Seite - 78 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3433 profile, the loadprofilewasdecomposedbyaweeklyseasonalityandvariationalmodedecomposition. Twodecompositionmethodscan identify featuressuchasseasonality, load increase/decreasepattern, andperiodicitywithoutanyexternaldata, suchas temperature. Thethree-stepregularizationprocessreducedthelong-termdependency,overfitting,andcovariate shiftproblemcausedby featuredecomposition,which increases thedata samplesanddimensions. The results also reveal the effectivenessof the long short-termmemoryneural networksbasedon variationalmodedecompositionwithdifferentpredictiontimescales.Weexpect theproposedmethod tobeakeytechniquefordemandsidemanagement,electricalpower theftdetection,energystorage systemscheduling,andenergytradingplatforms in futuresmartgrids. AuthorContributions: S.H.K.developed themain ideaanddesigned theproposedmodel; he conducted the simulation studies andwrote the paperwith the support of G.L. andG.-Y.K. under the supervision of the correspondingauthor,Y.-J.S.D.-I.K. contributedto theeditingof thepaper.Allauthorshavereadandapproved thefinalmanuscript. Acknowledgments:ThisworkwassupportedbyaNationalResearchFoundationofKorea (NRF)grant funded bytheMinistryofScience, ICT&FuturePlanning(No.NRF-2017R1A2A1A05001022)andthe frameworkof the internationalcooperationprogrammanagedbytheNRFofKorea (No.NRF-2017K1A4A3013579). This research wasalsosupportedbytheKoreaElectricPowerCorporation(KEPCO)(No.R18XA05). Conflictsof Interest:Theauthorsdeclarenoconflictsof interest. Nomenclature AMI advancedmeasuring infrastructure ANN artificialneuralnetwork LSTM longshort-termmemory EMD empiricalmodedecomposition VMD variationalmodedecomposition LTLF long-termloadforecasting MTLF medium-termloadforecasting STLF short-termloadforecasting USTLF ultra-short-termloadforecasting DSM demandsidemanagement ARIMA auto-regressive integratedmovingaverage GPR Gaussianprocessingregression GRU gatedrecurrentunit SVR supportvector regression RNN recurrentneuralnetwork NARX nonlinearautoregressiveexogenous CNN convolutionalneuralnetwork IMFs intrinsicmodefunctions k themodeindex vk kth intrinsicmode Wp(t) pthweekly loadprofile w frequencyofmode K the totalnumberofmodes xt the typical loadprofile xpt the load p thweeklyseasonality feature xpk,t k th IMFof the load pthweeklyseasonality feature δ theDiracdistribution Dp theweeklydecay’sexponent factor Nk the IMFnormalizationfactor BN batchnormalization st thememorycell stateofLSTM ft the forgetgateofLSTM 78