Enhancing Battery State of Health Estimation using Machine Learning Techniques

Abstract

Electric vehicles (EVs) have gained significant popularity due to their improved performance and contribution to addressing climate change. However, the cost of EVs remains a chal lenge, with lithium-ion batteries being one of the most expensive components. A correct SoH estimation allows the quantification of the residual market value of the battery pack, en abling customers to sell the battery before it exceeds its end-of-life in the target application, while still ensuring safety and reliability for reuse in other domains. This paper proposes the use of four ensemble learning algorithms to estimate the SoH of Li-ion batteries. Our approach also incorporates measures to ensure the applicability of the methods in real-world applications. This study leverages two datasets comprising charge profiles of lithium-ion batteries to develop and evaluate the proposed models. By employing popular ensemble al gorithms such as Random Forest, XGBoost, LightGBM, and CatBoost, alongside univariate and multivariate feature selection techniques, and an unsupervised anomaly detection tech nique, Isolation Forest, we address the challenge of precise battery SoH estimation. The results highlight the superior performance of LightGBM among the tested models using ap propriate features selected by feature selection techniques. By enhancing the accuracy of SoH estimation for lithium-ion batteries, this work promotes consumer confidence in the adoption of electric vehicles and renewable energy sources.