Performance degradation prediction of proton exchange membrane fuel cells based on CNN-LSTM network with squeeze-and-excitation attention mechanism

Durability issues have become the main obstacle to the commercialization process of proton exchange membrane fuel cells (PEMFCs), and accurately predicting the performance degradation of PEMFCs under dynamic conditions is significant yet challenging. Herein, this paper proposes a novel prediction model that fuses long short-term memory (LSTM), convolutional neural network (CNN), and squeeze-and-excitation attention mechanism (SEAM), namely CNN-LSTM-SEAM prediction model. Specifically, the combination of LSTM and CNN can effectively avoid over prediction, while the addition of SEAM eliminates the requirement to select characteristic parameters related to voltage degradation, further improving prediction accuracy. Additionally, three other deep learning prediction models (LSTM, CNN, and CNN-LSTM) are compared with CNN-LSTM-SEAM prediction model on the static operating condition dataset, and the CNN-LSTM-SEAM prediction model is further validated in different ratios on the dynamic operating condition dataset. The prediction results demonstrate that the proposed CNN-LSTM-SEAM prediction model achieves superior prediction accuracies with high stability. For instance, the proposed CNN-LSTM-SEAM prediction model can accurately predict the voltage degradation details of PEMFCs, with the optimal root mean square error (RMSE), mean absolute error (MAE) of 7.440 × 10−4 V and 6.095 × 10−4 V (under the static operating condition) and 1.317 × 10−3 V and 8.389 × 10−4 V (under the dynamic operating condition), respectively.