Page - 382 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1138 Table5.Numberofunits in theproposednetwork. Layer NumberofUnits Layer NumberofUnits INPUT:Pl (6, 16) INPUT:DpWp Tp 3 BATCHNORMALIZATION (6,16) DENSE1 100 GRU 300 DENSE2 100 MERGE 400 DROPOUT 400 DENSE3 100 DENSE4 100 OUTPUT:Pp 96 Table6.Parametersetting in theproposednetwork. Parameters Value InputTimeSteps 6 InputDimension 16 BatchSize 30 Epoch 200 Optimizer RMSprop LearningRate 0.001 Decay 0.9 DropoutRate 0.6 3.3. ComparisonofResults ofProposedMethod Theresultsoftheproposedmethodareshownasfollows. Intheexperiments, thetrainingsamples are recorded fromthe loaddata in theperiod fromOctober 2012 toSeptember 2013while the test samplesarerecordedfromloaddata in theperiodfromOctober toDecemberof2013. Thenumber of recorded training samples and test samples of each categories is 36,000 and 9000, respectively, with100customers inacategory. Theratioofsamplenumber is4:1.Meanabsolutepercentageerror (MAPE) is theclassicevaluation indexfor loadforecasting. Thecomputational formula isgivenby Equations (16) and (17), wheren=96 represents the dimension of samples andm represents the numberof test samples. MAPE= ∑mj=1 ∣∣Ej∣∣ m ×100% (16) Ej= ∑ni=1 ∣∣∣Pp,ij−Pl,ij∣∣∣ n (17) Customer53990001 is selectedfromCategory2for the forecastingcustomer. TheMAPEsduring atrainingperiodforCategory2areshowninFigure9whentheparametersaresetasshowninTable6. Thecomparedcurvesofactual loadandforecasting loadusing theproposedmethodon18November forCustomer53990001areshowninFigure10. TheMAPEforCustomer53990001on18November is 10.23%. The compared curves of actual load and forecasting load from18 to 24November for Customer53990001areshowninFigure11. TheMAPEforCustomer53990001 in thisweek is10.97%. InFigures10and11, theerror insamplepointsofoneday isbasicallyaverageandbecomes larger when the curve comes toapeak. It is reasonablebecause thehighor lowpeak isnot reachable in mostcases. Thenetworkshouldbalance thepredictionresults formostsituationsduringthe training process. According toFigure9, theMAPEdecreases toasteadystateas theepoch increases to200. According toFigures10and11, the forecastingcurve is close to theactual curve,whichproves the availabilityof theproposedmethod. 382