Seite - 65 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

energies Article DeepLearningBasedonMulti-Decompositionfor Short-TermLoadForecasting SeonHyeogKim ,GyulLee ,Gu-YoungKwon ,Do-InKimandYong-JuneShin* DepartmentofElectricalandElectronicEngineering,YonseiUniversity,Seoul03722,Korea; goodguy7@yonsei.ac.kr (S.H.K.); thyecho@yonsei.ac.kr (G.L.);kgy926@yonsei.ac.kr (G.-Y.K.); penpony109@yonsei.ac.kr (D.-I.K.) * Correspondence: yongjune@yonsei.ac.kr;Tel.: +82-2-2123-4625 Received: 31October2018;Accepted: 3December2018;Published: 7December2018 Abstract:Loadforecasting isakey issueforefficient real-timeenergymanagement insmartgrids. Tocontrol the loadusingdemandsidemanagementaccurately, loadforecastingshouldbepredicted in theshort term.Withtheadventofadvancedmeasuring infrastructure, it ispossible tomeasure energyconsumptionatsamplingratesuptoevery5minandanalyze the loadprofileofsmall-scale energygroups, suchas individualbuildings. Thispaperpresentsapplicationsofdeep learningusing featuredecompositionfor improvingtheaccuracyof loadforecasting. Theloadprofile isdecomposed intoaweekly loadprofileandthendecomposedinto intrinsicmodefunctionsbyvariationalmode decomposition to capture periodic features. Then, a long short-termmemorynetworkmodel is trainedbythree-dimensional inputdatawith three-stepregularization. Finally, thepredictionresults of all intrinsicmode functions are combinedwith advancedmeasuring infrastructuremeasured in the previous steps to determine an aggregated output for load forecasting. The results are validated by applications to real-world data from smart buildings, and the performance of the proposed approach is assessed by comparing the predicted results with those of conventional methods,nonlinearautoregressivenetworkswithexogenous inputs,andlongshort-termmemory network-basedfeaturedecomposition. Keywords:deeplearning;empiricalmodedecomposition(EMD); longshort-termmemory(LSTM); loadforecasting;neuralnetworks;variationalmodedecomposition(VMD);weeklydecomposition 1. Introduction Accurate loadforecastingoptimizespower loads, reducingcostsandstabilizingelectricpower distribution. Load forecasting accuracy depends on the time series data of non-stationary and non-linearity characteristics. These characteristics are influencedby theprediction time scale and energyconsumptionscale.Dependingonthepredictiontimescale, loadforecasting isclassifiedinto four types. Long-term load forecasting (LTLF) has a time scale ofmore than a year, medium-term load forecasting(MTLF)a timescale fromoneweektooneyear,andshort-termloadforecasting(STLF)a timescale fromonehour tooneweek. Systemoperators typicallyestimatedemandbyreferring to loadprofiles fromseveralhoursago.Ultra-short-termloadforecasting(USTLF) isakeyissueforsmart grids, real-timedemandsidemanagement (DSM),andenergytransactionsbecauseenergytrading in DSMrequiresprecise loadforecasting in theorderofminutes,andprofit is stronglyrelatedto forecast accuracy. Therefore, theUSTLFtimescale is fromseveralminutes toonehour [1]. Conventional loadforecastingmethodsusestatisticalmodelsbasedoninherentcharacteristics ofhistoricaldata. PreviousSTLFstudieshaveproposedauto-regressive integratedmovingaverage (ARIMA),Gaussianprocessingregression(GPR), supportvector regression(SVR),andneuralnetwork Energies2018,11, 3433;doi:10.3390/en11123433 www.mdpi.com/journal/energies65