Scaling Data Driven Building Energy Modeling using Large Language Models: Prompt Engineering and Agentic Workflow

Abstract

Data-driven building energy modeling (BEM) faces scalability challenges due to the complexity of diverse building and data types as well as integrating them into effective models. Large language models (LLMs) offer significant potential to enhance code generation and reasoning capabilities, which could facilitate broader adoption and implementation of data-driven BEM. In this paper, I hypothesize that LLMs can incorporate domain-specific knowledge into data processing and modeling, enabling automation of data-driven BEM across building types (residential and commercial), modeling output (zone temperature and energy consumption), and specific modeling needs. This paper leverages LLMs in the forms of prompt engineering and agentic workflow. A Machine Learning Operations (MLOps)-based prompt template is developed to systematically generate Python code for data-driven modeling. Experiments are carried out around four BEM scenarios and results indicate that both approaches are effectively scalable for implementing data-driven BMS solutions where bi-sequential prompting achieves the highest success rates of 95% in code accuracy. The agentic workflow, a paradigm where agents utilize planning, action, tools, and memory, further improves the automation, self-correction, and interaction with LLM, and resulting in improved accuracy to 100%.This framework can help energy engineers, facility managers, and sustainability consultants in automating BEM workflows, especially where coding expertise is limited, or scalability is a priority.