Multi-feature temporal sequence modeling for improved subway passenger flow prediction

Abstract Station-level subway passenger-flow forecasting must account for daily regularity, weather and day-type changes, and spatial coupling among stations. This paper develops TCSET-seq2seq, a multi-feature sequence-to-sequence model for short-term subway passenger-flow prediction. For each target day, the model retrieves similar days from earlier dates, combines rule-based calendar, temporal-distance, and weather matching with a neural similarity score, encodes selected passenger-flow sequences with condition-aware temporal convolutions, recalibrates similar-day and channel representations with a squeeze-and-excitation module, propagates features over the undirected station topology, and generates daily forecasts autoregressively. Experiments use 30-minute smart-card records from Hangzhou, Shanghai, and Zhengzhou metro systems within 06:30–22:00, with repeated day-type/weather-stratified splits, historical-only similar-day retrieval, and inverse-scale MAE, RMSE, and MAPE. With posthoc nonnegative clipping during inference, the complete TCSET configuration obtains RMSE values of 283.37, 355.37, and 271.46 on Hangzhou, Shanghai, and Zhengzhou, respectively. It gives the lowest RMSE on Hangzhou and Zhengzhou and is slightly higher than the strongest Shanghai baseline. Component ablation and hyper-parameter sensitivity analyses indicate that SET recalibration, topology propagation, and output clipping are most useful on the longer Zhengzhou dataset, whereas external features and hybrid similarity are less stable on the shorter Hangzhou and Shanghai datasets. The results suggest that TCSET-seq2seq is effective for station-level forecasting on the tested datasets, but the similar-day retrieval and external-feature settings should be adjusted to the available data.