Page - 153 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2080 0 0.5 1 1.5 2 2.5 3 3.5 4 1.30% 1.50% 1.70% 1.90% 2.10% 2.30% 2.50% 3 5 10 11 12 14 15 16 20 25 Networks Overall accuracy vs redundant networks OVERALL REGULAR SPECIAL HOT COLD Time (s) Figure12.Overall forecastingerror (RMSE)andexecutiontimewithdifferent redundantnetworks. 4.6. FrequencyofTraining TheresultsfromTable7showtheperformanceofbothmodelswhenthetrainingperiodischanged from3months to24months. The testingperiod remains the sameasdescribed inTable1, but the dataused to train themodel that forecastedeachblockchanges. Thereappear tobenosignificant improvementfromretrainingthemodelsmorefrequentlythanannually,asseenonFigure13.However, a trainingperiodlongerthanayearseemstocauseanincrease intheforecastingerror. Bothmodelsare affectedverysimilarlybythisparameter,withanincrease in theerrorofabout23%forbothmodels whenincreasingthe timeinbetweentrainings from12to24months. Table7.Forecastingerror (RMSE)withdifferent trainingfrequency. TypeofDay 3Months 6Months 12Months 24Months AR NN AR NN AR NN AR NN Overall 1.44% 1.54% 1.44% 1.56% 1.45% 1.55% 1.78% 2.07% Regular 1.39% 1.45% 1.40% 1.47% 1.40% 1.46% 1.74% 2.01% Special 1.78% 2.09% 1.79% 2.10% 1.81% 2.10% 2.13% 2.35% Hot 1.57% 1.96% 1.57% 1.97% 1.55% 1.93% 2.26% 3.11% Cold 1.70% 1.45% 1.69% 1.46% 1.72% 1.45% 1.96% 1.96% Testconditions: 7YT,4N,10RN,5TL,7/14LAG. 0.00% 0.50% 1.00% 1.50% 2.00% 2.50% 3.00% 3 months 6 months 12 months 24 months Accuracy vs training frequency AR NN Figure 13. Forecasting error (RMSE) for AR andNNmodels with training frequency from 3 to 24months. 153