Uncertainty quantification-based framework for predicting degradation trends of proton exchange membrane fuel cell

Accurately predicting the degradation trends of proton exchange membrane fuel cells (PEMFCs) can provide a solid basis for optimizing the control of vehicles and stations based on PEMFCs. However, most prediction methods do not consider factors such as measurement errors from experimental environments and the inherent cognitive uncertainty of the model. These methods can only offer point estimates, lacking credibility. This paper introduces a deep learning prediction framework that combines a bidirectional gated recurrent unit (BiGRU) model with a truncated Bayes by backpropagation through time (TB) algorithm. The TB algorithm reconstructs fixed parameters in the model into probability density distributions, transforming the output from point estimation to interval estimation with probability density distributions. Under dynamic conditions, the TB-BiGRU (truncated Bayes-based bidirectional gated recurrent unit) improves the mean absolute error (MAE) and root mean square error (RMSE) by 37.28% and 36.09%, respectively, compared to the TB-GRU (truncated Bayes-based gated recurrent unit). Compared with TB-GRU and B-GRU (Bayesian gated recurrent unit), TB-BiGRU has significantly improved uncertainty quantification ability. Under different working conditions and noise levels, the prediction accuracy of TB-BiGRU is superior to that of the other seven models, and it exhibits better noise resistance and stability. This method holds greater practical significance compared to other prediction approaches. Additionally, the paper proposes four effective evaluation metrics for uncertainty quantification, providing higher reference value in effectively characterizing the model’s prediction accuracy and uncertainty quantification capability.