Page - 311 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1561 Figure8.Recurrenceplotof thesamplesobtainedfromthefourquarters inVIC. As the figure shows, the second and third quarters in VIC display relatively homogeneous distributions,whileotherquartersdisplay isolatedbrightenedareas.Accordingto the theoryof the recurrenceplot, the instabilitiesof the former twosamples areweaker, and theother twosamples reveal stronger instabilities. Furthermore,wecanconcludethat theperformancesof the forecasting models shown inTable 5 are remarkablewhen thedataset is relatively stable,while theunstable dataset results inanunsatisfactoryperformance,whichcannotbeavoided. 5.2. SensitivityAnalysis Thehybridmodelproposedin this studyisbasedonthestructureof theneuralnetworkshown in Figure 1. In the hybridmodel, the hyperparameter is a key factor that influences themodel’s performance. Inmoststudiesonmachine learning, thesettingof thehyperparametersalwaysdepends on trials or empirical knowledge. This is the reasonwhymany experimental results cannot be reproduced andwhy a considerable amount of time and energy is spent on tuning parameters in industrial applications. At present, there is no absolutemethod todetermine the values of all typesofhyperparameters. In this study,wealsomainly reliedonexperiences and trials to set the hyperparameters, as shown inTable 1. Among thehyperparameters, several parameters need to behighlighted. Thefirstone is thenumberofsalppopulations inMOSSA. In theswarmheuristicoptimization algorithm, the number of swarms is usually a vital factor that needs to be considered. A larger populationmightprovidea largerprobability toreachthebest individual,butexceedingthedesired populationmaycauseanincrease in thecomplexityof thealgorithm,which is relatedto thenumberof algorithmic iterations.Considering thenumberofparameters inourproposedmodel, thepopulation numbers that ranged from10 to100witha stepof 10wereevaluated. Asa result,weselected the number50as thepopulationnumber(asshowninTable6)aftercomprehensivelyconsideringthetime complexityandmodelperformance. Thesecondtypeofhyperparameters thatneedtobeemphasizedare theupperandlowerbounds of individualparameters inMOSSA. Inoursimulation, thedatasetswerenormalizedwithin therange of−1 to1 inorder toavoid the influenceofdimensionand improve the trainingspeed. Therefore, theabsolutevalueofweightsandthresholdsofneuralnetworks in the trainingprocesswillnotbe too large. AsTable 6 shows,we set the initial upper and lower bounds to 2 and−2 according to the experiment trials. Excessive range limitsmay increase the difficulty of searching for the best parameterswitha limitednumberof iterations. Furthermore, thealgorithmthatoperatesbasedona small rangemaynotelicit theoptimalsolution. 311