A framework based on temporal causal inference graph neural networks for the probabilistic estimation of the remaining useful life of proton exchange membrane fuel cells

Proton exchange membrane fuel cells (PEMFCs) offer high efficiency and clean emissions for sustainable energy systems, yet accurate prediction of their Remaining Useful Life (RUL) remains challenging due to complex degradation mechanisms under dynamic operating conditions. ​This study introduces a temporal-causal inference framework that synergizes Graph Total Variation (GTV) regularization with Monte Carlo (MC) dropout to resolve the efficiency-accuracy trade-off in PEMFC prognostics.​​ The approach dynamically infers degradation causality via Peter-Clark momentary conditional independence algorithms to construct time-lagged interaction graphs, while the GTV convolutional layer suppresses spurious correlations through spectral-domain regularization. By integrating MC dropout without Bayesian computational overhead, the framework enables computationally efficient probabilistic RUL estimation. Validation on the PHM 2014 dataset demonstrates state-of-the-art performance, achieving a 0.738 α-Coverage at 95% confidence intervals for probabilistic reliability, reducing RMSE compared to graph convolution baselines for accuracy superiority, and enhancing robustness verified in ablation studies, collectively advancing reliable fuel cell deployment in practice.