Advanced building control methods like Model Predictive Control (MPC) and Reinforcement Learning offer energy savings but struggle with scalability and interpretability. We introduce DARLIN, a retrieval-augmented framework using Large Language Models for optimal cooling control without system-specific training or modeling. DARLIN queries a curated knowledge base with real-time data to generate optimal cooling setpoints and human-readable rationales. In calibrated building simulation, DARLIN achieved 7.06% cooling energy savings, approaching theoretically perfect MPC baseline (9.42%) while maintaining occupant comfort. Interpretability analysis of generated rationales confirmed its trend-aware, predictive, and corrective reasoning capabilities. DARLIN offers a scalable, interpretable paradigm for building system control.
Paper
Full text
DARLIN: Domain-guided Augmented Retrieval for LLM-based INterpretable HVAC Control
Semantic Scholar · Computer Science · 2025
Abstract
Advanced building control methods like Model Predictive Control (MPC) and Reinforcement Learning offer energy savings but struggle with scalability and interpretability. We introduce DARLIN, a retrieval-augmented framework using Large Language Models for optimal cooling control without system-specific training or modeling. DARLIN queries a curated knowledge base with real-time data to generate optimal cooling setpoints and human-readable rationales. In calibrated building simulation, DARLIN achieved 7.06% cooling energy savings, approaching theoretically perfect MPC baseline (9.42%) while maintaining occupant comfort. Interpretability analysis of generated rationales confirmed its trend-aware, predictive, and corrective reasoning capabilities. DARLIN offers a scalable, interpretable paradigm for building system control.