A deep learning and station feature fusion-based architecture for subway passenger flow prediction in intelligent transportation systems

Accurate short-term network-level passenger flow prediction is crucial for the development and operation of intelligent transportation systems, particularly urban subway systems. While graph neural networks (GNNs) have advanced this field, challenges remain in effectively modeling dynamic station characteristics and multiscale spatiotemporal dependencies. This study proposes a multi-graph convolutional-convolutional network (SF-MGCN-CNN) integrating comprehensive station features to address these limitations. The framework simultaneously captures temporal periodicity and trends, spatial connectivity, passenger flow pattern similarity, and transfer correlations through dedicated branch networks. Furthermore, it employs multi-class feature segmentation modules to better represent station heterogeneity based on passenger flow, land-use properties, and network attributes. Validation on the Chengdu subway network demonstrates superior prediction accuracy compared with baseline methods, while ablation studies confirm the individual contributions of each component. The proposed framework provides methodological support for short-term passenger flow prediction in urban subway operations by incorporating station heterogeneity and multiple spatial correlations.