Seite - 381 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1138 networkscanbe trainedwithsamples in thesamecategoryaccordingto thecustomer tobepredicted, so that the interferenceofelectricityusecharacteristicscanbereduced. 6DPSOLQJ 3RLQWV 6DPSOLQJ 3HULRG PLQXWHV Figure7.Loadcurvesof30customers inCategory2. 6DPSOLQJ 3RLQWV 6DPSOLQJ 3HULRG PLQXWHV Figure8.Loadcurvesof30customers inCategory3. 3.2. TheDetailedNetworkStructureandParameters ThedetailedstructureofwholenetworkareshowninTable5. Theparametersof thenetwork aresetas showninTable6. Thestructureandparametersare set forbetterperformanceaccording to themultipleexperiments forcustomers inWanjiangarea. The“RMSprop”optimizer ischosenfor itsbetterperformance inrecurrentneuralnetworks. Theparameterscanbeadjustedfor thedifferent practicalsituations. Inthispaper, thenumberofepochisset to200fortheproposedmethodandcanbe adjustedfor thecomparedmethods. The training is stoppedwhentheerrordecreases toasteadystate. 381