Data-Driven Estimation of Li-Ion Battery Health: Integrating GPT-4 With Distilled Lifelong Learning

Abstract

Data-driven methods have attracted significant interests to estimate the state-of-health (SOH) of lithium-ion batteries (LIBs). Yet, it is often laborious and computationally costly to robustly train and implement one machine learning model for different LIB chemistries across varied operations. In light of this issue, coupled with the recent technological breakthrough of large language models, this letter exploits the strong generalisation capability of the generative pre-trained transformer-4 (GPT-4) for SOH estimation. Since battery data usually arrives sequentially with varied distribution in the real world, the teacher-to-student model-based distillation of knowledge and lifelong learning are incorporated into GPT-4 to estimate SOH adaptively with minimal fine-tuning. Testing results of the proposed method on an embedded system show a very high estimation accuracy (mean RMSE of 0.64%) at low-compute cost.