Seite - 379 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1138 bepredicted locates in. The loadmeasurementdataof individualcustomers inonedayisextracted asa sample for short-term load forecasting, notedbyP. ThedimensionofP is 96with the15min samplingperiod. Then, thesamplesare reshaped into two-dimension for the inputofGRUneural networks.ConsideringtheinfluencingfactorsdateD,weatherWandtemperatureT,dateDp,weather Wp and temperatureTp on the forecastingday are added to the another input of theGRUneural networks.Consideringthegeneral factorofdate, thepredictioninterval isset tosevendays. Therefore, the loadmeasurementdataPl on theday in the lastweek fromthe forecastingday,Dp,Wp andTp, arerecordedas theoverall input. The loadmeasurementdataPponthe forecastingdayarerecorded as theoutput,whosedimension is96. Therefore, the inputX andoutputYofsamplesaregivenby Equations (14)and(15). X={Pl; Dp, Wp, Tp} (14) Y={Pp} (15) Figure5.SchematicdiagramofproposedframeworkbasedonGRUNeuralNetworks forshort-term loadforecasting,wherek is thenumberofhiddenunitsand t is the timestep. TheparametersofGRU unitsareclarified inSection2.1. The inputandoutputparametersareexplained in thenext subsection. The features fromGRUneuralnetworksandfullyconnectedneuralnetworkaremergedwith the concatenatingmodeandpasses throughbatchnormalizationanddropout layer toavoidoverfitting andincrease the learningefficiency. Theprinciple is thatbatchnormalizationcanavoidthegradient vanishingof falling into the saturatedzone, and that thebetterperformance infixedcombination isavoidedwhenrandomneuronsdonotwork inadropout layer. Then, two-layer fullyconnected neural networkare addedbefore theoutput for learningandgeneralizationability. With training byback-propagation throughtime, thewholenetwork implements theshort-termloadforecasting 379