Page - 300 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1561 4.1.DataDescriptions For each state,we considered thedatausinghalf anhour interval in fourquarters. Thedata used in thispapercanbeobtainedonthewebsiteofAustralianenergymarketoperator (http://www. aemo.com.au/).Wechosedata fromthewholeof2017from1January20170:30amto31December 20170:00amtoconstructdataset. Ineachstate, the total samplenumber is17,520. Foreachquarter, thenumberofsampleswere4320,4358,4416,4416respectively. Inorder tocontrol thecomparability, weselected1200samplestotest thetrainedmodel,andusedtherest ineachquarter totrainthemodels. Theproportionof trainsetsversus the test setswasapproximatelyequal to3:1. Thedescriptionof thedatacharacteristicsareshowninFigure4.Consideringthestructureof theneuralnetwork in this study,weset six inputneurons,13hiddenneurons,andthreeoutputneurons. Specifically, theoutput setwas formulated inaccordancewithFormula (17). During data preprocessing, the input data were divided into several IMFs depending on CEEMDAN,asdisplayed inFigure2.Accordingto theenergyentropyofeachIMFshowninFigure3, weignoredtheIMFswhichcontainedhighfrequencies,andsummedtherestof theIMFstoreconstruct the inputset, asshowninFigure1. Figure4.Boxplotof theentiresetofdatasamples. 4.2. PerformanceMetrics Inorder tocomprehensivelyassess theperformanceof themodels, somemetricswereemployed. Thesemetricsprimarily focusedonthecoverageof therealvalue in theprediction intervalandthe widthof the interval. 4.2.1.CoverageProbability Coverageprobability [50] is usually consideredas abasic feature of PIs andCP is calculated accordingto theratioof thenumberof targetvaluescoveredbyPIs: CP= 1 m m ∑ i=1 θi (19) 300