Degradation prediction and uncertainty quantification for PEMFC using NSGA-optimized deep learning method

Abstract

Reliable degradation prediction of the proton exchange membrane fuel cell (PEMFC) can provide sufficient decision support for its predictive maintenance. However, most prediction methods only focus on the point prediction and do not consider the uncertainty in the degradation prediction. In this paper, an optimized deep learning method with uncertainty quantification is initially proposed for PEMFC degradation prediction. Specifically, the deep learning method based on deep belief network (DBN) and extreme learning machine (ELM) is combined with the lower upper bound estimation method to construct a novel prediction model. It can quantify the PEMFC uncertainty through the established prediction interval (PI). On this basis, the non-dominated sorting genetic algorithm (NSGA) and a novel comprehensive cost function are used to optimize the DBN-ELM performance further. Finally, the performance of the NSGA- DBN-ELM method is verified by measured data, and the experimental results show it can provide not only accurate prediction results but also a highly reliable PI for PEMFC health management.