@article{GU2026, 
author = {Zhangqing GU and Maoyun YANG and Shijin REN and Guosheng HAO and Shiyu ZHANG and Shuai ZHONG and Xiaoguang ZHANG},
title = {Remaining useful life prediction for lithium batteries based on a multi-scale width-ensemble extreme learning machine},
year = {2026},
journal = {Journal of Chongqing University},
volume = {49},
number = {8},
pages = {66-82},
keywords = {lithium batteries, remaining useful life prediction, CEEMDAN, dilated convolution, Extreme Learning Machine, Elastic Net, Broad Learning System},
url = {https://www.sciopen.com/article/10.11835/j.issn.1000-582X.2026.08.007},
doi = {10.11835/j.issn.1000-582X.2026.08.007},
abstract = {Accurate prediction of the remaining useful life (RUL) of lithium batteries is crucial for predictive maintenance and health management. Existing methods often require large training datasets, exhibit low computational efficiency, and struggle to satisfy real-time application requirements due to variations in operating conditions and manufacturing processes. To address these limitations, this paper proposes a lithium battery RUL prediction method based on multi-scale feature extraction and an extreme learning machine (ELM) ensemble. To capture the spatio-temporal features of complex battery degradation data, complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and dilated convolution are employed to extract multi-scale features, enabling accurate identification of modal variations in battery lifetime behavior. An integrated ELM-based RUL prediction model is then built using elastic net (EN) regularization, while incorporating concepts from the broad learning system (BLS) and ensemble learning. This design improves prediction accuracy and robustness for complex sequential data while remaining suitable for resource-constrained scenarios. Experimental results on the university of maryland center for advanced life cycle engineering CALCE dataset reveal that the proposed method yields an 81.25% reduction in RMSE, a 60.43% decrease in MAPE, and a 0.74% improvement in R2 relative to the best-performing deep learning model. Compared to the optimal conventional machine learning model, training time increases by less than 1 ms, while prediction time drops by 45%. These results demonstrate that the proposed method significantly enhances prediction accuracy and robustness, enabling lightweight models to perform better in resource-constrained circumstances.}
}