Uncertainty-aware oriented lifetime prediction of proton exchange membrane fuel cells based on high-order time-frequency health indicator

Abstract

Accurate degradation prediction of proton exchange membrane fuel cells is essential for their reliability and durability. However, the sophisticated degradation mechanism introduces uncertainties that compromise the prediction accuracy of PEMFCs lifetime. To address this problem, an uncertainty-aware network is proposed for interval prediction of degradation, which leverages higher-order time-frequency health indicators. These indicators are derived from higher-order voltage polynomials, with coefficients determined by frequency features extracted from the distribution relaxation time. This approach facilitates the extraction of multi-order effective information. The uncertainty-aware network achieves interval prediction by incorporating global quantile regression layer into bidirectional long short-term memory neural network, which increases prediction accuracy and reliability. Moreover, the nature-inspired hippopotamus optimization algorithm is employed to fine-tune hyperparameters of uncertainty-aware network, reducing computational complexity. The performance of proposed method is demonstrated through experimental comparisons. The root-mean-square error of prediction was improved by more than 39.65% for both static and dynamic conditions, and the accuracy of remaining life prediction was improved by more than 32.8%. This method provides a high-order interpretable time-frequency health indicator for fuel cell degradation prediction, which provides strong support for fuel cell degradation prediction and long-time stable operation.