Large Language Models for HVAC Operations in Building Energy Systems: A Critical Review of Methods, Applications, and Deployment Readiness

Are LLMs ready to control HVAC systems in real buildings today?

According to this review of 66 peer-reviewed studies (2023–March 2026), no. Zero studies qualified as 'ready-now', three as 'near-term', and 63 as 'research-only'. The only real-office field evidence comes from one programme (Office-in-the-Loop, ≈7.5 weeks across two periods, 47.9% energy savings in the LLM-only condition), and it remained human-in-the-loop. The authors recommend LLMs only in bounded, human-supervised roles: point-name cleanup, document-grounded operator support, building-energy-modelling assistance, and advisory interfaces around physics-based controllers (MPC/RL).

What are LLMs actually good for in building operations, per this review?

The strongest supported uses are semantic and workflow tasks where language reasoning bridges fragmented information: mapping heterogeneous BAS point names to Brick/Haystack schemas (with human validation), answering operator questions grounded in manuals and building documents via RAG, generating and debugging EnergyPlus building models (e.g., expert authoring time cut from 4 hours to 9 minutes in one study), and explaining or tuning physics-based controllers. For high-frequency closed-loop control, sub-second actuation, and short-horizon load forecasting, conventional ML, MPC and RL remain the better-supported choice.

How was this review conducted and can I check its coding?

The authors ran a single reproducible Scopus query (LLM/GPT/RAG terms AND HVAC/building-energy terms; articles, reviews, conference papers; 2023 to 31 March 2026), screening 142 records down to 66 primary studies with a PRISMA flow diagram. Each paper was coded by one reviewer across application family (A1 FDD, A2 forecasting, A3 building energy modelling, A4 control, A5 comfort), method family (M1 prompting/fine-tuning, M2 context-grounded orchestration, M3 multimodal), evidence realism, deployment readiness and responsibility boundary. The complete paper-level coding is published as a supplementary XLSX (66 studies), so the assignments are auditable, though no second-coder reliability statistics are reported.

Why does the review say load forecasting conclusions are weak?

Only three primary-coded studies address energy load forecasting (A2, n=3), all on retrospective field data, which the authors explicitly state is too sparse for subfield-level conclusions. The one direct comparison in the corpus found classical time-series models (PatchTST) outperforming LLM-based approaches at short horizons, with relative performance possibly changing at longer horizons — but with three papers, neither the gap nor any LLM advantage can be established robustly.