Physics-informed neural network for long-term prognostics of proton exchange membrane fuel cells

Abstract

This study formulated a physics-informed neural network (PINN) to prognosticate the remaining useful life (RUL) of proton exchange membrane fuel cell (PEMFC), leveraging both prior knowledge of PEMFCs and aging test data. Governing equations elucidating membrane and catalyst degradation mechanisms were integrated into the PINN framework to assimilate the prior understanding of PEMFC degradation dynamics. Subsequently, operation time and current density were designated as input variables for the PINN to forecast output voltage, following a comprehensive analysis of PEMFC operational mechanisms and aging test data. Notably, a novel strategy involving replicating the current density from the training phase to the prediction phase was introduced, incorporating prior knowledge of PEMFC system variability into the PINN. Consequently, the proposed PINN demonstrated proficient prediction of PEMFC RUL while mitigating reliance on aging test data. This accomplishment, representing a 9.2 percentage points enhancement over the previously lowest reported data dependency of 35.2 %, substantiates the attainment of state-of-the-art status by the proposed PINN in reducing data dependency.