Multi-cycle charging information guided state of health estimation for lithium-ion batteries based on pre-trained large language model

Abstract

Continuous, stable, and accurate state of health (SOH) estimation is essential for the sustainable and reliable operation of lithium-ion batteries. However, conventional definitions and mainstream estimation methods encounter challenges in efficient implementation subject to rigorous feature engineering and complex engineering conditions. In this work, we explore feature combinations from multi-cycle charging information and employ a pre-trained large language model (PLM), which is prominent in natural language processing, for state estimation. Firstly, voltage-charge capacity curves are constructed by directly measurable data to identify candidate features across various fragmented charging processes. Posteriorly, considering the short-term stability of SOH, this paper proposes feature combinations from multi-cycle charging information to enhance the flexibility of feature engineering. Thereafter, we fine-tune the PLM to adapt to specific regression tasks, balancing prior knowledge and training efficiency. Compared to old-fashioned degradation features, the integrated multi-cycle feature combination does not require stringent prerequisites and exhibits exceptional correlation. Supplemented with GridSearch and large datasets, the proposed estimation method presents superior performance compared to other algorithms, achieving an optimal RMSE of only 0.0054. This work highlights the potential of fine-tuning the PLM for battery state estimation, leveraging innovative feature engineering technology.