MD-TLN: A Transformer--LSTM Hybrid Network for Metro Passenger Flow Prediction With Multi-Source Features

Abstract

Accurate metro passenger flow prediction is essential for enhancing daily crowd management and vehicle scheduling, especially amid the rapidly increasing demand for metro travel. However, existing mathematical and deep learning models struggle to capture the complex effects of heterogeneous external factors, resulting in limited modeling capacity and poor generalization in dynamic environments. To address these challenges, this paper proposes a hybrid deep learning framework, termed Multiple Dimension Transformer LSTM Network (MD-TLN), which integrates multi-source heterogeneous features within a unified structure. MD-TLN comprises three channels—historical, spatial, external and—organized into a structured five-dimensional tensor to balance expressiveness and computational efficiency. The model combines a convolutional feature encoder, a multi-patch transformer for temporal dependency modeling, and an attention-based fusion mechanism for dynamic feature weighting. Experiments on data from 263 Guangzhou Metro stations (October 1–31, 2024) show that MD-TLN significantly outperforms state-of-the-art baselines in MAE, MSE, and R². Extensive performance evaluation and ablation studies further validate the accuracy and robustness of the proposed framework, demonstrating its strong potential for real-world metro operations under complex disturbances. The implementation of MD-TLN is available at: https://github.com/Lilin-Chen/MD-TLN