An Unsupervised Domain Adaptation Framework for Cross-Conditions State of Charge Estimation of Lithium-Ion Batteries
Yunpeng Liu, Moin Ahmed, Jiangtao Feng, Zhiyu Mao, Zhongwei Chen
Abstract
With the rapid development of deep learning (DL), battery state of charge (SOC) estimation has made major strides. However, the batteries’ inconsistency and changing working conditions lead to the distribution discrepancy across domains, which further affects the prediction accuracy of the pre-trained model. Moreover, collecting sufficient and labeled data is labor-intensive to gain a well-performed SOC estimator. To overcome these drawbacks, this article proposes a novel SOC estimation framework based on adversarial domain adaptation. Firstly, a distinctive SOC estimator is constructed and trained to capture the mapping relationship between the original input and the battery SOC based on the offline source dataset with a specific working condition. Then, an adversarial network with a reconstruction module and maximum mean discrepancy (MMD) constraint is designed to extract the domain-invariant features and decrease distribution discrepancy across domains. Thus, the pre-trained model could be transferred to the different working conditions using only the limited and unlabeled target data. Experimental results demonstrate that the best cross-domain root mean square error (RMSE) of the proposed transfer framework is 1.33%, 2.57%, and 1.45% for fixed ambient temperatures, changing ambient temperatures, and changing battery type, respectively, indicating that this framework emerges as a promising solution for the precise battery SOC cross-domain estimation.