Lithium-Ion Battery State of Charge Estimation With Adaptability to Changing Conditions

• Published in: IEEE transactions on energy conversion Vol. 38; no. 4; pp. 2860 - 2870
• Main Authors: Meng, Zhan, Agyeman, Kofi Afrifa, Wang, Xiaoyu
• Format: Journal Article
• Language: English
• Published: New York IEEE 01-12-2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adversarial machine learning; Artificial intelligence; Battery aging; changing conditions; Data models; Deep learning; Lithium-ion batteries; lithium-ion batteries (LIBs); Rechargeable batteries; State of charge; state of charge (SOC) estimation; Transfer learning
• ISSN: 0885-8969; 1558-0059
• DOI: 10.1109/TEC.2023.3285405

Accurate state of charge (SOC) estimation provides critical information to ensure the safe operation of lithium-ion batteries (LIBs). With the evolution of artificial intelligence, the model-free deep learning-based SOC estimation has made extraordinary progress. However, collecting sufficient data to train an accurate SOC estimator of batteries' whole lifecycle is costly, and generalizing the estimator to different operating conditions is challenging. To this end, aiming to enhance the generalization ability of data-driven SOC estimators, this article proposes a novel SOC estimation framework based on adversarial transfer learning (TL), which transfers a well-trained SOC estimator with fewer data demands to fit different conditions. Specifically, a source SOC estimator is trained under a specific condition to learn the basic characteristics of the battery, which is defined as the source domain. Target domains correspond to measured data of the battery under different conditions. An adversarial TL-based training framework is developed to extract the domain invariants of the source and target domains, which is guided by minimizing the distribution discrepancies. The effectiveness of the proposed method is demonstrated through the use of two LIBs datasets. The successful generalization to multiple operating conditions and aging health states demonstrates the estimation accuracy, robustness, and good generalization.