Reliability assessment of PEMFC aging prediction based on probabilistic Bayesian mixed recurrent neural networks

The current deep learning-based aging prediction models for Proton Exchange Membrane Fuel Cells (PEMFC) are inherently uninterpretable, focusing solely on prediction accuracy. However, the credibility of aging prediction results is one of the key factors limiting their practical application. This paper proposes a Bayesian Mixed Gated Unit (B-MIXGU) model, which integrates Bayesian theory with the Mixed Gated Unit model (MIXGU) to provide both point estimates and interval estimates of PEMFC aging predictions. First, the model parameters of MIXGU are replaced with probability distributions derived from Bayesian theory. Next, the total uncertainty is quantified using the variance of the interval estimates, where cognitive uncertainty and arbitrary uncertainty are characterized by the posterior distribution of the parameters and the probability distribution of the output, respectively. Durability test data under dynamic load cycle conditions show that, in cases where training data is limited or domain shifts exist between training and testing data, the prediction accuracy of B-MIXGU significantly surpasses other improved neural network models. Compared to MIXGU model with an attention mechanism (AT-MIXGU), RMSE and MAE are reduced by 44 % and 29 %, respectively. For the first time, the credibility of PEMFC aging predictions is evaluated from the perspective of uncertainty sources.