Multifeature Battery State-of-Health Linear Estimation Based on Low-Cost Pretrained Inference LLM

Traditional battery state-of-health (SOH) estimation models often suffer from insufficient accuracy, weak generalization, and high deployment costs. To address these issues, an innovative multi-featur linear estimation model, DS-SOH, was proposed. This model is based on a pre-trained large language model and leverages its advanced capabilities to enhance SOH prediction. The DS-SOH transfers the capabilities of large language models to the field of battery SOH linear regression. It achieves this by extracting six key features related to battery aging. Furthermore, it establishes a dual evaluation system consisting of a reference dataset and a generalization validation set. In terms of research methodology, a multi-dimensional comparative experiment was designed: in the reference dataset test, DS-SOH achieved reductions of over 13% in error, while R² improved by 0.4%, demonstrating its advantage in linear regression. In the generalization capability validation, DS-SOH exhibited a breakthrough performance on unlearned data, reducing error by more than 44% and improving R² by 5.5%, highlighting its strong generalization ability. Deployment tests indicate that the model maintains efficient training and inference even in low-computational-power CPU environments. Research indicates that DS-SOH outperforms other linear models in terms of accuracy, generalization, and deployment adaptability. It provides a lightweight solution with practical application value for battery management systems.