Abstract
Abstract Due to its high efficiency and free from environmental pollution Proton exchange membrane fuel cells (PEMFCs) attracted a lot of attention for clean energy conversion in rural, hospital and vehicle power applications. But the nonlinear characteristics along with the difficulties associated in water management and gas flows make it hard to develop a real-time control strategy. The research develops an integrated model that combines physics modeling, data driven parameter estimation, and control using deep neural networks for PEMFC systems. In this study, an integrated modeling approach using physics modeling, parameter estimation using data-driven techniques, and control using Deep Neural Networks (DNNs) for PEMFC systems has been developed. The main contribution of this study is to integrate a physics-based five-state PEMFC model, an LSTM surrogate model for dynamic predictions, and a DNN Controller, trained from expert PI control policies in a common framework for efficient PEMFC modeling and control. A five-state dynamic model is formulated for simulating the behavior of oxygen, hydrogen, and water vapor inside the PEMFC along with inlet-valve dynamics. Simulation results are used to train an LSTM network to learn and predict the dynamic responses of the system. The trained LSTM model achieved an RMSE of 1.7 × 10⁻⁵ and provided faster simulations than the physics-based model by more than one order of magnitude. An expert PI controller was employed to train a feedforward NN controller which successfully learned the expert’s control strategy with an RMSE of 1.52 × 10− 4 mol/s for hydrogen and 9.85 × 10⁻⁵mol/s for oxygen. Closed-loop validation of the proposed modeling and control approach reveals superior performance of the proposed DNN Controller compared to the traditional PI controllers. Specifically, the DNN Controller showed reduced settling times for tracking (1.76 s vs. 2.64 s) and disturbance rejection (0.30 s vs. 0.57 s) while decreasing overshoot from 3.71% to 2.03% and maintaining zero steady state error. These results demonstrate the effectiveness of the proposed framework for real-time PEMFC modeling, control, and energy-management applications.