Comparative evaluation of large language model prompting configurations for occupant-centric thermal control

Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

Occupant thermal demands vary with age, gender, and body mass index (BMI), yet conventional fixedtemperature and Predicted Mean Vote (PMV)-based control strategies rely on standardized parameters that do not directly accommodate these characteristics. Large language models (LLMs) can interpret occupant information and indoor environmental variables to generate thermal control setpoints, but output quality is sensitive to prompting configuration. This study systematically evaluates eight LLM-based prompting configurations for occupant-centric thermal setpoint generation. Static performance was assessed using 1080 samples from the ASHRAE Global Thermal Comfort Database II, with numerical agreement against an observed neutral-condition reference and PMV Comfort Compliance Rate (PCCR), defined as PMV-based comfort-bound compliance. The static evaluation showed that Decomposition achieved the lowest mean absolute error (0.43 degrees C), whereas Modular Reasoning, Knowledge, and Language (MRKL)-style modular prompting achieved the highest PCCR (91.6 %); matched cross-model analysis showed the lowest numerical errors for Decomposition across all three models, while PCCR rankings varied by model. Dynamic performance was further examined in EnergyPlus for four selected configurations: the instruction-only baseline, Decomposition, MRKL-style modular prompting, and a hybrid of the latter two. Among these configurations, the static performance ordering was not preserved under closed-loop operation. The two PMV-anchored configurations showed higher PMV-based comfort satisfaction than the other two dynamically tested configurations, and the hybrid achieved the highest aggregate comfort satisfaction among the four. These findings indicate that numerical agreement, PMV-based comfort-bound compliance, and dynamic control behavior represent distinct evaluation dimensions and characterize configuration-dependent trade-offs under the tested conditions.

키워드

Large language models; Prompting configurations; Thermal comfort; EnergyPlus; ASHRAE global thermal comfort database II; COMFORT
제목
Comparative evaluation of large language model prompting configurations for occupant-centric thermal control
저자
Yun, Ji Young; Choi, Eun Ji; Moon, Jin Woo
DOI
10.1016/j.buildenv.2026.115159
발행일
2026-11
유형
Article
저널명
Building and Environment
권
305