Page - 374 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1138 In thispaper,amethodbasedonGRUneuralnetworkswithmulti-source inputdata isproposed forshort-termloadforecasting inpowergrids.Moreover, thispaper focusesonthe loadforecasting for individualcustomers,which isan importantandtoughproblembecauseof thehighvolatilityand uncertainty [30]. Therefore,before training thenetworks,wepreprocess thecustomers’ loaddatawith clusteringanalysis toreduce the interferenceof theelectricityusecharacteristics. Then, thecustomers are classified into three categories to form the training and test samples by K-means clustering algorithm. Toobtainnotonly the loadmeasurementdatabutalso the important factors including date,weatherandtemperature, the inputof thenetworkaresetas twoparts. Thetemporal features of loadmeasurementdataareextractedbyGRUneuralnetworks. Themerge layer isbuilt to fuse themulti-source features. Then,wecanget the forecasting results by training thewholenetwork. Themethodologies are described indetail in Section 2. Themain contributions of this paper are as follows. 1. Trainedsamplesare formedbyclusteringtoreducethe interferenceofdifferentcharacteristics ofcustomers. 2. Multi-sourcedata includingdate,weatherandtemperaturearequantifiedfor inputsothat the networksobtainmore informationfor loadforecasting. 3. TheGRUunitsare introducedformoreaccurateandfaster loadforecastingof individualcustomers. In general, the proposed method uses the clustering algorithm, quantified multi-source informationandGRUneuralnetwork for short-termload forecasting,whichpast researchdidnot considercomprehensively. The independentexperiments in thepaperverify theadvantagesof the proposedmethod. The rest of thepaper isorganizedas follows. ThemethodologybasedonGRU NeuralNetworks for short-term load forecasting is proposed in Section 2. Then, the results and discussionof thesimulationexperimentsaredescribedtoprove theavailabilityandsuperiorityof the proposedmethodinSection3. Finally, theconclusion ismade inSection4. 2.MethodologyBasedonGRUNeuralNetworks Inthissection, themethodologyisproposedforshort-termloadforecastingwithmulti-sourcedata usingGRUNeuralNetworks. First, thebasicmodelofGRUneuralnetworksare introduced[32]. Then, datadescriptionandprocessingareelaborated. The loaddataareclusteredbyK-meansclustering algorithmsothat the loadsampleswithsimilarcharacteristics ina fewcategoriesareobtained. This helps improve theperformanceof loadforecasting for individualcustomers. In the last subsection, the wholeproposedmodelbasedonGRUneuralnetworks is shownindetail. 2.1.Model ofGRUNeuralNetworks GatedrecurrentunitneuralnetworksaretheimprovementframeworkbasedonRNNs.RNNsare improvedartificialneuralnetworkswith the temporal inputandoutput.Originalneuralnetworks onlyhaveconnectionsbetweentheunits indifferent layers.However, inRNNs, thereareconnections betweenhiddenunits formingadirectedcycle in thesamelayer. Thenetworktransmits the temporal informationthroughtheseconnections. Therefore, theRNNsoutperformconventionalneuralnetworks inextractingthe temporal featuresbytheseconnections.Asimplestructure foranRNNisshownin Figure1. The inputandoutputare timesequences,which isdifferent fromoriginalneuralnetworks. Theprocessof forwardpropagation isshowninFigure1andgivenbyEquations (1)–(3). ath= I ∑ i=1 wihxti+ H ∑ h′=1 wh′hs t−1 h′ (1) sth= fh(a t h) (2) ato= H ∑ h=1 whosth (3) 374