ThermoLLM: Thermodynamics-Aware HVAC Control with Spatial-Semantic Knowledge Graph

Researchers present ThermoLLM, a Large Language Model-based framework for multi-zone HVAC control that integrates thermodynamic physics and spatial building semantics through a knowledge graph. The system outperforms standard baselines and competing LLM approaches by reasoning about zone coupling and thermal interactions, achieving superior energy-comfort trade-offs in building simulations.

ThermoLLM addresses a critical gap in AI-driven building automation by combining language models with domain-specific physics knowledge. Previous LLM-based HVAC controllers relied on unstructured text and isolated sensor readings, failing to capture how temperature changes propagate across interconnected zones and how building geometry influences thermal dynamics. This research bridges that divide through a Brick-style semantic knowledge graph that encodes spatial relationships, adjacency patterns, and historical thermal interactions alongside real-time building state data.

The approach reflects a broader maturation in AI system design where raw pattern recognition proves insufficient for complex physical systems. Building HVAC control demands reasoning about delayed thermal responses, zone coupling effects, and long-term energy efficiency—problems that benefit from structured domain knowledge rather than prompt engineering alone. The integration of Brick semantics, a standardized ontology for building data, enables the LLM to leverage existing industry standards rather than inventing new representation schemes.

For building operators and energy management companies, this framework has commercial potential in reducing operational costs while improving occupant comfort—measured here through PMV (Predicted Mean Vote) violations. The demonstrated energy-comfort trade-off represents a meaningful advance over greedy optimization strategies that sacrifice comfort for efficiency. Real-world deployment would require adaptation to different building geometries and climate zones, but the modular knowledge graph approach suggests transferability.

Future developments should focus on validation across diverse building types, integration with actual IoT sensor networks, and comparison against traditional control theory baselines. The framework's scalability to larger, more complex buildings remains an open question, as does its robustness to sensor failures or data quality issues in production environments.

• →ThermoLLM achieves superior energy-comfort trade-offs by grounding LLM control decisions in physics-informed spatial knowledge graphs derived from building semantics.
• →The framework incorporates zone coupling and thermal interaction modeling, addressing limitations of prior LLM-based HVAC controllers that ignored building structure.
• →Integration of Brick-style building ontology enables standardized semantic reasoning about spatial relationships and thermal dynamics across zones.
• →Results demonstrate the lowest PMV violations compared to standard baselines and alternative LLM approaches while maintaining energy efficiency.
• →The approach suggests a broader pattern where specialized domain knowledge graphs enhance LLM reasoning for complex physical systems beyond pure language tasks.