Degradation prediction based on physics-constrained data-driven framework for hydrogen fuel cell lifetime

Accurate lifetime prediction has become a critical research focus in the field of hydrogen fuel cell technology advancement towards commercialization. Data-driven methods represent a mainstream approach for lifetime forecasting, their performance generally relies on extensive historical data. To reduce dependence on large datasets and enhance model generalization, this study proposes a hybrid prediction method that integrates physical mechanisms with data-driven methods. The proposed approach incorporates a fuel cell voltage degradation mechanism model as a physical constraint within the data-driven framework. The hybrid model consists of three core components. Firstly, a gradient boosting decision tree (GBDT)-based data-driven module that uses operating time and voltage as inputs to generate initial voltage predictions. Secondly, a voltage degradation mechanism model has been established based on the kinetics equations of the membrane and catalytic materials, providing theoretical voltage estimates as physical constraints. Finally, a bidirectional long short-term memory (Bi-LSTM) network for error correction, which uses prediction deviations from the first two stages as inputs to effectively minimize systematic errors through compensatory adjustments. The hybrid method has been validated under operating conditions that were static, quasi-dynamic, and dynamic with varying training durations. The results indicate that its accuracy is consistently superior to predictions generated directly by the GBDT algorithm. In data-scarce scenarios characterized by steady-state operations with only 300 h of training, the proposed method achieved reductions of 56.2 % in the root mean square error (RMSE) and 64.4 % in mean absolute percentage error (MAPE), thereby validating its superior generalization capability and data efficiency. Among all the evaluated operating conditions, the optimal RMSE and MAPE values were achieved under the dynamic working condition with a 200-h training duration, reaching 0.0047 and 0.09 %, respectively.