Efficient fine-tuning of large language models for automated building energy modeling in complex cases

Abstract

Building energy modeling (BEM) requires extensive time and efforts in building development. Automated building energy modeling (ABEM) is significant to reduce this burden, promoting widespread adoption of building energy modeling in building design and operation practice. This paper presents an efficient fine-tuning approach to tailor large language models (LLMs) for ABEM. Leveraging Low-Rank Adapter (LoRA) and a comprehensive training dataset (490 k samples), the proposed approach enhances LLM customization while maintaining computational efficiency. Model quantization and mixed-precision training further optimize efficiency without compromising performance. Using this approach, the developed platform (EPlus-LLMv2) can auto-generate complex buildings with varying geometries, thermal zones, materials, hourly schedules, operation settings, etc. Tests on 402 modeling cases demonstrate 100 % accuracy of modeling while reducing modeling efforts by >98 %. Additionally, an interactive human-AI interface is developed to further enhance the platform's accessibility. Finally, insights and future work to customize LLMs in ABEM and other building applications are discussed.