A Dynamic Spatiotemporal Feature-driven Method For Freeway Network Traffic State Prediction

With the continuous growth of Chinese economy,freeways are becoming more and more important to the transportation system,and the increasing traffic congestion and traffic accidents has reduced the operating efficiency of the freeway transportation system.As the basis of control measures for intelligent transportation system,traffic flow prediction is of great significance to the traffic control and management of freeway network.Due to the complex spatiotemporal dependencies between traffic flows,the task of traffic flow prediction faces great challenges in parsing spatiotemporal relationships.At present,the existing representation of the network spatial connection does not comprehensively describe the spatial relationship between traffic nodes,and cannot reflect the dynamic change of the connection strength.In addition,most traffic flow prediction methods do not make full use of the spatiotemporal dependence,and it is difficult to mine the deep spatiotemporal correlation in the traffic flow data.Aiming at the incomplete description of network spatial relationship,this thesis proposes a spatial connectivity matrix,which combines the structural attributes with the traffic flow features to describe the connection relationship three-dimensionally.First,the structure of the freeway network is represented by a topological graph,and the deep adjacency relationship in the road network is described by the reachability matrix.On the basis of the adjacency relationship,the spatial connectivity matrix uses the traffic flow sequence to calculate the average correlation of each node,and captures the deep-level static connection characteristics of the freeway network.Moreover,in order to describe the change trend of traffic flow and capture the dynamic change of connection strength,this thesis proposes a new traffic flow parameter,the fluctuation features of traffic flow.In addition to the characterization of spatial relationships,there is also the problem of insufficient utilization of spatiotemporal dependencies in prediction.Therefore,this thesis proposes a spatiotemporal graph network model driven by dynamic spatiotemporal features for traffic volume prediction task.Based on spatial connectivity matrix,using a graph attention network to capture the dynamic spatial dependencies between traffic flows,where the attention weights are calculated by using the fluctuation features of traffic flow.In addition,the temporal distribution characteristics are captured by a gated recurrent unit,and the recent level,day level,and week level temporal dependence are extracted respectively considering the different scale of temporal correlation,and the temporal attention mechanism is used to learn the multi-level correlation adaptively.By extracting the spatial and temporal distribution features,it can make full use of the temporal and spatial dependence of traffic flow to predict.This thesis conducted model performance evaluation experiment on the Jiangsu freeway data set.The evaluation metrics is used to compare the prediction performance of the proposed model and other benchmark models,and the ablation analysis experiment is carried out for each module.The results of comparative experiments demonstrate the effectiveness of the proposed dynamic spatiotemporal feature-driven traffic volume prediction model.Ablation analysis experiments proved the effectiveness of each module in the prediction model.