Page - 283 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1605 hand, short-termforecastingmodels, suchasBPNNandSVM,providedexcellentperformance for one-stepforecastingtask.However, thesemodelsperformedpoorlyorsufferedseveredegradation whenappliedto thegeneralmultistepproblems. Ingeneral, theperformanceofensemble forecasting models (e.g., combiningshort-termandlong-termapproaches)wasbetterwhencomparedtosingle models. Therefore,a forecastingcombinationcanbenefit fromperformanceadvantagesofshort-term andlong-termmodels,whileavoidingtheirdisadvantages. Furthermore, toovercometheshortcoming of a static combinationapproach, adynamic combinationof short- and long-termforecasts canbe employedbyusinghorizondependentweights. Table7.SummaryofevaluationmeasuresamongallmodelsonGOCdata. NO. Model EvaluationMatrix Score1 IndexedRank2 T-Time MAPE(%) ED DA(%) 1 LR 0.04 2.4 0.074 82.59 29 8 2 BPNN 0.05 1.42 0.035 89.03 20 6 3 SVR 0.03 1.24 0.034 89.90 19 5 4 Bagging 0.06 1.66 0.026 66.17 18 4 5 AR 0.07 1.41 0.059 82.59 22 7 6 1stSMLE 0.09 0.9 0.028 88.50 15 3 7 2ndSMLE 0.13 0.78 0.024 90.69 9 2 8 3rdSMLE 0.17 0.74 0.020 91.24 8 1 1 Score: sumofrankvalues from(1–8) foreachmodeldependsonperformance inrelatedmeasure. 2Ordervalue foreachmodeldependingontotal score, forexamplerankno1means thefirstmodel. Figure9. IllustratedT.TimeMAPE,ED,andADevaluationmeasuresofallmodelson10-aheadGOC prediction. (Theorderofmodelswere arranged from1–8according toTable 6, asLR,BPNNSVR, bagging,AR,1st, 2nd,and3rdSMLE,respectively). 283