Page - 411 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies 2018,11, 242 For theother fourcomparativepredictors, theirparameterconfigurations for theofficebuilding energyconsumptionpredictionare listedas follows: • FortheBPNN,therewere200neurons inthehiddenlayer. Furthermore, thesigmoidfunctionwas chosentorealize thenonlinear transformationof features.Additionally,werantheBPalgorithm 1000times toobtain thefinaloutputs. • For theGRBFNN,the5-foldcross-validationwasutilizedtodetermine theoptimizedspreadof theradialbasis function. Furthermore, thespreadwaschosenfrom0.01 to2witha0.1step length. • For theELM, therewere150neurons in thehidden layer,andthehardlimfunctionwaschosenas theactivationfunctionforconvertingtheoriginal features intoanotherspace. • For theSVR, thepenaltycoefficientwasset tobe10andthesigmoidfunctionwaschosenas the kernel functiontorealize thenonlinear transformationof input features. 4.4.3. ExperimentalResults For the testingdataof theofficebuilding,partsof theprediction results of thefivepredictors are illustrated inFigure12.Again, forbettervisualization, thepredictionerrorhistogramsof thefive predictorsareshowninFigure13. D 0'%1 $FWXDO YDOXH H 695 $FWXDO YDOXH E %311 $FWXDO YDOXH G (/0 $FWXDO YDOXH F *5%)11 $FWXDO YDOXH 7LPH RI VHYHQ WHVWLQJ GD\V RQH XQLW PLQXWHV 7LPH RI VHYHQ WHVWLQJ GD\V RQH XQLW PLQXWHV 7LPH RI VHYHQ WHVWLQJ GD\V RQH XQLW PLQXWHV 7LPH RI VHYHQ WHVWLQJ GD\V RQH XQLW PLQXWHV 7LPH RI VHYHQ WHVWLQJ GD\V RQH XQLW PLQXWHV Figure 12. Parts of the prediction results of the five predictors constructed by utilizing the energy-consumingpattern: (a)hybridDBNmodel; (b)BPNN;(c)GRBFNN;(d)ELM;and(e)SVR. 411