An interpretable uncertainty quantification framework for proton exchange membrane fuel cell degradation prediction

Accurate prediction of fuel cell degradation characteristics is crucial for its control and health diagnosis. However, existing mainstream data-driven methods often fail to adequately account for uncertainty factors during model training and the transient changes in fuel cell characteristics, resulting in predictions that typically provide only point estimates, lacking both credibility and interpretability. To address this, this paper proposes a framework that integrates multi-head attention mechanisms and Bayesian 1D convolution (Bayes-1DCNN-MSH), aimed at quantifying uncertainty in the predictions of various deep learning models while offering interpretable prediction information. This method uses multi-head attention to capture critical multi-dimensional features impacting fuel cell output voltage and applies Bayesian 1D convolution to quantify uncertainty, offering predictions with probability density distributions and confidence intervals for degradation forecasting. The Bayes-1DCNN-MSH structure, validated across six deep learning models, shows a 100 % improvement rate under diverse conditions, significantly enhancing noise resistance and reliability. Experimental results confirm that the framework effectively captures transient feature changes and explains prediction uncertainty through feature probability distributions. This approach improves the reliability and interpretability of fuel cell prediction models, supporting safer integration into sectors like transportation.