Abstract
Accurate early prediction of lithium-ion battery degradation is essential for safe and reliable energy storage systems, yet it remains challenging due to complex aging mechanisms and prediction uncertainty. This paper proposes a novel two-stage multi-task early prediction framework with uncertainty quantification. The model decomposes the prediction task into inflection point identification, cycle life prediction, and trajectory reconstruction, ensuring the reconstructed SOH curve strictly follows physical decay patterns. Three characteristic curves (V-Q, IC-V, DV-Q) are integrated by using a multi-scale architecture combining parallel CNNs and iTransformer for local and global feature extraction. Furthermore, conformal prediction is employed to provide not only high-accuracy point estimates but also prediction intervals with a guaranteed 95% confidence level, offering quantifiable risk indicators. Validated on the HUST and XJTU datasets, our results demonstrate that the proposed model achieves superior prediction accuracy and provides well-calibrated uncertainty bounds, representing a significant advancement towards trustworthy and intelligent battery management systems. The errors of the model with complete input of three characteristic curves are the lowest compared with other three single characteristic curves. The MAE and the RMSE of the proposed framework on XJTU datasets achieved to 0.0058% and 0.0071%.