Digital Twin for Capacity Estimation of Lithium-Ion Batteries With Large Language Models

Accurately online estimating the state of charge (SOC) of lithium-ion batteries is critical for the reliability, safe operation, and optimal performance of battery management systems (BMSs). However, cross-domain battery capacity estimation is facing challenges like: limited adaptability to diverse degradation patterns across operating conditions, imbalanced convergence during model training, and the inability to provide real-time feedback on battery health. To tackle these problems, this article introduces cosine similarity (CS) into the cross-domain prediction model UniTime, combining the strong generalization ability of large language models (LLM) with the online prediction advantages of digital twin (DT) technology. Specifically, we enhance cross-domain adaptability by introducing the CSUniTime model, which leverages domain instructions and a language transformer to achieve effective alignment between time-series data and textual modalities. To address convergence imbalance, we design a CS attention mechanism that efficiently extracts degradation features while maintaining lightweight complexity. Furthermore, a dynamic degradation simulator based on DT technology is developed to realize online estimation using only charging-phase data, thereby enabling real-time feedback of the battery’s health status. Extensive experiments on two public datasets show that DT-CSUniTime consistently outperforms state-of-the-art baselines, achieving the best mean squared error (mse) on 20 of 21 batteries and delivering superior accuracy and generalization.