Research On LSTM-based Correlated Time Series Prediction

Correlated time series are ubiquitous in various real-world scenarios,such as time series of traffic passenger demand in different areas and time series of air pollutant concentration of multiple monitoring stations.Accurate prediction of correlated time series has important research and application significance.For example,in the transportation domain,if passenger demand can be predicted accurately,it will greatly help to sense market dynamics in advance,to pre-allocate transportation resources,and to enhance passenger travel experience.If concentration of air pollutants can be predicted effectively,it will help people to take preventive measures beforehand so as to prevent the occurrence of disease.However,correlated time series prediction is a very challenging task.Since correlated time series have strong intrinsic features.There are both inner-sequence temporal dependencies and inter-sequence spatial dependencies in correlated time series.Effectively modeling the spatio-temporal dependencies of correlated time series is the key for correlated time series prediction.While most of the traditional time series prediction methods only focus on single sequence individually,it is difficult for them to capture the spatial dependencies among correlated time series.Therefore,the correlated time series are considered as a whole in this thesis,to analyze and model the intrinsic spatio-temporal features of correlated time series.The main research contents and innovations of this thesis are summarized as follows:This thesis first analyzes the intrinsic features of correlated time series,and finds that temporal dependencies and spatial dependencies both play an important role in correlated time series prediction tasks.Based on the advantages of LSTM in modeling time series on the temporal dimension,a novel dual channel LSTM is proposed to model the spatio-temporal correlations of correlated time series.The dual channel LSTM modifies the internal structure of the traditional LSTM,employing two channels to explicitly capture spatio-temporal correlations,and utilizing a priori relational matrices to describe the relations among sequences.Experiments on air quality datasets and traffic datasets show the versatility and effectiveness of the dual channel LSTM method on modeling spatio-temporal correlation of correlated time series.There are extensive applications of correlated time series predictions in the traffic domain.Oriented towards the application tasks in the traffic domain,a correlated time series prediction method(CTS-LSTM)is proposed in this thesis.Tightly integrated with the traffic business scenario,based on the dual channel LSTM,CTS-LSTM can adaptively learn the spatial dependence in traffic domain,and effectively integrate the temporal and spatial correlation of traffic sequences in order to model the intrinsic characteristics of traffic sequences more accurately.Besides an external factor modeling module is designed to capture the impact of external information on traffic sequences.The final prediction results are jointly determined by the intrinsic characteristics of the traffic sequence and the external factors.Experiments on two real-world traffic prediction tasks show that the CTS-LSTM proposed in this thesis can comprehensively model both the complex spatio-ternporal patterns of traffic data and external influence factors,and its prediction performance is significantly better than the existing methods.