Traffic Congestion Signal Inference Based on Trajectory Correlation: A Trajectory Similarity-Based Graph Convolutional Network

Urban traffic flow often exhibits significant dynamic spatiotemporal correlations, leading to uncertain and complex traffic states. Accurately forecasting congestion propagation, which reflects the network dynamics of the congestion scenarios, is crucial for mitigating severe traffic jams and improving the overall traffic situation. Vehicle trajectory features at flow-level traffic provide new insights into spatiotemporal dependencies. To characterize and forecast the dynamic propagation rules among different roads, this work proposes a propagation inference framework based on a trajectory correlation-based graph convolutional network (TrajGCN) model and a propagation pattern module. Because single and fixed adjacency matrix always limit the expressiveness of GCN models in diverse scenarios, the proposed dynamic trajectory module integrates multiple trajectory features, including speed, demand overlap rate, distance, and connectivity, to describe vehicle mobility interactions. Then, a spatial-temporal block is developed to enhance both dynamic and static GCN modules in capturing spatiotemporal characteristics. Furthermore, urban congestion propagation patterns are simulated by learning potential propagation rules. Experimental results of a regional road network in Beijing show that the TrajGCN model outperforms baseline methods in most cases. Rather than relying on predefined features such as distance or connectivity, this study discusses the propagation patterns of congestion signals within traffic flow-level network structure. The proposed method can improve the understanding of propagation patterns, facilitating traffic management strategies such as signal control and lane allocation to mitigate congestion generation and signal propagation growth.