Graph-Based Deep Learning for Pedestrian Flow Prediction in Transport Hub: A Simulation-Informed Approach

Transport hubs play a major role in the national transport network. Larger and deeper subway stations attract a considerable influx of passengers, particularly during peak hours. Therefore, long-term pedestrian flow prediction is important for subway station management. This paper proposes an end-to-end learning framework based on the Graph Convolutional Network and the Attention Mechanism. In the first phase, key functional facilities within the subway station are abstracted into a graph structure. Subsequently, pedestrian flow data for graph nodes are obtained using Anylogic pedestrian simulation. In the second phase, a Graph Convolutional Network is employed to extract spatial features among nodes to capture the spatial dependencies of each node. The temporal attention operation is utilized to capture long-term temporal dependencies among nodes. This aids in the rational allocation of subway personnel and reducing congestion. Results demonstrate the validity and robustness of our model for pedestrian prediction.