Early-cycle lifetime prediction of lithium-ion batteries using a probability-anchored parallel framework with SHAP-based interpretability

Accurate lifetime prediction of lithium-ion batteries in early cycles is critical for failure warning and quality grading, yet it remains challenging due to the significant heterogeneity in degradation patterns caused by diverse operating conditions. To address the limitations of traditional unified regression models, this work proposes a probability-anchored parallel framework (PAPF) that integrates degradation pattern clustering with a specialized multi-regressor architecture. By extracting statistical features from voltage, temperature, and incremental capacity profiles within the first 100 cycles, the framework utilizes a classifier's probability distribution to dynamically weight predictions from parallel Gaussian Process Regressors (GPR). Experimental results on a heterogeneous dataset demonstrate that the proposed PAPF achieves superior performance with a Root Mean Square Error (RMSE) of 105.92 cycles and an R 2 of 0.8960. This represents an accuracy improvement of approximately 23.2% compared to the primary benchmark method, while systematic SHAP analysis further validates the effectiveness of this heterogeneity-aware modeling strategy in capturing distinct degradation mechanisms.