An LLM Framework for Modeling U.S. Thermal Comfort Perceptions

Abstract

This study explores the use of large language models (LLMs) to analyze thermal comfort discourse on social media as a proxy for real-time thermal satisfaction in the United States. Using the OpenAI API, a longitudinal sentiment analysis was conducted on 68,396 U.S.-based tweets posted in 2019 and 2020 to examine changes before and during the COVID-19 disruption. The methodological pipeline included temporal grouping, text preprocessing, semantic interpretation, relevance classification, sentiment labeling, confidence scoring, and final filtering of thermal comfort-related content. Of the full dataset, 55,448 tweets were classified as relevant to thermal comfort, including 33,081 tweets in 2019 and 22,367 in 2020. Results show a marked decline in tweet volume in 2020, consistent with pandemic-related changes in mobility, occupancy, and daily routines. Despite this reduction, sentiment patterns remained stable: neutral tweets constituted the largest share, negative sentiment exceeded positive sentiment, and seasonal and weekday variations persisted. Lexical analysis further showed that thermal comfort discourse was dominated by direct experiential terms such as cold, hot, humidity, and weather, with greater visibility of home- and air-related terms in 2020. The findings indicate that LLM-based sentiment analysis can capture large-scale shifts in public thermal perception and provide a novel methodological framework for using digital sentiment as an indicator of social feedback on environmental comfort. This approach offers practical value for human-building interaction research and for the development of more resilient and human-centered building performance standards.