Accurate prediction of the remaining useful life (RUL) of lithium-ion batteries is essential for ensuring safe operation, reducing maintenance costs, and improving energy efficiency in electric vehicles, aerospace equipment, and industrial applications. However, existing methods often struggle to capture long-term temporal dependencies, local capacity regeneration, and nonlinear degradation characteristics. In a study published in ENG. Chem. Eng., researchers at Qingdao University and collaborators propose CA-Mamba2, a novel framework that addresses these challenges through state-space modeling enhanced by coordinate attention and gated residual networks.
The CA-Mamba2 framework consists of three core components. The coordinate feature attention network (CFAN) extracts direction-specific information along both temporal and variable dimensions, generating bidirectional attention weights that highlight key degradation stages and critical health variables. The Mamba2 backbone models long-range degradation evolution using a selective state-space mechanism that dynamically adjusts parameters based on input content, efficiently capturing global features while maintaining low computational overhead. The swiGLU-gated residual network (SGRN) introduces nonlinear gated transformations before the final prediction layer, improving the decoder’s ability to map complex degradation features to RUL while preserving original representations through residual connections.
The model was evaluated on three public datasets from NASA, Tongji University (TJU), and Xi’an Jiaotong University (XJTU), covering different battery chemistries and test conditions. Under single-variable input on the NASA dataset (B0005 test), CA-Mamba2 achieved MAE values of 0.0089–0.0098 and RMSE of 0.0140–0.0172 across prediction starting points SP = 50, 70, and 90, with R² values of 0.949–0.989 and absolute errors of 1.2–2.3 cycles – outperforming LSTM, Transformer, PathFormer, TimeMixer, and baseline Mamba.
Under multivariate input on the TJU dataset (17 features including voltage/current statistics and charging characteristics), CA-Mamba2 achieved MAE of 0.0014–0.0016, RMSE of 0.0022–0.0024, R² of 0.9995–0.9998, and absolute errors of 2.4–2.5 cycles across SP = 200, 300, and 400 – significantly surpassing all comparison methods. Notably, while the absolute error of TimeMixer increased from 12.0 to 50.7 cycles as SP shifted, CA-Mamba2 remained stable, demonstrating insensitivity to prediction starting point changes.
On the XJTU dataset, cross-dataset generalization tests showed MAE of 0.0049–0.0083 and RMSE of 0.0066–0.0111, with R² consistently above 0.986, confirming strong cross-battery and cross-stage generalization capabilities.
Computational efficiency analysis showed CA-Mamba2 required only 57.848 s for training and 0.476 s for inference on the TJU dataset – significantly lower than Transformer (423.765 s training, 1.528 s inference) and other baselines. Ablation experiments confirmed that both CFAN and SGRN contribute positively, with their combination achieving the best overall performance.
CA-Mamba2 provides a robust, efficient, and generalizable solution for lithium-ion battery RUL prediction, with strong potential for deployment in battery management systems for electric vehicles and energy storage applications.
DOI
10.1007/s11705-026-2687-x