Large language model agents capture heterogeneous occupant behaviors in building energy simulation of shared office spaces
Year: 2026
Type: article
Abstract: Occupant behavior is a major source of uncertainty in building energy simulation, yet existing models struggle to effectively capture the behavioral diversity within shared office spaces. Specifically, traditional probabilistic models fail to characterize individual differences, while large language model agents lack sufficient exploration in multi-occupant social contexts. To bridge this gap, this study couples physics-based building energy simulation with large language model agents to represent heterogeneous occupants with distinct thermal comfort preferences, psychological expectations, and energy-saving intentions. A closed-loop framework links environmental sensing, agent reasoning and control actions in real time. Tested through a four-person shared office scenario, the simulation successfully demonstrates the technical feasibility of this framework, which yields profile-dependent behavioral dynamics. Quantitative analysis reveals a highly concentrated distribution of operational authority, where one dominant occupant executed 62.1% of all control actions, whereas the least active occupant contributed only 1.3%. Within this simulated case, this study observes a sequential propagation chain from individual differences through group interactions and action execution to energy consumption feedback. Furthermore, the proposed framework provides a computational basis for investigating socially mediated occupant behavior in human-centric building operations.
Source: Building and Environment
Language: English
Subfield: Building and Construction
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $4,280
OpenAlex ID: W7214501457