A Multitask Learning Framework With LSTM-TPA for Dynamic Modeling of Automotive Fuel Cell Systems

Proton exchange membrane fuel cell (PEMFC) systems exhibit complex dynamic nonlinearities and multi-timescale coupling effects, challenging traditional physics-based modeling. To improve modeling accuracy and adaptability, this paper proposes a multi-task deep learning modeling framework based on the temporal pattern attention mechanism and long short-term memory networks (Multi-LSTM-TPA), providing an end-to-end dynamic modeling solution for PEMFC systems. Through a feature-response grouping method based on subsystem control logic and electrochemical coupling mechanism, the model internally implements independent training and updates of multiple subtasks. The TPA mechanism dynamically optimizes the weights of the LSTM across different time steps, effectively capturing the dynamic characteristics at various timescales. The 70-30 train-test split demonstrated an optimal trade-off between training cost and prediction accuracy, achieving 95.2% regression accuracy on the test set while showing high robustness against ratio variations. A PCA-based projection compression method was developed, which reduced memory usage by 88% while maintaining comparable accuracy, thereby enhancing computational efficiency and reducing deployment complexity. Finally, the model’s high-fidelity representation and practical utility were validated through Simulink simulations.