Remaining useful life prediction of lithium-ion batteries based on trend-residual decomposition and large-kernel attention Mamba model

Accurate prediction of a battery system's remaining useful life (RUL) is essential for maintaining operational safety and supporting effective maintenance planning. It also helps extend battery life and improve the operational reliability and maintenance efficiency of energy storage systems. Battery degradation signals exhibit a slowly evolving long-term decline combined with rapidly fluctuating local variations. These components are strongly coupled in the temporal domain and are susceptible to noise, which poses challenges in preserving long-term consistency while capturing transient dynamics. To tackle this problem, we propose a trend–residual decomposition-based, large-kernel-attention-enhanced Mamba framework (LEMamba). We employ Difference-Aware Exponential Smoothing (DAES) to extract the long-term degradation trend and adaptively adjust the smoothing coefficients based on local variations. Simultaneously, we design a feature modulation mechanism that reduces redundancy and stabilizes the feature distribution by employing pooling and linear projection, thereby preserving key dynamic information while suppressing noise. The residual pathway incorporates large-kernel attention block (LKAB) to enhance the model's sensitivity to short-term fluctuations and, by leveraging its large receptive field, capture cross-temporal associative structures for more accurate local dynamic modeling. Experimental results on the CALCE, NASA, and XJTU datasets demonstrate that our LEMamba achieves better performance in the RUL prediction task, outperforming existing mainstream methods overall.