Page - 222 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1900 3.5. LoadPredictionwith theWeatherForecastData 3.5.1.DataPreprocessingBasedonMCM Werecordeddailyweather forecastdata for Julyexcept forweekendswitha totalof171samples. Throughanalysis, it is foundthat theerrorsbetweenweather forecastdataandrealweatherdataobey thenormaldistributionN[μ,σ2]. In order to study the influenceof theuncertaintyof theweather forecastdataontheaccuracyof loadforecasting,weuse theMCMtomodify the inputparametersof theSVMmodel,namely, themeteorologicaldata. Then, thepreprocessedweather forecastdataare input into themodel for loadforecasting. The IBM SPSS Statistics 19.0 software is used to analyze the error distribution of seven meteorological inputparameters in turn. Figure8showstheerrorprobabilitydistributionbetweenthe weather forecast temperatureT(h–1)andtheactual temperatureonehourbefore thepredictedtime. Themeanvalue isμ=0.588andthestandarddeviation isσ=1.799. Figure8.TheerrorprobabilitydistributionofT(h–1). Next, wewrite the programofMCM intoMATLAB to implement theMonteCarlo random sampling of its error Δw, setting the number of simulationsMas 1000, anduse a corresponding calculationprogramtoobtainasetof revisedweather forecastdataT(h–1)*. The formula isas follows: T(h−1)∗=T(h−1)+Δw (22) Forexample, for forecasting the loadat9o’clockon3July, it isknownthat the1-h-aheadweather forecastdrybulb temperatureof thepredictionmomentT(h–1) is27.2 ◦C.Theresultof therandom samplingforT(h–1)usingtheMCMbasedon1000runsof themodel is showninFigure9. Themost frequent values ofT(h–1) in the results of randomsampling simulationarenear 26.7 ◦C.Actually, the errorΔwobtainedby randomsampling is−0.6 through calculation, and the revisedweather forecastdataT(h–1)* is26.6 ◦C,whichmeans that theexpectedvalueofT(h–1) is26.6 ◦Candiscloser to therealweatherdata, i.e., 26.0 ◦C. 222