Interpretable degradation forecasting of fuel cells under steady, quasi-dynamic and dynamic conditions: An ALA optimized TimesNet model based on ICEEMDAN decomposition

To improve the degradation prediction precision for fuel cells (FCs) under different operating conditions, a hybrid method was proposed by combining data decomposition with the artificial lemming algorithm (ALA) optimized TimesNet model (ICEEMDAN-ALA-TimesNet). The ICEEMDAN-ALA-TimesNet model exhibits substantially improved predictive performance over 12 benchmark methods across steady-state, quasi-dynamic, and dynamic conditions. Point prediction results for the FC1, FC2, and FC3 datasets indicate that, using only 30% of the training data, the ICEEMDAN-ALA-TimesNet model reaches MAE, MAPE, and RMSE of 0.017%–0.0966%, 0.0184–0.0298%, and 0.025%–0.1265%, respectively, with R2 ranging from 0.9951 to 0.9983. Even under a 60% missing data ratio, the model maintains strong robustness, with MAE, MAPE, and RMSE remaining within 0.0166%–0.1158%, 0.018%–0.0359%, and 0.0243%–0.1572%, respectively, and R2 values preserved between 0.9818 and 0.9987, highlighting its stability and generalization capability. A 95% confidence interval-based interval forecasting framework is further constructed to assess the reliability of predictions. Results indicate that the model achieves an average coverage of 95.17%–95.79% with an interval width (MS) narrowed to 0.025%–0.13%, thereby striking an optimal balance between Coverage and MS. This study provides an accurate and reliable reference for PEMFC degradation trend prediction and remaining useful life assessment.