Towards real-world state of health estimation, Part 1: Cell-level method using lithium-ion battery laboratory data

Abstract

Accurate and rapid state of health (SOH) estimation is crucial for battery management systems (BMS) in lithium-ion batteries (LIBs). Given the variability in battery types and operating conditions, along with limited data samples, conventional data-driven methods are inadequate to meet the requirements, especially in real-world applications, e.g., electric vehicles and energy storage systems. To this end, we develop a meta-learning-based method with a Gated Convolutional Neural Networks-Model-Agnostic Meta-Learning (GCNNs-MAML) model to seek proper initial parameters that can rapidly adapt to new given teat samples with few-shot training. It uses multiple existing historical datasets for meta-training, and then the initial parameters of the trained model are used for meta-testing on new small-scale data. With only random 800 s charging segments from 5% of the cycling data employed for training, the GCNNs-MAML model yields a SOH estimation with a mean RMSE of 1.8% and a minimal RMSE of 1.3% on the remaining 95% testing samples. The results indicate that it remarkably outperforms the feature-based and learning-based methods. The meta-learning-based method exhibits high precision, robustness, and strong generalization capacity, implying its enormous potential for real-world applications and few-shot conditions.