Research On Traffic Flow Data Analysis And Prediction Based On Deep Spatio-Temporal Network

With the development of modern society and the promotion of urbanization,the number of population and vehicles is increasing rapidly,and the urban transportation system is facing great pressure.Based on the massive traffic flow data,Intelligent Transportation Systems(ITS)can effectively and comprehensively apply advanced science and technology to the traffic field,dig out the potential traffic change rules,and effectively improve the efficiency of utilization of traffic network,avoid traffic congestion,save traffic resources and so on.In the study of Intelligent Transportation Systems,the processing and analysis of traffic flow data and its application are particularly important.Urban traffic flow data has stochastic and non-linear characteristics,and also possesses strong spatio-temporal correlation.Existing methods are often unable to excavate the hidden temporal and spatial laws in the data well enough to cope with the current complex traffic network conditions.Therefore,this thesis proposes to analyze the spatio-temporal characteristics of traffic flow data through the deep spatio-temporal network,and based on this,carry out specific research on the two problems of traffic data restoration and traffic flow prediction.The main research contents include the following aspects:(1)This thesis proposes a data restoration model combining temporal and spatial feature analysis.For massive traffic data,due to sensor failure,communication failure,storage loss and other problems,the acquired data will inevitably be missing.Therefore,the model proposed in this thesis combines spatial self-attention mechanism with bi-directional recurrent neural network to mine the spatio-temporal features of data.Based on the auto-encoder model and the gating unit,the data which is close to the real distribution is generated to improve the repair effect.(2)This thesis proposes a traffic flow prediction model based on a multi-graph gated graph convolution framework.An accurate and effective traffic flow prediction system not only can effectively avoid traffic problems such as traffic congestion,can also provide data for other complex task,so in order to improve the feature mining capability of the model and obtain higher prediction accuracy,this thesis proposes to model historical traffic data separately by constructing temporal and spatial modules.The temporal module extracts temporal correlation through a one-dimensional convolutional neural network based on channel attention mechanism and ”inception” structure.The spatial module extracts complex spatial correlations through an interpretable multi-graph gated graph convolution framework.Finally,the prediction accuracy of the model is further improved by stacking spatio-temporal blocks to deepen the network.In conclusion,this thesis mainly research on the restoration and prediction of traffic data,in order to verify the validity of the proposed models,a large number of experiments were carried out in real traffic data sets,the results show that compared to the baseline models,the proposed models is able to extract features of traffic flow data better,which helps in the subsequent traffic control tasks and thus provides a feasible idea for relevant applications of intelligent transportation.