An interpretable data-driven method for degradation prediction of proton exchange membrane fuel cells based on temporal fusion transformer and covariates

Abstract

The durability of proton exchange membrane (PEM) fuel cells is an obstacle to industrialization. PEM fuel cells need a robust incentive to improve their durability. Traditional data-driven methods combined with deep learning for degradation prediction methods lack interpretability, which prevents users from fully trusting the results and limits the application. Moreover, traditional data-driven methods not making full use of the valuable information in the experimental data. This study presents an easy to interpret data-driven method that improves the reliability of the model, the heterogeneity of input features, and the effects of significant covariates. The method's primary objective is to provide a long- and short-term degradation prediction of PEM fuel cells for a temporal fusion transformer. This method achieves high accuracy in predicting the degradation performance of PEM fuel cells with a root mean square error of 0.0067. The model's root means square error is reduced by 58% compared to the attention-based gated recurrent unit model. Based on the interpretability analysis, the degradation performance directly correlates with power, the cathode inlet temperature, the anode inlet temperature, time, the cathode inlet pressure, and the cathode outlet pressure. The method can also be applied to fuel cell systems for optimizing control, management, and maintenance.