Integrated modeling, identification, and neural network control of PEM fuel cells

Abstract Proton exchange membrane fuel cells (PEMFCs) are promising candidates for clean energy conversion in transportation, stationary, and distributed power applications. However, their nonlinear dynamics, coupled with water management and gas flow constraints, present significant challenges for real-time simulation and control. This work develops an integrated framework that combines physics-based modeling, data-driven identification, and neural network–based control for PEMFC systems. A dynamic five-state mass balance model is formulated to describe hydrogen, oxygen, and water mole inventories together with inlet valve dynamics. Synthetic datasets generated from the nonlinear plant are then used to train a long short-term memory (LSTM) network, enabling sequence-to-sequence prediction of PEMFC states. The LSTM surrogate achieves high fidelity with mean root mean square error (RMSE) below 1% on validation data, while reducing computation time by more than an order of magnitude compared to direct physics-based simulations. For control, an expert proportional–integral (PI) strategy is designed to regulate hydrogen and oxygen supply, and its operation is used to train a feedforward neural network (NN) controller via supervised learning. Quantitative evaluation shows that the NN closely reproduces the expert controller policy with RMSE of 0.52 mol/s for hydrogen and 0.47 mol/s for oxygen. Closed-loop validation demonstrates superior performance of the NN controller: under step input tracking, it achieves faster settling time (1.76 s vs. 2.64 s for PI) with negligible overshoot and zero steady-state error, while under disturbance rejection it reduces overshoot (2.03% vs. 3.71%) and halves settling time (0.30 s vs. 0.57 s). These results confirm that the NN controller’s faster response, reduced overshoot, and improved robustness, highlighting its suitability for real-time PEMFC energy management and automotive deployment.