Hybrid time series–Exponential feature model for PEMFC degradation prediction

Proton exchange membrane fuel cells (PEMFCs) are becoming key to large-scale renewable energy integration. Accurate degradation prediction is vital for extending their lifespan. Model-based methods struggle to dynamically update parameters from historical data, while data-driven approaches find it hard to integrate physical prior knowledge. To address this challenge, this paper proposes a hybrid degradation prediction framework (LSTM-Transformer-ExpPR) for PEMFCs, enhancing accuracy by alternately integrating time-series and exponential degradation features. Firstly, based on the polarization resistance change patterns revealed through dynamic condition aging experiments, an exponential semi-empirical degradation model (ExpPR) was established. Secondly, a LSTM-Transformer model was constructed to learn time-series features from historical operational data and predict voltage observations. Finally, leveraging a particle filtering algorithm, the ExpPR model dynamically updates its degradation parameters using these observations and outputs the final voltage prediction. Results demonstrate that under dynamic conditions, the hybrid method achieves a prediction Root Mean Square Error (RMSE) of 5.5437. When the dynamic conditions are mapped to equivalent static conditions for prediction, the RMSE significantly decreases to 0.2332. Under static and quasi-static operating conditions, the RMSE values are as low as 0.0165 and 0.0227, respectively. The results confirm the method's outstanding accuracy across scenarios.