资源描述：
《Time Series Forecasting Based on Augmented LSTM》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。
1、TimeSeriesForecastingBasedonAugmentedLongShort-TermMemoryDanielHsuJuly7,2017AbstractInthispaper,weuserecurrentautoencodermodeltopredictthetimeseriesinsingleandmultiplestepsahead.Previouspredictionmethods,suchasrecurrentneuralnetwork(RNN)anddeepbeliefnetwork(DBN)models,cannotlearnlong
2、termdependencies.Andconventionallongshort-termmemory(LSTM)modeldoesn'trememberrecentinputs.Com-biningLSTMandautoencoder(AE),theproposedmodelcancapturelong-termdependen-ciesacrossdatapointsandusesfeaturesextractedfromrecentobservationsforaugmentingLSTMatthesametime.Basedoncomprehensiv
3、eexperiments,weshowthatthepro-posedmethodssigni cantlyimprovesthestate-of-artperformanceonchaotictimeseriesbenchmarkandalsohasbetterperformanceonreal-worlddata.Bothsingle-outputandmultiple-outputpredictionsareinvestigated.1IntroductionTimeseriesforecastingandmodelingisanimportantinte
4、rdisciplinary eldofresearch,involvingamongothersComputerSciences,Statistics,andEconometrics.MadepopularbyBoxandJenkins[1]inthe1970s,traditionalmodelingprocedurescombinelinearautoregression(AR)andmovingaverage.But,sincedataarenowadaysabundantlyavailable,oftencomplexpatternsthatarenotl
5、inearcanbeextracted.So,theneedfornonlinearforecastingproceduresarises.Recently,neuralnetworkswithdeeparchitectureshaveproventobeverysuccessfulinimage,video,audioandlanguageleaningtasks[6].Intimeseriesforecastingarea,thoughtraditionallyshallowneuralnetworksaregenerallyadopted,thedeepn
6、euralnetworkshavealsoarousedenormousinterestsamongresearchers.Deepbeliefnetworks(DBN)arefrequentlyemployedincurrentshort-termtracforecasting[7][8],andpre-trainingstrategieswithunsupervisedlearningalgorithmssuchasRestrictedBoltzmannmachine(RBM)[9]andStackedAutoEncoder(SAE)[11]arealso
7、used.However,thesedeeparchitecturescannotcapturethelongdependenciesacrossdatapointswhicharebeyondinputobservations.RNNsareparticularlysuitableformodelingdynamicalsystemsastheyoperateoninputarXiv:1707.00666v2[cs.NE]6Jul2017informationaswellasatraceofpreviouslyacquiredinformation(dueto
8、recurrentcon
