COSBY: Clustering-guided online sequential BYOL-based transfer learning framework for Lithium-ion battery capacity prediction

Cross-domain Lithium-ion battery capacity prediction remains challenging due to distribution heterogeneity and anomalous capacity points. To address this, we propose a clustering-guided online sequential BYOL-based (COSBY) transfer learning framework. First, we design a cluster number-driven mixture of experts (MoE) initialization method: density-based spatial clustering of applications with noise and Gaussian constraint (DBSCAN-GC) is designed to generate spherical clusters with explicit centers and radii, which initialize the MoE expert count and provide a geometric reference for gating network regularization, ensuring rational expert allocation. Second, a noise-aware MoE (NOA-MoE) distinguishes noise-free and noise samples via their distance to cluster centers: noise-free samples are processed by the noise-free MoE, while noise samples are independently modeled by a dedicated expert, effectively isolating interference from anomalous points. Finally, a phased online sequential update strategy integrating BYOL and extreme learning machine (ELM) is employed: ELM is embedded into BYOL, with the OS module frozen initially to stabilize representation convergence and activated later to update ELM weights, dynamically fitting global cross-domain distributions and avoiding misleading predictions from early-stage representations. Experiments on NASA and CALCE datasets demonstrate that COSBY outperforms advanced baselines, reducing the root mean square error by 0.2191 and mean absolute error by 0.0906 compared to state-of-the-art methods, fully verifying its cross-domain adaptability. Ablation experiments further confirm the effectiveness of each component.