Spatio-Temporal Attention Convolution Network for Short-Term Subway Passenger Flow Forecasting

With the rapid growth of urban rail systems, accurate passenger flow forecasting is crucial for optimizing capacity, reducing congestion, and improving service quality. To overcome the limitations of methods relying on static topologies, we propose the Spatio-Temporal Attention Convolution Network (STACN), which requires no predefined graph structure. STACN combines the global modeling strength of attention with the local feature extraction of convolution, using temporal and spatial attention convolution blocks to capture dynamic inflow and outflow dependencies. A bidirectional interaction strategy and joint spatiotemporal extraction further enhance the modeling of flow dynamics. Finally, a two-layer fully connected network outputs the predicted inflows and outflows. Experiments on metro data from Hangzhou and Shanghai show that STACN outperforms traditional and state-of-the-art models in accuracy, efficiency, and stability.