Generative Occupants: LLM Agents for Building Energy Simulation

A building’s energy use depends enormously on the people inside it — when they arrive, whether they open a window, turn the heating up or switch on equipment — yet most building simulations still represent occupants with fixed schedules or simple rules that miss how people really behave. This project is developing generative occupants: building-occupant agents, powered by large language models, that are given a persona and memory, sense a physics-based simulation of their building, and decide how to act and negotiate over shared spaces in everyday language. The aim is to capture more of the human side of energy use — habits, preferences and trade-offs — and so model buildings more realistically than schedule- or rule-based methods allow. Planned work focuses on validating the agents against real buildings and occupant data, extending the simulation beyond a single space, and moving to efficient, locally run models — building towards a tool that can give practical, simulation-backed advice on reducing building energy use, and ultimately agents that learn individuals’ preferences and represent them when building-wide decisions are made.