Accurate remaining useful life (RUL) prediction of lithium-ion battery packs is essential for ensuring the reliability and safety of energy storage systems. However, the cross-domain transfer prognosis from single cells to packs faces two major challenges: nonlinear and asynchronous degradation behaviors, and multimodal noise that violates the single-Gaussian assumption commonly used in traditional models. To address these issues, this study proposes an integrated ASDMD-GMM-UT framework for lithium-ion battery packs RUL prediction, which synergistically combines an adaptive step-size differential model based on local curvature (ASDMD), and a Gaussian mixture model-based Unscented Transformation (GMM-UT). The ASDMD adaptively regulates differential step sizes via local curvature estimation, ensuring the reduction of noise interference in smooth degradation phases and accurate capture of inflection points in steep degradation phases. The GMM is employed to model the statistical distributions of both measurement noise and differential noise, while the UT realizes probabilistic propagation of multimodal uncertainty across domains. Three comparison methods were conducted to verify the accuracy and robustness of the proposed method. Experimental results showed that the proposed method significantly improved the prediction performance for battery packs. In the migration prediction from battery cells to battery packs, the root mean square error for RUL prediction was less than 4%, providing reliable guidance for maintenance scheduling and second-life utilization of battery systems.

lithium-ion battery packs; Adaptive differential model; Gaussian mixture model; RUL prediction