Large Language Models for HVAC Operations in Buildi... | AI Research

Large Language Models for HVAC Operations in Building Energy Systems: A Critical Review of Methods, Applications, and Deployment Readiness
This paper provides a systematic review of 66 peer-reviewed studies published between 2023 and March 2026, investigating how Large Language Models (LLMs) can improve the operation of Heating, Ventilation, and Air Conditioning (HVAC) systems. While modern buildings generate vast amounts of sensor data, this information is often difficult to use due to inconsistent naming, missing metadata, and fragmented documentation. The authors evaluate whether LLMs can bridge this gap by acting as a "semantic layer" that helps operators and engineers interpret data and manage building workflows more effectively. To see openai in practice, Gemini's now Generates Files! walks through a concrete example.

The Role of LLMs in HVAC

The research suggests that LLMs are best suited to act as a bridge between fragmented information sources—such as equipment manuals, maintenance logs, and building automation system (BAS) data—and the human operators who need to act on them. Rather than replacing existing physics-based controllers, LLMs can help normalize point names, provide document-grounded support for technicians, and assist in building energy modeling (BEM) workflows. By translating complex technical information into accessible language, LLMs can help overcome the "insight-poor" nature of current building management systems.

How the Research is Categorized

The authors classify the 66 studies into five application areas and three primary method families to understand how these models are being applied:

• Application Families: Fault detection and diagnosis (FDD), energy load forecasting, building energy modeling (BEM), HVAC control and optimization, and thermal comfort/occupant interaction.

• Method Families: * M1 (Prompting/Supervised Adaptation): Using pre-trained models with minimal task-specific training.

• M2 (Context-Grounded Orchestration): Connecting LLMs to external tools, databases, and retrieval systems to ground their outputs in real-world data.

• M3 (Multimodal Inputs): Incorporating non-textual data like floor plans, heatmaps, and diagrams. The ai agents story also surfaces in OpenAI Unveils GPT-Red an Automated Model..., adding another angle.

Current Deployment Readiness

A key finding of the review is that the field is still in an early, research-oriented stage. Of the 66 studies analyzed, none were classified as "ready-now" for industry adoption. Only three were considered "near-term," while the vast majority (63) remain strictly in the research phase. The authors note that while LLMs show promise, they currently lack the field-validated benchmarks and safety guardrails required for autonomous operation in real-world, safety-critical HVAC environments.

Future Directions and Limitations

The paper emphasizes that HVAC operations are physics-constrained and time-dependent, meaning that high accuracy in a simulation does not guarantee success in a real building. The authors argue that future work must prioritize "governed integration," where LLMs are used in bounded, human-in-the-loop roles. They recommend that researchers focus on creating field-validated benchmarks, improving the reliability of orchestration workflows, and ensuring that any LLM-driven system includes verifiable safety properties and human oversight before being deployed in physical building environments. The ai agents story also surfaces in OpenAI agents break out of sandbox..., adding another angle. as detailed in the full paper on Arxiv