Estimating lithium-ion battery capacity from relaxation voltage using a machine learning approach

Abstract Lithium-ion batteries are pivotal in the development of electric vehicles and energy storage systems, with their State of Health (SOH) being crucial for both academic research and industry applications. Estimating battery capacity accurately presents significant challenges due to the complex aging mechanisms involved. In this study, we introduce a novel approach using a one-dimensional convolutional neural network (1D CNN) that leverages relaxation voltage data to predict battery capacity. The model is mainly structured with two convolutional layers, one maxpool layer, and two fully connected layers, each specially optimized to meet the unique requirements of this application. A key feature of this model is the use of a unit kernel size in the initial layer, which enhances the capture of non-linearities in the data. Our results demonstrate a prediction percentage error of -0.03% ± 0.77%, outperforming many state-of-the-art models in terms of accuracy and robustness. Furthermore, the model’s compactness, with approximately 5k parameters, suggests its suitability for edge deployment in future applications, promising significant advancements in real-time battery management.