Page - 169 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2038 Inordertocomparetheaccuracyforthedifferenttypesofday, thedaysofthetestdata(2016)were dividedintwogroups: specialdays,which includeweekends,August (officialacademicholidays), andalldays thataredeterminedbythedummyvariablesFH1, . . . , FH5inTable2;and, regulardays, which include therestof thedays. Resultsareexposed inTable7.Notice that the lowestMAPEscores arealwaysreachedforregulardays. Table7.MAPE(%)for regularandspecialdays in2016. Pred.Horizon=48h Bagging RForest CForest XGBoost MAPEregulardays (149) 9.07 8.60 10.44 8.15 MAPEspecialdays (217) 9.97 9.83 11.08 9.57 MAPEtotaldays (366) 9.60 9.33 10.82 8.99 Figure1a,b showthemonthly evolutionof twogoodness-of-fitmeasures (RMSEandMAPE). Remark that accuracies of random forest andXGBoost are quite similar,with greatest differences in January andMarch (due to lackof accuracy inChristmas andEasterndays). Also, themodels fitbetter fornighthours (from10p.m. to5a.m.) due to theabsenceof activityduring thatperiod (seeFigure2a,b). Figure1.Goodness-of-fitmeasures foreachmonth in2016andeachensemblemethod: (a)usingroot meansquareerror (RMSE) (kWh);and, (b)usingmeanabsolutepercentageerror (MAPE) (%). Figure2.Goodness-of-fitmeasures foreachensemblemethodbyhourof theday in2016: (a)using RMSE(kWh); (b)usingMAPE(%). Asanexample,Figure3showstheactualandprediction loadforacompleteweek inMay2016. 169