Large language model-based methodology for data-driven health prediction of Lithium-ion batteries

Abstract

Accurate prediction of lithium-ion battery health is critical for the performance and safety of electric vertical takeoff and landing (eVTOL) vehicles. Traditional machine learning approaches require significant expertise in data preprocessing and model development, which limits their accessibility. This study introduces an innovative large language model (LLM)-based technique to automate the implementation and optimization of machine learning algorithms for battery state-of-health (SOH) forecasting. The proposed framework integrates ChatGPT into the complete machine learning pipeline, including data pre-processing, determining importance of characteristics, model recommendation and selection based on learning from reference studies, hyperparameter tuning, and performance evaluation. The LLM-driven approach involves iterative refinement of the model through structured prompts, ensuring continuous improvement and adaptation to the specific requirements of the SOH estimation. The study utilized a publicly available dataset of a lithium-ion battery used in the propulsion system of an eVTOL vehicle, which includes comprehensive flight missions and structured charge–discharge cycles. Three machine learning algorithms, i.e., Random Forest, XGBoost, and CatBoost, were implemented and optimized using ChatGPT. The performance of the LLM-driven models was benchmarked against conventional methods, demonstrating a 52% reduction in Mean Absolute Percentage Error (MAPE) compared to traditional approaches. The findings highlight the potential of LLM-driven machine learning in enhancing battery health prediction, making advanced techniques more accessible to a broader audience. This study demonstrates that integrating ChatGPT into the machine learning workflow can significantly improve the accuracy and efficiency of SOH estimation for eVTOL applications.