel,VMD-LSTM,showedthebestaccuracywithanMAEof1.95kWh, RMSEof4.28kWh,andMAPEof0.71%. Inaddition, EMD-LSTMandVMD-LSTMshowedbetter accuracyon thepreviousdaywhen compared to the36stepsahead(3h)andoneday to threedaysahead(24h,48h,72h) timescales. The24h,48h,and72hcasesshowthatRNN-basedmodelshadhigheraccuracies thanARIMAor GPR,buteventuallyshowedsimilarerrors,andtheirperformancesweresaturated. This resultwas obtainedbecausethereferenceloadprofilewaslearnedasadominantinputaccordingtotheprediction timescale toreflect thepowerconsumption trend, so the288stepsaheadcaseand576stepsahead case,whichhadsimilarpatterns,wereslightlymoreaccurate thanthe36stepsaheadcase (3h). Table2.Loadforecastingerrorsofdifferentmodels. Prediction Horizon Index ARIMA GPR SVR NARX EMD NARX VMD NARX LSTM EMD LSTM VMD LSTM 1step ahead (5min) MAE 7.45 6.03 3.43 7.52 7.33 3.25 2.92 5.53 1.95 RMSE 11.77 10.21 6.89 11.89 11.21 6.62 4.98 8.72 4.28 MAPE(%) 3.46 2.67 1.96 3.61 3.39 1.84 1.12 2.21 0.71 12steps ahead (1h) MAE 17.28 16.11 14.76 17.71 17.02 15.12 9.01 11.69 4.81 RMSE 22.12 20.94 20.12 24.12 22.49 19.31 12.87 15.08 7.53 MAPE(%) 6.20 6.06 5.70 6.35 6.27 5.43 3.54 4.27 1.90 36steps ahead (3h) MAE 57.14 53.96 48.72 58.85 56.54 50.69 30.25 38.52 16.27 RMSE 64.50 61.35 59.31 70.64 66.80 56.38 38.05 43.66 22.40 MAPE(%) 20.62 19.91 18.17 21.79 20.97 20.03 11.63 14.26 6.01 288steps ahead (24h) MAE 51.22 48.55 43.25 52.68 51.50 45.12 28.19 32.65 15.60 RMSE 59.38 58.81 56.88 58.12 57.24 56.85 35.98 37.38 21.80 MAPE(%) 18.90 17.91 16.13 19.16 19.06 16.71 10.62 11.78 5.75 576steps ahead (48h) MAE 57.24 52.87 46.57 57.48 55.65 48.23 28.60 32.48 15.85 RMSE 63.28 60.31 59.72 62.51 61.49 57.53 36.24 42.27 22.11 MAPE(%) 22.08 19.72 17.76 21.49 20.99 17.18 10.92 12.18 5.89 864steps ahead (72h) MAE 60.45 58.55 51.92 59.75 59.01 53.44 29.12 34.37 16.09 RMSE 68.24 62.42 58.35 67.38 66.26 58.72 36.85 43.52 22.18 MAPE(%) 24.14 21.76 18.72 22.43 21.19 19.05 11.05 12.86 5.96 7.Conclusions This paper proposes short-term load forecasting using deep learning based on multi-decomposition. The results of the proposed approach were validated by applications to real-worlddata fromabusiness building, and theperformance of theproposed approachwas assessedbycomparingthepredictedresultswith thoseofothermodels. Tomonitor small-scale loadanddemand sidemanagement, an enhancedAMI that provides three-timesmore sample data points per hour than conventional AMIwas used, increasing the accuracyof the loadforecastingusingdeeplearning. In thisstudy, todetect the featuresof the load 77