Seite - 310 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1561 5.Discussion In this section,wediscuss some factorswhichmayhaveaneffect on theperformancesof the proposedmodels inorder to improve thepracticabilityofourhybridmodel. The factors involved mainly includethe featuresof thedatasetsandthesettingof thehyperparameters in thealgorithm. 5.1.DatasetFeatures The feature and quality of the datasets have a significant effect on the performance of the predictionmodels. InSTLF, thedata showsperiodicityandvolatility. Theperiodicity is attributed to theregularity in theactualuseofelectricity,andthevolatility isattributedto therandomnessand occasionaluseofelectricity. Therefore, the linearcomponentandthenon-linearcomponentsoperate simultaneouslyduringtheforecastingof themodel. Specifically, someoutliersmayhaveanegative effect in theprocessofprediction. AsFigure4 shows, thedataset featuresof thedifferent samplesarevarious. According to the boxplot theory, thedatapoints that are larger thanQ3+1.5IQRor smaller thanQ1− 1.5IQRare regardedasoutliers. For thefirstandfourthquarters inNSW,andthefirstandfourthquarters inVIC, thedistributionsof thedatasetsdisplayedanumberofoutliers.Additionally, theresultsof themodels showninTables2–5demonstrate that themodelperformanceof thesamplewhosedistribution isnot desiredmaybeunremarkable. Theseoutliersare important factors that leadtosuchresults, even if the CEEMDANmodelhasbeenapplied indatapreprocessing. Another set of data features that may cause an unsatisfactory result are the non-linear characteristicsof thedataset. It iswellknownthat in traditional research, thepredictionof regularand linear timeseriesareeasy toreachthedesiredaccuracy.However,unstableandnon-linear timeseries aremoredifficult to forecast inspiteof theapplicationsofnovelmodels, suchas thecaseofmachine learningalgorithms. Amethodused tomeasure the instabilityofdata series is the recurrenceplot (RP) [56].Arecurrenceplot isanadvancedtechniqueofnon-lineardataanalyses. It is thevisualization (oragraph)ofasquarematrix inwhichthematrixelementscorrespondtothose timesatwhichthe stateof thedynamical systemrecurs. Stationarysystemswilldeliverhomogeneousrecurrenceplots, andunstablesystemscausechanges inthedistributionofrecurrencepoints in theplot,which isvisible and identifiableby thebrightenedareas. In this study,weselectedVICasanexample toverify the influenceof instability. Beforedrawingtherecurrenceplot, the timedelayandthedimensionof the embeddedmatrixweredeterminedbytheC–Cmethod.Dependingonthe“CRPToolbox”releasedby NorbertMarwan[57], therecurrenceplotof the fourdatasetsof thedifferentquarters inVICisshown inFigure8. Figure8.Cont. 310