Page - 128 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3283 Table2.Statisticsofpowerconsumptiondata. Statistics Cluster# ClusterA ClusterB ClusterC Numberofvalidcases 1826 1826 1826 Mean 63,094.97 68,860.93 30,472.31 Variance 246,836,473 269,528,278 32,820,509 Standarddeviation 15,711.03 16,417.31 5728.92 Maximum 100,222.56 109,595.52 46,641.6 Minimum 23,617.92 26,417.76 14,330.88 Lowerquartile 52,202.4 56,678.88 26,288.82 Median 63,946.32 66,996.72 30,343.14 Upperquartile 76,386.24 79,209.96 34,719.45 5.2. ForecastingModelConfiguration In this study, we used the LSTMnetworksmethod to show the repeating pattern of power consumptions depending on the day of theweek. We tested diverse cases and investigated the accuracyof loadforecasting for the test cases todetermine thebest inputdataselection. Asshown in Figure 4, the input variables consist of four electrical loads fromoneweek ago to fourweeks ago as apart at aweekly interval to reflect the cycle of onemonth. In the feature scalingprocess, werescaled the rangeof themeasuredvalues from0 to1.Weused tanhas theactivation function andcalculatedthe lossbyusingthemeanabsoluteerror.Weusedtheadaptivemomentestimation (Adam)method,whichcombinesmomentumandrootmeansquarepropagation(RMSProp),as the optimizationmethod. TheAdamoptimizationtechniqueweighsthetimeseriesdataandmaintainsthe relativesizedifferencebetweenthevariables. In theconfigurationof theremaininghyper-parameters of themodel,weset thenumberofhiddenunits to60,epochs to300,andbatchsize to12. Figure4.SystemarchitectureofLSTMnetworks. We experimentedwith the LSTMmodel [52] by changing the time step fromone cycle to a maximumof30cycles. Table3showsthemeanabsolutepercentageerror (MAPE)ofeachcluster for eachtimestep. In the table, thepredictedresultswith thebestaccuracyaremarkedinbold. Table3 shows that the 27th time step indicates themost accuratepredictionperformance. Ingeneral, the electricitydemandisrelativelyhigh insummerandwinter, comparedto that inspringandautumn. Inotherwords, ithasariseandfall curve inahalf-yearcycle,andthe27th timestepcorresponds toa weeknumberofabouthalfayear. Weperformedsimilara timeseriespatternanalysisbasedonthedecisiontree through10-fold cross-validation for the trainingset. Amongseveraloptionsprovidedbyscikit-learn toconstructa decisiontree,weconsideredthecriterion,maxdepth,andmaxfeatures. Thecriterion isa functionfor measuringthequalityofasplit. In thispaper,weusethe“mae”criterionforour forecastingmodel since itgives thesmallesterrorratebetweentheactualandtheclassificationvalue.Maxdepth is the maximumdepthof the tree.Wesetmaxdepthto3, suchthat thenumberof leaves is8. Inotherwords, thedecision tree classifies the trainingdatasets into eight similar timeseries. Max features are the numberof features to considerwhen looking for thebest split. Wehavechosen the“auto”option 128