Fine-tuning for rapid capacity estimation of lithium-ion batteries

Abstract

The widespread adoption of large-scale battery-powered technologies, such as electric vehicles and renewable energy storage systems, has led to growing interest in assessing their remaining usability after years of operation. As these systems age, state-of-health estimation becomes crucial for ensuring reliability and safety, and for extending life through second-use applications. However, current methods — spanning physics-based, empirical, and data-driven approaches — face challenges, including insufficient labeled data, high resource costs, and poor generalizability across diverse usage conditions. Data-driven models, in particular, struggle to extrapolate beyond their training domain, limiting their applicability in real-world scenarios. This work develops a fine-tuning framework to address these challenges, enabling rapid capacity estimation using short-duration ($\le 100$ s) features. Tested on two battery chemistries (LFP/Gr and NMC/Gr), the fine-tuned model achieves average mean-absolute-percent-errors of 2.592% and 3.094% on datasets collected from the respective chemistries. Compared to two baseline approaches, direct-transfer and target-only modeling, fine-tuning achieves a 25% reduction in estimation error in the target domain, on average. Domain differences are quantified using statistical measures such as Kullback–Leibler divergence and maximum mean discrepancy. These measures are found to correlate with fine-tuning performance, offering insights into domain compatibility. This study also analyzes the impact of feature selection, hyper-parameter tuning, and labeled data availability on fine-tuning efficacy, providing practical guidelines for real-world applications.