Knowledge-guided Intelligent decoupling control for PEMFC air supply based on end-to-end dynamic modeling

The air supply system of fuel cell vehicles exhibits strong multivariable coupling, time delays, and nonlinear dynamic behavior, which pose significant challenges for precise and robust control. To address these issues, this study develops an end-to-end intelligent air management framework based on a 100 kW-class proton exchange membrane fuel cell (PEMFC) system, integrating a Long Short-Term Memory (LSTM)–based surrogate model and a knowledge-guided Deep Deterministic Policy Gradient (DDPG) controller. The LSTM model captures system dynamics from raw sensor signals with 99.7% prediction accuracy for air mass flow and pressure, providing a reliable interaction platform for intelligent control. A composite knowledge-guided reward mechanism is designed to incorporate actuator constraints, exploration efficiency, and control precision, enabling efficient autonomous learning and intelligent decoupling of multivariable interactions. Ablation experiments confirm that the knowledge-guided mechanism improves convergence speed by a factor of two. Comparative simulation results show that the proposed DDPG controller converges 17% faster than TD3, achieves a 1.5-fold improvement in response speed under load disturbances, and maintains steady-state errors within 2 g/s and 1.5 kPa. The controller also demonstrates strong generalization capability under an unseen WLTP driving cycle. Additional decoupling scenarios further verify effective decoupling performance, with response times below 1 s. Hardware-in-the-loop experiments demonstrate that the proposed knowledge-guided DDPG controller satisfies real-time execution requirements and maintains stable control performance under sensor noise disturbances. These results highlight the effectiveness, robustness, and practical applicability of the proposed framework, offering new insights into intelligent air supply management for fuel cell vehicles.