Abstract
Liquid metal batteries (LMBs) represent an emerging safe energy storage technology. Accurate prediction of their degradation trajectories is essential for battery safety management and reliable large-scale deployment. However, their degradation behavior differs markedly from that of conventional batteries, exhibiting pronounced nonstationarity and complex features such as early-stage capacity drops. These distinctive degradation mechanisms make conventional data-driven models highly susceptible to error accumulation and overfitting during long-horizon extrapolation. To address these challenges, this study proposes a degradation prediction framework based on stationary wavelet-based decoupling and component-specific modeling. The framework first performs multiscale decomposition according to the intrinsic characteristics of historical aging data to decouple degradation patterns. Each decomposed component is then matched with a dedicated model based on its specific characteristics, enabling component-level prediction. By integrating the collaborative predictions of all submodels, the proposed framework accurately captures both gradual capacity fading and abrupt capacity drops. The framework is validated on two types of LMBs with capacities of 20 Ah and 50 Ah under multiple prediction starting points. Even with limited training data, it achieves excellent predictive performance, with root mean square errors (RMSE) as low as 0.057 Ah and 0.114 Ah, respectively. The experimental results demonstrate that, without incorporating future operating information, explicit multiscale degradation decoupling combined with component-specific modeling can effectively overcome the bottleneck of error accumulation under complex nonlinear degradation, substantially improving the stability and accuracy of long-horizon degradation prediction for LMBs. This framework also provides effective technical support for health-state management of LMBs.