Multi‐Resolution Deep Learning for Coupler Force Prediction in 20,000‐Ton Heavy‐Haul Trains

ABSTRACT Heavy‐haul trains play a crucial role in long‐distance bulk transportation, yet their enormous mass and kilometer‐scale length lead to complex longitudinal interactions and high coupler forces, which threaten operational safety. Conventional mechanism‐based models, while accurate, are computationally expensive and unsuitable for real‐time prediction. To address this limitation, this study develops a data‐driven prediction framework that combines physics‐based modelling and deep learning. A detailed longitudinal dynamics model of a 20,000‐ton train operating on the Shuohuang Railway is constructed, incorporating traction, electrical braking, and resistance characteristics to compute coupler forces under varying gradients and curvature conditions. Based on this model, a QP‐based optimization algorithm and a high‐fidelity simulation platform are used to generate multi‐strategy operating datasets that balance energy efficiency, punctuality, and ride comfort. The resulting data are processed using normalization and sliding‐window segmentation to form supervised learning samples. A multi‐resolution dual‐stream LSTM (MRDS‐LSTM) and its attention‐enhanced variant (MRDS‐LSTM–Attn) are then proposed to capture both short‐term fluctuations and long‐term temporal trends. Compared with RNN, GRU, LSTM, Bi‐LSTM, NLSTM, CNN‐LSTM, CNN‐NLSTM, CapNet‐NLSTM, Transformer, and Informer baselines, the proposed model achieves the highest prediction accuracy with MRDS‐LSTM‐Attn achieves an MAPE of 2.57%, and of 0.9888. The results demonstrate that the proposed framework effectively bridges physical modelling and data‐driven prediction, achieving up to 706 faster inference than traditional solvers. It provides a practical foundation for intelligent heavy‐haul train operation, supporting real‐time coupler force monitoring, predictive safety control, and future extensions to pneumatic braking and field data validation.