Seite - 373 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1138 anpowerautoregressiveconditionalheteroskedasticity (PARCH)methodwaspresented forbetter performance [6]. These statisticmodels need fewer historical data and have a small amount of calculation.However, theyrequireahigherstabilityof theoriginal timesequencesanddonotconsider the uncertain factors such asweather andholidays. Therefore, artificial intelligentmodels about forecasting, suchasneuralnetworks [7,8], fuzzy logicmethod[9]andsupportvector regression[10], wereproposedwith thedevelopmentofcomputerscienceandsmartgrids. Recently,neuralnetworks arebecominganactiveresearch topic in theareaofartificial intelligence for its self-learningandfault tolerantability. Someeffectivemethodologies for loadforecastingbasedonneuralnetworkshavebeen proposed inrecentyears.Aneuralnetworkbasedmethodfor theconstructionofprediction intervals wasproposedbyQuanetal. [7]. Lowerupperboundestimationwasappliedandextendedtodevelop prediction intervalsusingneuralnetworkmodels. Themethodresulted inhigherquality fordifferent typesofpredictiontasks.Ding[8]usedseparatepredictivemodelsbasedonneuralnetworks for the dailyaveragepowerandthedaypowervariationforecasting indistributionsystems. Theprediction accuracywas improvedwithrespect tonaivemodelsandtimesequencemodels. The improvement of forecastingaccuracy cannotbe ignored, but,with the increasingcomplexityandscaleofpower grids,high-accuracy loadforecastingwithadvancednetworkmodelandmulti-source information is required. Deeplearning,proposedbyHinton[11,12],madeagreat impactonmanyresearchareas including fault diagnosis [13,14] and load forecasting [15–17] by its strong learning ability. Recurrent neural network (RNNs), adeep learning framework, aregoodatdealingwith temporaldatabecauseof its interconnectedhiddenunits. Ithasprovensuccessful inapplicationsforspeechrecognition[18,19], image captioning [20,21], and natural language processing [22,23]. Similarly, during the process of load forecasting,weneed tomine andanalyse large quantities of temporal data tomake aprediction of timesequences. Therefore,RNNsareaneffectivemethod for load forecasting inpowergrids [24,25]. However, thevanishinggradientproblemlimits theperformanceoforiginalRNNs.Thelater timenodes’ perceptionof thepreviousonesdecreaseswhenRNNsbecomedeep.Tosolvethisproblem,animproved networkarchitecturecalled longshort-termmemory(LSTM)networks [26]wereproposed,andhave provensuccessful indealingwithtimesequencesforpowergridsfaults [27,28].Researchonshort-term loadforecastingbasedonLSTMnetworkswasput forward.Gensleretal. [29]showedthecompared results forsolarpowerforecastingaboutphysicalphotovoltaic forecastingmodel,multi-layerperception, deepbeliefnetworksandauto-LSTMnetworks. ItprovedtheLSTMnetworkswithautoencoderhad thelowesterror.Zhengetal. [30] tackledthechallengeofshort-termloadforecastingwithproposinga novelschemebasedonLSTMnetworks.TheresultsshowedthatLSTM-basedforecastingmethodcan outperformtraditional forecastingmethods.Aimingatshort-termloadforecastingforbothindividual andaggregatedresidential loads,Kongetal. [31]proposedanLSTMrecurrentneuralnetworkbased frameworkwiththe inputofdayindicesandholidaymarks.Multiplebenchmarkswere testedinthe real-worlddataset and theproposedLSTMframeworkachieved thebestperformance. The research worksmentionedaboveindicatethesuccessfulapplicationofLSTMfor loadforecastinginpowergrids. However, loadforecastingneedstobefastandaccurate. TheprincipleandstructureofLSTMarecomplex with inputgate, outputgate, forgetgateandcell, so thecalculation isheavyfor forecasting ina large scalegrid.Gatedrecurrentunit (GRU)neuralnetworkswasproposedin2014[32],whichcombinedthe inputgateandforgetgatetoasinglegatecalledupdategate. ThemodelofaGRUissimplercompared withanLSTMblock. Itwasprovedonmusicdatasetsandubisolfdatasets thatGRU’sperformanceis betterwith lessparametersaboutconvergence timeandrequired trainingepoches [33]. Luetal. [34] proposedamulti-layerself-normalizingGRUmodelforshort-termelectricityloadforecastingtoovercome theexplodingandvanishinggradientproblem.However,short-termloadforecastingforcustomers is influencedbyfactors includingdate,weatherandtemperature,whichpreviousresearchdidnotconsider seriously. Peoplemayneedmoreenergywhenthedayiscoldorhot. Enterprisesor factoriesmayreduce theirpowerconsumptiononholidays. 373