A conditional generative adversarial network-based synthetic data augmentation technique for battery state-of-charge estimation

• Published in: Applied soft computing Vol. 142; p. 110281
• Main Authors: Qiu, Xianghui, Wang, Shuangfeng, Chen, Kai
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.07.2023
• Subjects: Data augmentation; Generative adversarial network; Lithium-ion battery; Long short-term memory; State of charge; Long short-term memory; Lithium-ion battery; Data augmentation; State of charge; Generative adversarial network
• ISSN: 1568-4946; 1872-9681
• DOI: 10.1016/j.asoc.2023.110281

This paper proposes a battery data augmentation approach to enrich training data for state-of-charge (SOC) estimation algorithm. The approach is evaluated on battery datasets collected under various conditions to test its effectiveness. Visual comparison and low Kullback–Leibler divergence values prove that synthetic data is indistinguishable from real battery data. The computation results show that the performance of the SOC estimator can be greatly improved by adding synthetic data to the training data, and the accuracy of the estimator is even better than that of our previously proposed advanced white-box method. This data augmentation approach provides a credible way to enrich training data for SOC estimation algorithm and we have confidence that it will further accelerate the development of accurate SOC estimators. The proposed generative method is also an universal method to generate multi-type time series, rather than a method only applicable to battery data augmentation. •A generative adversarial network based battery data augmentation approach is proposed.•The proposed generative model can generate battery sequences subject to some conditional inputs.•Wasserstein distance and gradient penalty are introduced to improve the loss functions.•Visual comparison proves that the synthetic data is indistinguishable from real battery data under different conditions.•The performance of the SOC estimator can be greatly improved by adding synthetic data to the training data.