Page - 74 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3433 Figure5.Weekly loadprofiledecompositionusingEMD(k=10). EMF,EMDIMF. Figure6.Weekly loadprofiledecompositionusingVMD(k=10).VMF,VMDIMF. The first VMF (VMF-1) is effectively the DC bias (Figure 6), i.e., the average daily load consumption.VMF-2andVMF-3showhighcorrelationsignalperiodicities.Officebuildings typically exhibit a commuteperiod, and this appears inVMF-2. ThisR&Dbuildinghas twopeaks around the commute time, and this pattern appears in VMF-3. On the other hand, EMF-10 and EMF-9 show high correlation trends, whereas the other EMFs show low correlations. High frequency EMFs(EMF-5–EMF-10)also includeend-pointproblems,whereasVMDdecomposes thesignal into band-limitedsignals;hence,VMFshavenoend-point issues. Table1showsthecorrelations foreachIMF.TheVMFscapturesimilar frequencysignalsbetter than the EMFs and decompose high frequency signals well. As VMD is done mathematically, thecorrelationbetweenVMFsisgraduallyreduced,whereasEMDIMFsare irregular. Therefore, in the case of high samplingor short prediction time scales, VMDshowsbetter performance thanEMD becauseVMDcanreflect thehighfrequencycharacteristicsof thedataset. 74