Seite - 273 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1605 Figure2.Comparisonof (a) theactualandpredictedconsumptionwith theuseof theSVR,BPNNand LRsingle learners (b) errorsofall singlemodels. 2.3.2.Models Asabove-mentioned,weapplied theensembleSMLEmodel topredict theGOCdatasetafter combiningtheheterogeneousmodels,Table2 lists the learners’parameters thathavebeeninvestigated in thispaper. Theserelatedmethodsarepresentedbrieflyas follows: 1. TheBPNNalgorithmconsistsofmultiple layersofnodeswithnonlinearactivationfunctionsand canbeconsideredas thegeneralizationof thesinger-layerperceptron. Ithasbeendemonstrated tobeaneffectivealternative to traditional statistical techniques inpatternrecognitionandcan beusedforapproximatinganysmoothandmeasurable functions [49]. Thismethodhassome superiorabilities, suchas itsnonlinearmappingcapability, self-learningandadaptivecapabilities, andgeneralizationability. Besides these features, theability to learn fromexperience through trainingmakesMLPanessential typeofneuralnetworksandit iswidelyappliedto timeseries analysis [50]. 2. TheSVMalgorithmisalwaysconsideredauseful tool forclassificationandregressionproblems dueto theability toapproximatea function. Furthermore, thekernel function isutilized inthe SVRtoavoidthecalculations inhigh-dimensional space.Asaresult, it canperformwellwhen the input featureshavehighdimensionality. It separates thepositive andnegative examples asmuchaspossiblebyconstructingahyperplaneas thedecisionsurface. Thesupportvector regression(SVR)is theregressionextensionofSVM,whichprovidesanalternativeandpromising methodtosolve timeseriesmodelingandforecasting[51,52]. 3. LR is a popular statistical method for regression and prediction. It utilizes the ordinary least-squaresmethod or generalized least-squares tominimize the sumof squares of errors (SSE) forobtainingtheoptimal regressionfunction[53]. 273