Proton Exchange Membrane Fuel Cells (PEMFC) are a promising solution in the energy transition focused on reducing CO2 and utilizing renewable energies. As part of the development of these components, enhancing both performance and durability is crucial as they are expected to achieve longer lifespans. This research develops an optimization framework to optimize operational parameters within the constraints of physical laws, thus maximizing PEMFC performance in high-power applications. The study involves the analysis of over 500 hours of experimental data, leading to the creation of a surrogate model for the FC using Recurrent Neural Networks (RNN), which serves as a black-box function. These functions do not possess an analytical expression, lack derivatives and require computational-costly evaluations. Therefore, Bayesian Optimization (BO) is adopted as a solver for physics-constrained optimization, given its demonstrated effectiveness in optimizing black-box functions. The findings reveal that a Gated Recurrent Unit (GRU) surrogate shows high precision and reliability as a model, achieving an accuracy of over 98% in replicating the FC's internal state regarding physical parameters. When integrated with the optimization framework, it identifies optimal parameter combinations within the feasible region defined by the physics constraints, leading to an improvement in FC performance up to 5 %.
Bodon et al. (Mon,) studied this question.