Evaluation of Thermal Comfort in Urban Commercial Space with Vision–Language-Model-Based Agent Model

Thermal comfort in urban commercial spaces significantly impacts both business performance and public well-being. Traditional evaluation methods relying on field surveys and expert assessments are often time-consuming and labor-intensive. This study proposes a novel vision–language model (VLM)-based agent system for thermal comfort assessment in commercial spaces, simulating eight distinct heat-sensitive roles with varied demographic backgrounds through prompt engineering using ChatGPT-4o. Taking Harbin Central Street, China as a case study, we first validated model accuracy through ASHRAE scale evaluations of 30% samples (167 images) by 50 experts, and then conducted thermal comfort simulations of eight heat-sensitive roles followed by spatial and interpretability analyses. Key findings include (1) a significant correlation between VLM assessments and expert evaluations (r = 0.815, p < 0.001), confirming method feasibility; (2) notable heterogeneity in thermal comfort evaluations across eight agents, demonstrating the VLMs’ capacity to capture perceptual differences among social groups; (3) spatial analysis revealing higher thermal comfort in eastern regions compared to western and central areas despite inter-role variations, demonstrating consistency among agents; and (4) the shade and vegetation being identified as primary influencing factors that contribute to the agent’s decision making. This research validates VLM-based agents’ effectiveness in urban thermal comfort evaluation, showcasing their dual capability in replicating traditional methods while capturing social group differences. The proposed approach establishes a novel paradigm for efficient, comprehensive, and multi-perspective thermal comfort assessments in urban commercial environments.