Research on Predicting Performance Degradation of Proton Membrane Fuel Cells Using Deep Learning

Background: Proton exchange membrane fuel cells (PEMFCs) face performance degradation due to complex mechanisms like catalyst sintering and membrane drying, limiting long-term reliability. Traditional models struggle with multivariate interactions, while data-driven approaches lack interpretability. Methods: We propose a hybrid deep learning framework using LSTM and Transformer models on a 570,000record dataset from an 800 -hour PEMFC test, incorporating sliding windows, MSE loss, regularization, and SHAP for interpretability. Results: The Transformer model achieves MAE of 0.0125, RMSE of 0.0202, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{R}^{\mathbf{2}}$</tex> of 0.971 in regression, and 95.6% accuracy in classification, outperforming LSTM, SVR, and linear regression. SHAP highlights temperature and voltage as key factors. Concluding Remarks: This framework provides accurate, interpretable predictions for PEMFC maintenance, demonstrating stability and real-time feasibility.