.png){kind=icon}
.png){kind=icon}
Return
Conversational preference learning for personalized thermal comfort control with a lightweight large language model
.png){kind=icon}
PRE
AI
D
X
Y
DOI：10.1016/j.enbuild.2026.117181.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Traditional personalized thermal comfort models typically rely on scalar feedback (e.g., −3 to + 3 votes) or fixed schedules, which fail to capture the nuanced semantic information embedded in human perception. While conversational interfaces offer a natural way to interact with buildings, current systems often treat user utterances as isolated commands, lacking the ability to learn stable preferences from longitudinal interaction trajectories. To address this limitation, we propose a trajectory-aware preference learning framework that leverages Large Language Models (LLMs) to infer individualized comfort temperatures from multi-round dialogue history. We constructed a dataset of 411 valid conversational trajectories, each consisting of sequential AC setpoints, measured local temperatures, and natural language feedback. Using this dataset, we fine-tuned a compact instruction-tuned LLM (Gemma-2-2B) via Low-Rank Adaptation (LoRA) to map the interaction history to a statistically optimized comfort target. Experimental results demonstrate that the fine-tuned model achieves a Mean Absolute Error (MAE) of 0.1220°C on the held-out test set. This represents a significant reduction in regression error on the test set, reducing prediction error by approximately 76% compared to a temperature-only regression baseline (MAE = 0.4988°C) and 84% compared to the zero-shot base LLM (MAE = 0.7560°C). Qualitative analysis further reveals that the fine-tuned model effectively bridges the semantic gap in human-building interaction, correctly interpreting the intensity of feedback (e.g., distinguishing “freezing” from “slightly cool”) and handling non-linear control dynamics such as overshoot corrections. By aligning linguistic priors with physical causality, our approach enables precise preference estimation, supporting more responsive and human-centric HVAC control without requiring extensive data collection for each user.
Keywords:
thermal comfort
preference learning
conversational interface
large language models
HVAC control
Journal
IF:
7.1
Papers:
1.5W
Citations:
6.8W