PEMFC model based on recurrent neural network considering hysteresis effect on polarization for energy management strategies applications

This paper focuses on the design of an EMS through a comprehensive methodology encompassing system modelling, controller design, and experimental validation on a fuel-cell (FC)/battery hybrid test bench. The modelling methodology is data-driven and constructed from measurements obtained on our test bench. The primary objective is to reproduce the dynamic behaviour of the FC, with particular emphasis on its polarization characteristics. The proposed modelling approach relies on recurrent neural networks, which are well suited for capturing the temporal inherent dependencies in FC behaviour, thereby allowing accurate prediction of the voltage response for a given current input. When integrated into the overall vehicle model, this data-driven representation provides the foundation for training a reinforcement-learning (RL) based EMS with the state of the art Soft Actor-Critic algorithm. The modelling achieves a root mean squared error of 0.30 V and successfully captured the hysteresis effect of the fuel cell. Combined with an appropriately designed reward function, the EMS effectively addresses the battery charge-sustaining requirement, thereby allowing the optimization target to focus on improving FC efficiency. This combined approach results in stable and high-efficiency FC operation and enables a reduction in hydrogen consumption of 3.5% under experimental conditions. Additionally, the proposed approach consistently drives the system towards the upper region of the fuel-cell efficiency curve where hysteresis effects are reduced, leading to lower consumption and potentially reduced degradation, as supported by the observed system behaviour.