When Data Handling Alters Physical Interpretation: HVAC’s Missing Evidence Layer

Key Takeaways

• Missing data inflates temperature importance by 135%
• Accurate predictions can mask physical mis‑interpretation
• Imputation reduces variability, collapsing distinct thermal variables
• AI learns collection patterns, not true environmental physics
• Industry needs immutable evidence layer before model inference

Summary

A 2026 Energy and AI study shows that HVAC thermal‑comfort models produce wildly different variable importance ratios when missing data is handled differently, shifting the air‑temperature to mean‑radiant‑temperature (Ta:MRT) ratio from 1.9:1 to 4.46:1 – a 135% change. The researchers used the same dataset and LightGBM model, demonstrating that data imputation can create artificial correlations and suppress radiant effects. The most accurate models, judged by prediction error, were also the most misleading in physical interpretation. The authors argue that the industry must move from a data‑centric to a truth‑centric workflow that preserves complete environmental evidence at capture.

Pulse Analysis

The recent study highlights a hidden vulnerability in modern HVAC analytics: the way missing sensor readings are treated can fundamentally rewrite the physics the model believes it is observing. When mean radiant temperature (MRT) is absent, practitioners often substitute air temperature or let machine‑learning algorithms infer the gap, inflating the apparent influence of air temperature by more than double. This phenomenon is not limited to HVAC; any built‑environment system that relies on sparse sensor networks faces the same risk of drawing unstable conclusions from incomplete evidence.

Machine‑learning platforms such as LightGBM treat missing values as informative signals, allowing the model to learn patterns of data collection rather than true environmental dynamics. The result is a high‑accuracy model that, paradoxically, misrepresents the underlying physics, leading designers to undervalue radiant heating or cooling solutions and to over‑optimize for air‑side control. The industry’s current “collect‑clean‑impute‑model” pipeline therefore amplifies bias, making AI‑driven decisions fragile and difficult to defend in regulatory or performance‑verification contexts.

The path forward requires a shift toward a "truth system" that captures, timestamps, and preserves raw environmental measurements before any reconstruction occurs. Continuous, append‑only recording, coupled with an evidence‑governance framework, can ensure that every model input is admissible and that AI algorithms learn from genuine physical signals. By institutionalizing immutable evidence layers, building owners and operators can achieve more reliable energy models, robust control strategies, and ultimately, greener, cost‑effective facilities.