Short-term station-level forecasting is difficult because passenger demand is shaped by several interacting forces. Daily routines create strong regularity, but weather shifts and day-type changes—such as workdays versus holidays—alter ridership patterns, and flows at one station are coupled to flows at others through the network. TCSET-seq2seq addresses these forces in stages. For each target day, the model first retrieves similar days from earlier dates, combining rule-based matching on calendar attributes, temporal distance, and weather with a learned neural similarity score. The passenger-flow sequences of those selected days are then encoded with condition-aware temporal convolutions, which let the network condition its representation on the external features of the target day.

A squeeze-and-excitation module then recalibrates both the similar-day representations and the channel-wise features, effectively reweighting which signals matter most before prediction. The model propagates features across an undirected graph of the station topology, allowing information about crowding at neighboring or connected stations to inform forecasts at each stop. Finally, the daily forecast is generated autoregressively, with each prediction step feeding into the next. The authors trained and tested the model using repeated day-type and weather-stratified data splits, and restricted similar-day retrieval to historical data only, a design choice intended to prevent information leakage from the future into the training process.

Performance was measured with mean absolute error, root-mean-square error, and mean absolute percentage error using an inverse-scale weighting scheme. The complete TCSET configuration recorded the lowest RMSE among the tested methods on Hangzhou and Zhengzhou, while on Shanghai it came in slightly above the strongest baseline. Component ablation experiments and hyper-parameter sensitivity analyses showed that the squeeze-and-excitation recalibration, the topology propagation over the station graph, and the output clipping step were most useful on the longer Zhengzhou dataset. By contrast, the external features and the hybrid similarity mechanism proved less stable on the shorter Hangzhou and Shanghai datasets, suggesting that their benefit depends on having enough data to learn from.

The authors conclude that TCSET-seq2seq is effective for station-level forecasting on the datasets tested, but they caution that the similar-day retrieval and external-feature settings should be adjusted to the amount and character of data available in a given metro system. For operators, the practical implication is that graph-based spatial modeling and recalibration modules can carry substantial weight in forecast accuracy, particularly where long historical records exist, while feature-rich retrieval strategies may need careful tuning on smaller networks. The study was released as an open-access article on 26 October 2026, with the published version noted as citable through a permanent digital object identifier and subject to further editorial refinement before the final version of record replaces it.

Subject of Research: Multi-feature temporal sequence modeling for station-level subway passenger flow prediction

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

Article References: Hu, J., Zhong, R., Hu, S., Zeng, Z., Gan, W., & Tan, H. (2026). Multi-feature temporal sequence modeling for improved subway passenger flow prediction. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02505-y

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02505-y

Keywords: subway passenger flow, sequence-to-sequence model, temporal convolution, squeeze-and-excitation, graph propagation, smart card data, Hangzhou metro, Shanghai metro, Zhengzhou metro, short-term forecasting, Multi-feature, temporal