Abstract
Efficient health indicators (HI) and prediction methods are crucial for assessing the remaining useful life (RUL) of fuel cells. However, obtaining HI under dynamic conditions with frequently changing loads is highly challenging. Therefore, this study proposes a prediction framework based on dynamic conditions. A method combining complete ensemble empirical mode decomposition with adaptive noise, power spectral density, and energy analysis (CPE) is proposed to extract HI under dynamic conditions from the perspectives of frequency and energy. Furthermore, the time convolution network with adaptive Bayesian optimization (AB-TCN) is introduced to address parameter optimization and prediction challenges. Effective feature parameters of the data are identified using random forest and used to train the AB-TCN. Results show that the extracted HI can effectively determine the end-of-life. The AB-TCN achieves accurate RUL estimation with a prediction error of only 6.825% and shows strong adaptability to various prediction tasks.