A method for quantifying fuel cell degradation power loss mechanisms based on an identifiability-driven physics-informed neural network

The output power of proton exchange membrane fuel cells (PEMFC) decreases after long-term operation. Existing diagnostic methods require shutdown testing or dedicated equipment, which severely disrupts production operations; consequently, the entire stack is often replaced directly in engineering practice, resulting in substantial economic waste. To address these issues, this paper proposes an identifiability-driven physics-informed neural network (ID-PINN) method. Specifically, a multi-mechanism coupled electrochemical model covering five degradation pathways is embedded as a differentiable physics-constrained layer. A dual-branch architecture comprising a monotonic baseline and operating-condition residuals separates irreversible aging from operating-condition disturbances, identifies the evolutionary trajectories of 5 degradation-sensitive parameters using only routine operational data, and decomposes mechanism-level power losses under a unified rated operating condition. Validation using 994.88 h of long-term test data from an in-house stack and a public dataset from Tongji University shows that the test-segment voltage-prediction MAPEs of the two datasets are 1.89% and 1.39%, respectively. The ohmic, activation, and mass-transfer degradation trends reflected by 83 measured impedance spectra are consistent with the model identification results. The cumulative stack power loss is approximately 282.84 W, of which the model attributes 92.2% to catalyst activity decay, providing priority directions for stack inspection and maintenance.