Page - 72 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3433 4.3. LongShort-TermMemoryLayer TheLSTMcan capture long-termdependencies in time-stamps; therefore, it can address the vanishinggradientproblems. In theproposedmethod, thenumberofhiddenlayers increasesdueto thedecompositionof inputdata,but thevanishinggradientproblemissolvedthroughthememory cell structurewiththree-stepregularization. Inaddition, tominimizethecovariateshiftproblem,batch normalization isperformedprior to theactivationphaseof the input. IMFsandreference loadprofiles are trainedateachLSTMlayerandhavepredictivevalues, all ofwhicharesummedtopredict the loadprofile. 4.4.ModelConstruction 4.4.1.HyperparameterTuningandTrainingOptions TheLSTMmodelhas several hyperparameters suchas thenumberof inputneurons, hidden layers, inputwindowsize, numberof epochs, regularizationweight, batch size, and learning rate. Thewindow size of input and output parameters depends on the time scale of load forecasting. The inputneuronparameter isdeterminedbythedimensionsof the inputdata. The inputdimension of theproposedmethodis11,which is thesumof thereferenceprofileand10IMFsignals.Weselected thehyperparametersandusedADAMoptimization,oneof theoptimization techniquesused indeep learning[30–40]. 4.4.2. TrainingandTesting TheoverallAMIdatasetofeachdayisdividedintoaratioof70:15:15 for thepurposesofmodel training,validation,andtesting, respectively. 4.4.3. PerformanceMeasures Therootmeansquarederror (RMSE) isusedtocomparedifferencesbetweenthepredictedvalue yˆt andmeasuredvalueyt andiscomputedforT (which is thenumberofsamplesof theweekly load profile)differentpredictionsas thesquarerootof themeanof thesquaresof thedeviations: RMSE= √ ∑Tt=1(yˆt−yt)2 T . (5) Themeanabsoluteerror (MAE) isoneofanumberofwaysof comparing forecastswith their eventualoutcomes. MAE= 1 T T ∑ t=1 |yt− yˆt| . (6) Themeanabsolutepercenterror (MAPE) isalsowidelyusedtoevaluateaccuracy.Accuracycan becomparedviaMAPEusingpercentageswhenthescaleof the loads isdifferent [37–40]. MAPE= 100 T T ∑ t=1 ∣∣∣∣yt− yˆtyt ∣∣∣∣ . (7) 5. LoadProfileAnalysisbyMulti-DecompositionMethods 5.1.WeeklySeasonality This study used real-world load profile data from the R&D business building that utilized enhancedAMIfordemandsidemanagement. Figure4showsthereal-worldloadprofileofthebusiness building. Thebuildinggenerates288samplesperday,2016samplesperweek,and8640samplesper month. The loadprofile ismeasuredandstored indatastorage. 72