Analyzing Battery Discharging Capacity using Large Language Models

Lithium-ion batteries (LIBs) play a critical role in electric vehicles and energy storage solutions, emphasizing the necessity for accurate and reliable State of Health (SoH) prediction. Traditional regression-based methods, such as Recurrent Neural Networks (RNN) and XGBoost, often depend on large datasets and explicit sequence patterns, limiting their performance in scenarios with limited training data. This paper explores the applicability and advantages of employing Large Language Models (LLMs) to predict battery discharging capacity and SoH. We demonstrate that a DeepSeek-distilled 32B Qwen model (DS-Qwen) effectively leverages extensive pre-training to achieve superior predictive accuracy in low-data scenarios and offers enhanced extrapolation capabilities compared to traditional methods. Through comparative studies on distinct experimental settings, we identify conditions under which LLM-based methods outperform conventional approaches, analyze their limitations, and discuss underlying factors contributing to their performance advantages. This work provides valuable insights into the potential integration of advanced LLM frameworks for robust and interpretable battery health prediction.