The Membrane Water Distribution Estimation of Fuel Cell based on Physics-informed Neutral Network

The proton exchange membrane fuel cell (PEMFC) is becoming an indispensable component of the future green energy infrastructure. However, accurately estimating its internal states, particularly for high-dimensional distributions, presents a significant challenge for advancing health status management strategies. To address this, a physics-informed neutral network (PINN) is proposed to estimate the two-dimensional membrane water distribution within PEMFCs. The method integrates multiple loss functions, including data loss, partial differential equation (PDE) loss, initial condition loss, and boundary condition loss. The PDE loss explicitly incorporates the heterogeneity and time-varying parameters of the membrane water diffusion mechanism, thereby enhancing both the accuracy and interpretability of the model. Additionally, Bayesian optimization is employed to fine-tune the weights of the four loss components, effectively improving convergence speed. The comparative simulation between proposed method and PINN with fixed parameters PDE is executed, which demonstrates a 16.20% decrease in the mean absolute percentage error, from 0.3491 to 0.3004.