BATTERYLLM: A MULTI-MODAL TEMPORAL FUSION AND PHYSICS PRIOR ENHANCED FRAMEWORK FOR BATTERY MANAGEMENT

Traditional battery management systems often exhibit insufficient generalization capabilities in complex scenarios and fail to effectively integrate multimodal data with physical and chemical knowledge. To address these challenges, this study proposes Battery LLM, a novel battery management framework that leverage large - language models (LLMs) with physical and chemical prior knowledge. Specifically, we introduce a multimodal integrated input paradigm that deeply combines battery voltage data with user instructions. This paradigm retains the physical characteristics of raw data while incorporating user-provided semantic information, forming a personalized input pattern. Furthermore, we employ Retrieval-Augmented Generation (RAG) technology to enhance user’s input. Key parameters and principles from electrochemical degradation models are encoded into structured semantic units and stored in a vectorized knowledge base, forming a domain-specific repository grounded in physical and chemical theory. At the same time, we design a multiscale temporal modeling architecture where Transformer networks capture global degradation trends, while Convolutional Neural Networks (CNNs) extract local anomaly features. A dynamic weight allocation mechanism optimizes the collaboration between these two components. Finally, the fused features are fed into the LLM for the subsequent joint prediction of SOC and SOH. Experimental results demonstrate that the proposed method outperforms state-of-the-art approaches in cross-scenario battery prediction for both SOC and SOH. This work provides a battery management solution that balances interpretability and generalization, offering a promising advancement in the field.