Seite - 66 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3433 models [2].ARIMAisacommonmethodfor linear timeseries-basedmethods.GPRandSVRprovide alternativemethods tomodel timeseries loads,usingexternaldatasuchasweatherdata toconsider non-linearityandnon-stationarity.GPRisasupervisedmachine learningmodelbasedonstatistical regressionandakernel functionthat refinesvarianceandstep length[3]. Toreduce thenonlinearityof the timeseriesdataandtoanalyze their statistical characteristics, a seasonalanalysiscombinedpredictionmethodisused[4,5]. Recent researchactivitiesdivideprofiles intosub-profilesaccording to the loadpatternsof customersbasedonhumanfactor, contract type, andregion.Afterdividingtheprofiles intosub-profiles,aclusteringalgorithmisusedforhierarchical classification[6–9]. To improvetheaccuracyof loadforecastingusingexternaldatasuchas temperature,humidity, weather information,andelectricityprices, amethodhasbeenproposed[10–13].However,measuring such data is a challenging task for low-level distribution and small-scale loads. Furthermore, data processing anddata storage of eachpiece of the dataset are required because the resolution of time-series data is different. Therefore, recent research trends use the technique of feature selection[14,15]ordecomposingthe loadprofile toextract thecharacteristicsof the loadusingsignal processingtheory[16–24]. Waveletdecompositionwithneuralnetworks [1,16–18]hasbeenemployedto increaseprediction accuracy. In [1],awaveletalgorithmdealtwith thenoiseof theactualelectrical loaddata,andload forecastingbasedonartificialneuralnetworks (ANN)wasproposed. Empiricalmodedecomposition (EMD)withmachine learninghasalsobeenproposed for load forecasting,windspeed, or energy prices [19–22]. However, EMD lacks amathematical definition andhasweaknesses that diverge at end-pointswhendecomposing the signal. Toovercome theweaknessofEMD, load forecasting studiesusingvariationalmodedecomposition(VMD)havebeenproposed[22–25]. Existingregression methodswithvariousdecompositions, clusteringalgorithms,andprobabilisticanalyseshavebeen investigated,as theycanbeusedto identify loadcharacteristics;however, they increase thedimension of the input [26–28].Clusteringanddecompositionmethodsareapplied in thepre-processingstage to improvetheaccuracyof the loadprediction,andcurrentstate-of-art loadforecastingstudieshave improvedtheperformanceof thepredictionmodel throughdeeplearning[29–33]. A recurrent neural network (RNN) has amemory structure and a hidden layer suitable for processing big data using deep learning techniques. However, an RNNhas vanishing gradient problemscausedbyan increase in thenumberof layers.Nonlinearautoregressiveexogenous (NARX) RNNsofferanorthogonalmechanismfordealingwith thevanishinggradientproblembyallowing direct connectionsordelays fromthedistantpastdata [34,35].However,NARXRNNshavea limited impactonvanishinggradients,andthedelaystructureincreasesthecomputationtime.Mostsuccessful RNNarchitectureshave longshort-termmemory (LSTM),whichusesnearlyadditive connections between states, to alleviate the vanishing gradient problem [36–41]. In [42], gated recurrent unit (GRU) neural networkswithK-mean clusteringwere proposed. AGRU is a variant LSTMwith a simpler structure, but it has similar performance, and convolutional neural networks (CNNs) arealsowidelyusedindeeplearningfor imageclassification[43,44].As the loadpredictionmodel becomesmoresophisticated,shorterpredictiontimescales[1,5]andlowerlevelfeedersofdistributions, suchasbehind-the-meter individual load,businessbuildings,andhouseholdelectricusage,arebeing studied[6,26–28]. This paper proposes a deep learning method whereby features are extracted through multi-decompositionforshort-termloadforecasting. Thescaleof thepredicted loadisa feeder-level business building. Feeder-level load forecasting ismore complicated than that of an aggregated loadbecause the statistical characteristics aregreatly changedevenwitha slight change inpower consumption. The proposed decompositionmethod significantly captures intrinsic load pattern componentsandperiodic features. AloadforecastingmethodbasedonLSTMwithVMDisdesignedandimplemented in thispaper. Theproposedtwo-stagedecompositionanalysis identifies thecharacteristicsof the loadprofilewith 66