A hybrid degradation prediction method for PEMFC integrating model-based degradation index extraction and Bayesian-optimized Bi-directional long short-term memory

• A generalized hybrid degradation prediction framework for PEMFC is proposed • DRT and polarization curve model are integrated for degradation characterization. • The weight coefficients of the hybrid degradation index are quantified. • Bayesian optimization algorithm is employed to obtain the optimal hyperparameters. Accurately estimating the state of health of proton exchange membrane fuel cell (PEMFC) and predicting the degradation trend are essential prerequisites for effective health management to enhance durability. This paper proposes a generalized hybrid degradation prediction method for PEMFC that is applicable to diverse operating conditions. Firstly, the internal polarization dynamics are characterized via the distribution of relaxation times method, and a third-order equivalent circuit model is established to quantify the polarization losses. The voltage losses are quantified using a polarization curve model. Degradation characteristic analysis from both approaches consistently reveals that deterioration in mass transfer kinetics and charge transfer kinetics is the primary cause of performance degradation. Subsequently, component-level degradation indexes are extracted based on degradation models, and a novel weighted fusion method is proposed to construct a hybrid degradation index characterizing the overall degradation state of PEMFC. Finally, the Bayesian-optimized Bi-directional long short-term memory (Bi-LSTM) model is employed to predict PEMFC degradation trend under various prediction horizons, enabling accurate estimation of remaining useful life (RUL). The results show that the optimized Bi-LSTM achieves higher RUL estimation accuracy than the baseline Bi-LSTM, and the hybrid method outperforms the AutoML-based method and the cascaded echo state network reported in previous studies. For the first stack, the estimation error remains below 7.78%, with a minimum error of 0.50%. For the second stack, the estimation error does not exceed 12.28% overall and drops below 10% when the prediction horizon is within 300 h, with a minimum error of 2.67%.