Yunlong Liu, Pengfei Zhou, Hao Wang, Zhiqiang Qin, Ji Wu, Wangqiu Zhou, Gang Luo, Qi Song, Xiangyang Li
Data Mining and Knowledge Discovery (DMKD)Published
The reuse of retired batteries offers substantial environmental and economic benefits, with battery capacity estimation serving as a pivotal component in this process. Although deep learning methods have achieved remarkable progress in battery capacity estimation, their development has also revealed three key challenges. First, existing datasets remain limited in size, hindering models from effectively learning diverse degradation patterns. Second, the absence of standardized datasets and evaluation protocols impedes fair and reproducible performance comparisons. Finally, existing models struggle to balance high estimation accuracy with the lightweight design required for industrial deployment. To address these challenges, we introduce ReLiB-100k, a large-scale dataset and benchmark for battery capacity estimation. ReLiB-100k contains single-shot charge and discharge data from over 100,000 battery cells, spanning eight nominal capacities and various states of health, thereby providing a comprehensive resource for learning degradation patterns. Using this dataset, we systematically evaluate the performance of ten representative and emerging time series estimation models, establishing a unified and reproducible benchmark for future research. Additionally, we propose a lightweight model, CapCLR, which achieves a lower average MAPE than the current state-of-the-art method while using only 1/11 of its parameters.
Yunlong Liu, Pengfei Zhou, Hao Wang, Zhiqiang Qin, Ji Wu, Wangqiu Zhou, Gang Luo, Qi Song, Xiangyang Li
Data Mining and Knowledge Discovery (DMKD)Published
The reuse of retired batteries offers substantial environmental and economic benefits, with battery capacity estimation serving as a pivotal component in this process. Although deep learning methods have achieved remarkable progress in battery capacity estimation, their development has also revealed three key challenges. First, existing datasets remain limited in size, hindering models from effectively learning diverse degradation patterns. Second, the absence of standardized datasets and evaluation protocols impedes fair and reproducible performance comparisons. Finally, existing models struggle to balance high estimation accuracy with the lightweight design required for industrial deployment. To address these challenges, we introduce ReLiB-100k, a large-scale dataset and benchmark for battery capacity estimation. ReLiB-100k contains single-shot charge and discharge data from over 100,000 battery cells, spanning eight nominal capacities and various states of health, thereby providing a comprehensive resource for learning degradation patterns. Using this dataset, we systematically evaluate the performance of ten representative and emerging time series estimation models, establishing a unified and reproducible benchmark for future research. Additionally, we propose a lightweight model, CapCLR, which achieves a lower average MAPE than the current state-of-the-art method while using only 1/11 of its parameters.