Seite - 71 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3433 4. Experiments Thissectiondescribes theprocessusedtoobtain timeseriesmodels for loadforecasting. Figure3 shows the proposed load forecasting model using LSTM with multi-decomposition for feature extraction.Wewilldiscusseachstep indetail. Figure3.Deep-learning loadforecastingbasedonthemulti-decompositionmethod. 4.1. Predictionof theTimeScale Reference loadprofiles reflect the loadprofiles thatareclose toreal-time loadprofilesbeforeh stepsago,wherehdetermines thepredictiontimescales,whichdependonthepurposeof the load forecasting. STLF techniques canbeused foravarietyofpurposesbyenabling smaller scales and fasterprediction.USTLF,whichpredicts the loadwithinafewminutes toonehour, canbeusedfor electricity theftdetectionorcanprovide informationforemergencypowersupply[47]. STLF,which predicts the loadfromonehour toaday, canbeusedforelectricity transactionsoreconomicdispatch of renewableenergyresources [2]. 4.2. ExtractFeatureLayer Through the multi-decomposition method, the features of time-series data are extracted. Thenumberofdecomposition levels (K) is10,which is thevalueobtainedwhenthedecomposition loss rate is0.1%or less. Theweightof theweekly loadprofile (Dp) considers the trendof loadpatterns accordingtoseasonalchanges. EachIMFdecomposedthroughtheVMDhasafrequencycharacteristic andisnormalizedtomakethe featurestandout (Nk). 71