Page - 298 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1561 wasdifficult. Therefore,weemployedMOSSAtorealize themulti-objectiveparameteroptimization. Furthermore, theoptimizationproblemcanbeexpressedas, argmin { PIW(θ) 1/CP(θ) (18) whereθ isasetofparameters inE–LUBE, includingtheweightandbias. Whentheparametersaredeterminedinthe trainingprocess, theentiremodelcanbeappliedto the test set toverify theperformanceof intervalprediction. Figure2.Forecastingflowchartof theproposedhybridmodel. 4. SimulationsandAnalyses Inorder tovalidate theperformanceof theproposedhybridmodel inSTLF, fourelectrical load datasets collected from four states inAustralia are used in our research. The four states include NewSouthWales (NSW),Tasmania (TAX),Queensland(QLD)andVictoria (VIC), andthespecific location is showed in Figure 3. The experiments in this study consist of twoparts: experiment I andexperiment II. For experiment I, the loaddata of four states aremodeledwith intervalwidth coefficientα=0.05,andfor theexperiment II, the intervalwidthcoefficientα is setas0.025 for further analysis. Inorder toverify thesuperiorityof theproposedhybridmodel, severalbenchmarkmodels which includebasicLUBE(LUBE),LUBEwithElmanneuralnetwork(E–LUBE),E–LUBEwithpoint optimization (PO–E–LUBE),E–LUBEwith intervaloptimization (IO–E–LUBE),andmodels integrated withCEEMDAN,areexhibited. Forpersuasivecomparabilityandfairness, thehyper-parameters in eachmodelareconsistent, asshowninTable1.Allexperimentshavebeencarriedout inMATLAB 298