STD-former: spatial-temporal dynamic transformer-based model for urban rail transit passenger flow prediction

Accurate passenger flow forecasting for urban rail transit is crucial for network planning, smart mobility, and operations. However, existing methods still fall short in modeling heterogeneous time series and dynamic graphs. To this end, we propose a passenger flow forecasting method based on the Spatio-Temporal Dynamic Transformer (STD-Former). In the temporal dimension, STD-Former explicitly separates multi-scale temporal patterns using the discrete wavelet transform through its time-decoupling layer. In the spatial dimension, we introduce a dynamic graph clustering module and a temporal graph convolution module to enable dynamic station grouping based on spatiotemporal similarity and to model time-varying graph structures. Finally, the model achieves precise multi-step short-term predictions by processing high-dimensional spatio-temporal features through its encoder-decoder. Experiments on two real-world datasets demonstrate that STD-Former outperforms existing baselines, with visualisations and ablation studies confirming its effectiveness and the necessity of each component.