ResStock-LLM: A Multi-Agent Framework for Climate-Adaptive Residential Retrofit Decisions

Residential building retrofits are essential for improving energy efficiency and reducing greenhouse gas emissions, yet identifying effective retrofit actions for building stocks remains challenging. Current methods often compare pre- and post-retrofit simulations, thereby ignoring regional differences and the distribution of efficiency gaps. To improve automation in retrofit decision-making, this study introduces ResStock-LLM, a large language model (LLM) framework that integrates multiple agents with ResStock data to analyze household descriptions, forecast building energy-efficiency percentiles, and develop climate-specific retrofit strategies. ResStock-LLM links pre-trained building energy-efficiency classifiers to assess retrofit needs and compares them against national and zone-level building-stock data from ResStock. The identified retrofit targets are directed to a report agent that generates code-compliant retrofit reports. Using the structural knowledge base and ResStock data, this approach demonstrates that LLM agents can produce reliable recommendations. Tests on a representative single-family residential building show that ResStock-LLM primarily identifies roof insulation, heating setpoint, and shading improvements as key factors influencing retrofit potential. In general, ResStock-LLM offers a scalable, climate-adaptive decision-support tool that integrates building stock models with a pre-trained machine learning classifier and language model-based reasoning to facilitate efficient retrofit planning.