PEMFC Degradation Prediction and Probabilistic State Assessment Using a Multiscale Periodic TimesNet With Laplace Loss

Accurate prediction of voltage degradation in proton exchange membrane fuel cells is crucial for ensuring their reliability and enabling predictive management. While existing methods predominantly focus on point prediction of voltage, the quantification of predictive uncertainty has not been adequately addressed. To tackle this issue, a Lp-TimesNet model is proposed as a novel approach for probabilistic prediction of voltage degradation. This model integrates the TimesNet module to capture multi-period temporal dependencies and incorporates a Laplace distribution-based output layer to quantify predictive uncertainty. Extensive evaluations under dynamic conditions demonstrate that the proposed model achieves a 19.4% reduction in root mean square error compared to the baseline TimesNet model, along with sharper and more reliable confidence intervals than other probabilistic models. Under both static and quasi-static conditions, the model also exhibits improved predictive accuracy and better uncertainty calibration. These results indicate that the proposed model possesses strong generalization capabilities, yielding accurate voltage predictions and reliable interval estimates across diverse operating conditions.