State of charge estimation for lithium-ion batteries based on data augmentation with generative adversarial network

Abstract

Ever-increasing efforts in state of charge (SOC) of lithium-ion battery estimation techniques have been centered on machine learning-based methods, and a prerequisite guaranteeing the effectiveness of these methods demands an abundance of high-quality datasets. However, performing battery-related experiments is time-consuming and expensive, and manufacturers are reluctant to release their datasets due to confidentiality constraints. This limited data availability has hindered further research in this area. This paper proposes a time series augmentation model (TS-DCGAN) based on a deep convolutional generative adversarial framework. By converting the data generation problem into image generation and utilizing the advancements in GAN-based image generation, the proposed method can produce high-quality battery data. The distributional similarity and diversity of generated datasets are evaluated quantitatively and qualitatively with several metrics. The SOC estimator trained on the synthetic datasets achieved comparable execution to the actual data. Additionally, when integrated with a few real datasets, the performance of baseline SOC estimators improved to varying degrees. Extensive experiments demonstrate that the proposed method can generate synthetic data that accurately reflects the original feature distributions and temporal dynamics. This has the potential to greatly assist researchers in developing reliable battery SOC estimators.