Traffic Speed Prediction Based On Temporal Graph Convolutional Network

Alleviating road congestion,improving road traffic efficiency and road travel environment have become the new requirements of traffic management departments for the construction of new urban intelligent transportation system.With the rapid development of digital control,computer intelligence,communication and other technologies and the rapid popularization of internet applications,it provides massive data support for urban traffic state prediction.Timely,efficient and accurate forecasting of traffic is essential for the regulation and direction of urban traffic,which can provide traffic managers with scientific urban operation traffic planning and powerful decision-making basis,and help travelers choose clearer travel roads,so as to reduce traffic congestion and improve daily commuting efficiency.In this thesis,based on the traffic data of the expressway network,a dynamic weighted complex network containing topological spatial information and traffic state is constructed,and the characteristics of the weighted complex network are analyzed.Combined with the theoretical method of graph neural network,a multi-step traffic speed prediction model of gated temporal graph convolution network based on hybrid deep learning framework is constructed.The specific research contents are as follow:(1)Based on the big data of highway traffic and combined with complex network theory,the topological characteristics of the traffic network are studied.The traffic network based on weighted complex network is constructed with sensors as nodes and the distance between nodes and vehicle speed as weights.respectively,The topological characteristics of the traffic network are studied by the degree distribution,point intensity,weighted clustering coefficient,weighted average node degree and other metrics.The degree distribution characteristics indicate that the traffic network is neither a random network nor a scale-free network;the study of point intensity,weighted clustering coefficient,weighted average node degree and other metrics indicate that the traffic network structure shows dynamic evolution characteristics and the network topology changes with time.(2)The expression ability of multi-layer automatic feature extraction of deep learning is deeply fused with the nonlinear,multimodal,and spatio-temporal correlation characteristics of traffic,and a multi-step traffic speed prediction model based on a hybrid deep learning framework is constructed by gated temporal graph convolution network(GT-GCN).Firstly,the dynamic weighted graph network is constructed according to the traffic network structure.Then,the graph convolution network is introduced to process the graph structure data of the traffic network,and the spatial characteristics of the traffic state are mined.Furthermore,gated time convolutional networks are used to capture the short-range and long-range time dependencies of traffic speed.Finally,the spatial characteristics of road network and the spatio-temporal characteristics of traffic speed are integrated.Experiments on real data show that the proposed model can describe the spatio-temporal evolution characteristics of traffic speed,and has good multi-step prediction performance compared with the baseline models.