A Trajectory-Driven Multi-Layer Spatiotemporal Graph Neural Network for Predicting Short-Term Urban Traffic State
The short-term prediction of urban traffic states serves as a fundamental application supporting traffic management and online navigation.Floating car trajectory data,characterized by low cost and high spatio-temporal coverage,has been widely employed for real-time traffic state extraction,thereby facilitating short-term traffic prediction.However,the spatio-temporal coverage of floating car trajectories is extremely uneven,resulting in a significant amount of missing or insufficient coverage of road networks over many time periods.It is difficult to accurately estimate the traffic states of all segments of the entire road network directly using the trajectory data of floating cars.The accuracy and reliability cannot meet the needs of real-time traffic state estimation and short-term prediction.This issue therefore poses a significant technical challenge for online short-term prediction of traffic states based on uneven trajectory data across the entire road network in large cities,hindering the refinement of traffic monitoring and management.To address the issue of uneven spatiotemporal distribution in trajectory data,this study proposes a dynamic layering method for the road network,dividing it into multiple layers based on the spatio-temporal distribution of trajectories,including a main road network with high-quality trajectories and other secondary road networks with sparse trajectory distributions.Building upon the layered road network,we propose a trajectory-driven multi-level spatio-temporal graph neural network method for short-term traffic state prediction.Leveraging the hierarchical structure of road networks,the proposed prediction method incorporates intra-and inter-layer message passing mechanisms that consider the spatio-temporal distribution of trajectories.We apply dilated causal convolution and graph attentional mechanisms to describe the complex spatio-temporal correlations of traffic states among road networks.Based on the correlation representation,we develop a unified graph neural network prediction model that integrates "representation" and "prediction" into an end-to-end learning scheme.The proposed method enables online prediction of speed and congestion state of all road segment in the road network simultaneously,significantly enhancing the predictive performance of traffic state for road segments with sparse trajectories.Through testing on real trajectory data from the large road network of Wuhan City,the proposed method showcases a notable improvement in prediction accuracy compared to recent popular baseline methods,particularly achieving satisfactory performance in segments with severe trajectory data deficiencies.The training efficiency is also significantly improved.Experimental results indicate that the proposed multi-level spatio-temporal graph neural network prediction method can effectively handle the prediction challenges caused by the uneven distribution of trajectories.
NETL
NSTL
万方数据
