Adaptive confidence-calibrated semi-supervised ensemble for early-cycle battery lifetime prediction

Accurate prediction of lithium-ion battery lifetime during the early cycles is essential for accelerating technology development and ensuring operational safety. However, the minimal capacity degradation observed in the initial cycles makes it challenging to extract informative features from early-stage data. Furthermore, the high cost and time consumption of battery aging experiments severely restrict the availability of labeled samples, posing significant challenges to early-life prediction.To address these issues, this study proposes an Adaptive Confidence-Calibrated Semi-Supervised Ensemble (ACCSE) framework. The framework extracts thirteen degradation-sensitive features from the first 100 charge-discharge cycles, systematically capturing early battery aging behavior through capacity-voltage derivatives, incremental capacity shifts, thermal responses, and voltage polarization indicators. An adaptive pseudo-labeling strategy, guided by model uncertainty, is introduced to automatically identify high-confidence unlabeled battery samples, thereby alleviating the constraint of limited labeled data. In addition, the ACCSE framework dynamically determines an optimal subset of base learners via greedy forward selection and integrates them using RidgeCV-based meta-regression, effectively leveraging their complementary strengths to improve prediction accuracy and robustness. Experimental results demonstrate that, using only the first 100 cycles, the ACCSE method achieves a root mean square error (RMSE) of 69 cycles in lifetime prediction. This represents an improvement of over 12% compared to several state-of-the-art data-driven benchmarks. Notably, even under conditions of extreme label scarcity with only 10% labeled data, the proposed method maintains robust performance with an RMSE of 212 cycles.