Page - 383 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1138 (SRFK Figure9.MAPEsduringatrainingperiodfor theCategory2. 6DPSOLQJ 3RLQWV 6DPSOLQJ 3HULRG PLQXWHV $FWXDO ORDG )RUHFDVWLQJ ORDG Figure10.Comparedcurvesofactual loadandforecasting loadinadayforCustomer53990001. 6DPSOLQJ 3RLQWV 6DPSOLQJ 3HULRG PLQXWHV $FWXDO ORDG )RUHFDVWLQJ ORDG Figure11.Comparedcurvesofactual loadandforecasting loadinaweekforCustomer53990001. The samples are preprocessed byK-means clustering algorithm to form three categories for training.Weperformedacomparativeexperimentwithvariable-controllingapproachaboutclustering. ThecomparedresultsofCustomer53990001onfourdifferentdaysofNovember2013areshownin Figure12. ThecomparedMAPEsofpredictionon18November forninecustomers in threecategories fromdifferent feedersareshowninTable7. It canbeconcludedthat the forecastingcurvewithout clusteringdeviates from the actual curve and that itsMAPE is larger. The reason is thatdifferent 383