Data-Driven Mean Radiant Temperature Measurement Using Ensemble Machine Learning: A Multi-Site Validation Study
Year: 2026
Type: conference-paper
Abstract: Accurate estimation of Mean Radiant Temperature (MRT) is essential for thermal comfort assessment in buildings, yet conventional methods based on globe thermometers or dense surface-temperature sensor networks are costly and hardly scalable. This paper proposes a data-driven framework based on ensemble machine learning — Random Forest (RF) and Gradient Boosting (GB) — that estimates MRT from low-cost environmental inputs enriched with physics-informed features for solar geometry and surface-specific irradiance. Primary validation is conducted on nine months of data from two prefabricated climate test cells at the CARTIF Technology Centre (Spain), including a stringent cross-cell generalization experiment; the framework is further confirmed through a real-office deployment at Università Politecnica delle Marche (Italy). Across both sites, the models achieve$R^{2}=0.85-0.91$; on the CARTIF cross-cell generalization experiment RF yields$R^{2}=0.89$and$\text{R M S E}=1.09^{\circ} \mathrm{C}$, while on the UNIVPM real-office test the maximum absolute error is$0.72^{\circ} \mathrm{C}$and$\text{M A E}=0.19^{\circ} \mathrm{C}$. The approach overcomes the limitations of traditional MRT instrumentation by using Internet of Things (IoT) sensors and open weather data, making it suitable for HVAC optimization in smart buildings and for integration into digital twin frameworks for climate-resilient built environments.
Institutions: Marche Polytechnic University, CARTIF
Language: English
Subfield: Building and Construction
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204163256