Seite - 54 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2019,12, 164 Figure4.Supervised learningof theANN. Forasetoffinite input-targetpairs,once theweightsareadaptivelyadjustedasperMARA[43], the forecastmodule returns the forecast error signal;meanabsolutepercentage error‘MAPE(i) = 1 m ∑ m j=1 |pa(i,j)−pf(i,j)| pa(i,j) ’, to theoptimizationmodule.Where p a(i, j) is theactual loadvalueand pf(i, j) is theforecastedloadvalue. Stepwiseoperationsof theproposedforecastmoduleareshowninFigure5a. ! " # $ % ' ' ( (a)Forecastmodule !"# $ $ !"# $ $ $ $ $ $ %& ' ( ) $ $ %& '*( ) + %& ',( - . / $$ %& '0( ! . $ 1 2 3 5 (b)Optimizationmodule Figure5.Flowchartsofourmodularapproach. 3.3.OptimizationModule Basedonthenatureof theoverall forecast strategy, thebasicobjectiveofoptimizationmodule is tominimize the forecasterror,EF(.), minimize Ith,Rth MAPE(i) (4) where i ∈ [1,m], Ith and Rth represent thresholds for irrelevancy and redundancy, respectively. Optimizationmodulegives Ith’sandRth’soptimizedvalues to theMI-basedfeatureselectionmodule 54