Research on Modeling Method of Metro Passenger Flow Congestion Propagation with Dynamic Spatiotemporal Graph Convolutional Networks

With the rapid development of metro network operation, metro passenger flow congestion propagation occurs frequently. Accurately modeling passenger flow congestion propagation is crucial for alleviating metro passenger flow congestion and formulating corresponding control strategies. Traditional modeling methods struggle to effectively capture the complex spatiotemporal dependency relationships in metro networks. To improve the accuracy of congestion propagation modeling, this paper proposes a Dynamic Spatiotemporal Graph Convolutional Network (DSTGCN). The model integrates node attributes and temporal encoding through a dynamic adjacency matrix generation module, uses multi-head attention mechanisms to adaptively learn the time-varying propagation intensity between nodes, and combines static topology to construct dynamic adjacency matrices. A multi-scale spatiotemporal feature extraction module is designed, employing temporal convolution and spatial attention mechanisms to mine periodic and local correlation features, and aggregating historical states with different time lags through stacked graph convolutions. Experimental results on real metro datasets verify the effectiveness of each module of the model and reveal the inherent laws of passenger flow congestion propagation in metro networks. The research provides theoretical support for congestion early warning and dynamic regulation in metro network operation.