Accurate prediction of short-term passenger inflow and outflow at subway stations is important for the real-time operation and effective management of urban transit systems. Existing studies widely acknowledge that such prediction relies on capturing the temporal dependencies inherent in historical flow data and the spatial correlations between the target station and surrounding factors influencing passenger demand. Building on this principle, we propose a dynamic spatial-temporal decomposition and gate fusion network that explicitly accounts for subway-bus transfer behaviours and effectively captures the spatial-temporal dependencies. The main contributions of this paper are two-fold: (1) We develop a gated cross-modal attention fusion module to capture the interactions between subway and bus stations through a dual-stream architecture. Unlike existing approaches that focus on general multi-modal flow prediction, this module is specifically designed to model the unique characteristics of subway-bus transfer behaviour; and (2) In contrast to prevailing deep learning methods that adopt unified architectures to model spatial and temporal dependencies, we propose a novel dual-path spatial-temporal adaptive gate fusion module to better capture the heterogeneity between these dependencies across subway and bus modes. Furthermore, we introduce a temporal decomposition module and a dynamic compressed causal graph convolution module to enhance temporal sensitivity and overcome the structural limitations of static network topologies, respectively. Numerical experiments on real-world data from Beijing, China, demonstrate that our model consistently outperforms state-of-the-art methods. The datasets and code implementation are released on https://github.com/liushan-seu/DST-DFN.
Liu et al. (Mon,) studied this question.