Large Language Models in HVAC Systems: Methods, Current State, and Deployment Readiness – A Comprehensive Review

Background and Context

Building automation systems generate vast quantities of sensor data daily, yet they frequently suffer from a critical paradox: data abundance coupled with insight scarcity. This operational bottleneck stems primarily from heterogeneous point naming conventions, missing metadata, and fragmented documentation systems, which severely hinder the direct conversion of raw data into operational value. To address these systemic inefficiencies, a comprehensive systematic review analyzed 66 peer-reviewed studies published between 2023 and March 2026, focusing on the application of Large Language Models within Heating, Ventilation, and Air Conditioning operations. The research represents the first attempt to evaluate the maturity of this technology across three distinct dimensions: methodology, application scenarios, and deployment readiness.

The review categorizes existing literature into five application families and three LLM methodology families, introducing the novel concept of responsibility boundaries to analyze the interaction between LLMs and physical HVAC decision-making. This multidimensional classification reveals that current implementations view LLMs not as autonomous controllers, but as semantic layers and workflow assistance tools. By mapping the specific roles LLMs play in bridging unstructured data with structured control systems, the study provides a clear roadmap for future research, highlighting the gap between theoretical potential and practical industrial utility.

Deep Analysis

From a technical perspective, the review details the specific implementation paths of LLMs in HVAC operations. While traditional machine learning, Model Predictive Control, Reinforcement Learning, and ontology-based tools continue to dominate high-frequency control, short-term numerical prediction, and well-structured ontology mapping, LLMs offer unique advantages in semantic processing. The primary technical applications identified include point name standardization, operator support based on documentation, Building Energy Modeling workflow assistance, and consultation interfaces surrounding physics-based controllers. These approaches do not attempt to replace underlying control algorithms but instead position LLMs as intermediaries that translate complex, unstructured information into actionable insights for human operators or standardized data formats.

Experimental settings and key results expose the stark reality of the current research landscape. The analysis shows a heavy concentration of resources on Building Energy Modeling, which accounts for 32 of the 66 papers, whereas load forecasting remains sparse due to data limitations, preventing definitive conclusions in that sub-field. Crucially, only four studies achieved pilot-level evidence, and none reported continuous operational deployment. In terms of deployment readiness, no study was classified as immediately ready for industry adoption, with only three deemed near-term ready. The remaining 63 studies remain in the pure research phase, indicating that while proof-of-concept demonstrations are promising, significant barriers to real-world integration persist.

Ablation analysis further clarifies which applications are most viable in the near future. Scenarios with well-defined boundaries and human-in-the-loop supervision, such as point name standardization and document support, are more likely to reach pilot trials. Conversely, attempts to enable LLMs to fully autonomously control HVAC systems lack empirical support and remain theoretical. The review emphasizes that autonomous agent operations and unverified occupant proxy models are still in experimental stages, underscoring the necessity of human intervention to ensure safety and reliability in physical environments. This distinction is vital for understanding the current limitations of LLMs in safety-critical infrastructure.

Industry Impact

The implications of this review for the building energy sector are profound, offering a pragmatic guide for the deployment of Large Language Models. The evidence strongly supports positioning LLMs as semantic understanding and workflow enhancement layers rather than autonomous HVAC controllers. For the open-source community and industrial stakeholders, this shift in perspective is crucial. It suggests that immediate efforts should focus on data governance, document retrieval, and operational assistance, rather than pursuing fully autonomous control systems that lack robust safety validations. This realistic assessment helps mitigate deployment risks and aligns technological development with actual operational needs.

The review highlights that the integration of LLMs into existing building management systems requires a rethinking of how data is structured and utilized. By automating the normalization of point names and interpreting complex system documentation, LLMs can significantly reduce the manual effort required for system maintenance and troubleshooting. This capability addresses one of the most persistent pain points in the industry: the difficulty of onboarding new personnel and maintaining legacy systems with inconsistent documentation. Consequently, the immediate value of LLMs lies in their ability to lower the barrier to entry for effective system operation, rather than in replacing expert human judgment.

Furthermore, the findings challenge the prevailing narrative that AI will rapidly replace traditional control methods in the near term. Instead, they suggest a complementary relationship where LLMs handle high-level semantic tasks while traditional algorithms manage precise, low-latency physical controls. This division of labor ensures that the reliability and determinism required for HVAC operations are maintained, while leveraging the flexibility and interpretability of language models for administrative and analytical tasks. Such an approach fosters a more resilient and adaptable building automation ecosystem, capable of evolving with changing operational requirements without compromising safety.

Outlook

Looking ahead, the review recommends prioritizing the development of field-validated benchmarks to objectively assess LLM performance in HVAC contexts. Current evaluations are largely theoretical or based on synthetic data, lacking the rigor needed for industrial adoption. Establishing standardized benchmarks will enable researchers and practitioners to compare models effectively and identify best practices for specific operational challenges. Additionally, the study advocates for exploring hybrid architectures that combine LLMs with Model Predictive Control and Reinforcement Learning. These hybrid systems would leverage the high-level semantic reasoning of LLMs alongside the low-latency, high-reliability of traditional control algorithms, ensuring verifiable safety attributes.

Future research must also focus on orchestration evaluation under operational constraints, particularly regarding latency and safety. The integration of LLMs into real-time control loops introduces new variables that must be carefully managed to prevent system instability. Developing frameworks that can dynamically adjust the level of LLM involvement based on system state and risk assessment will be critical for successful deployment. This adaptive approach allows for the benefits of AI-driven insights while maintaining strict control over physical system parameters, ensuring that safety is never compromised for the sake of automation.

Ultimately, the path to industrial deployment lies in a cautious, incremental integration of LLMs into building automation systems. By focusing on auxiliary roles that enhance human decision-making and streamline data management, the industry can harness the power of Large Language Models without exposing itself to unnecessary risks. As the technology matures and more field-validated evidence becomes available, the scope of LLM applications is expected to expand, potentially leading to more intelligent, responsive, and efficient building energy management systems. The review serves as a foundational guide for navigating this transition, emphasizing the need for rigorous validation and thoughtful architectural design in the pursuit of smarter, safer buildings.