State-of-charge estimation across battery chemistries: A novel regression-based method and insights from unsupervised domain adaptation

Abstract

Battery management systems play a critical role in the ongoing efforts for decarbonization by enhancing the effectiveness of energy storage solutions. A central problem within this domain focuses on the estimation of state-of-charge (SOC), pivotal for preserving battery health and averting unforeseen failures. While most methods perform well in controlled environments, deploying SOC estimation methods in industrial applications introduces significant challenges due to the inherent diversity in battery chemistry and operating conditions encountered in real-world scenarios. Current approaches often mandate extensive data gathering capabilities tailored to specific battery chemistries and operating conditions, entailing costly and time-intensive processes. These hurdles are compounded by limited access to ground truth data and the dynamic evolution of operational conditions, further complicating the viability of existing methodologies. In this study, we introduce a novel SOC estimation method that leverages regression-based unsupervised domain adaptation for cross-battery SOC estimation. Through a comprehensive comparative analysis with existing classification-based domain adaptation methods, we demonstrate the superior predictive accuracy of our approach. Our investigation also unveils trends in transfer learning capability across various conditions and methods. The findings underscore the substantial enhancements offered by the regression-based unsupervised domain adaptation method over conventional classification-based approaches in cross-battery SOC estimation.