The twin-federated-loop architecture: a privacy-preserving and personalized approach to thermal comfort management

Buildings account for a large share of global energy consumption, with HVAC systems being the primary contributors. Improving HVAC energy efficiency while maintaining occupant thermal comfort remains a major challenge, particularly due to the subjective and dynamic nature of comfort preferences and increasing data privacy concerns. This paper proposes the Twin-Federated-Loop (TFL) architecture, a novel framework for personalized and privacy-preserving thermal comfort management in smart buildings. The approach combines a Bipartite Digital Twin that separates physical building dynamics from occupant comfort modeling, a personalized federated learning strategy (pFedAvg), and an integrated Human-in-the-Loop mechanism. This design enables adaptive comfort prediction while keeping sensitive occupant data local. The proposed framework was validated using authentic data gathered from the NAITCD dataset and further supported by deployment in an industrial company, enabling a hybrid evaluation of simulation and operational conditions. Experimental results demonstrate that the TFL architecture achieves a 18.5% reduction in HVAC energy consumption and an approximately 15–25% reduction in the Predicted Percentage of Dissatisfied (PPD), compared with a conventional baseline control strategy. These results confirm that personalized and occupant-centric control can simultaneously enhance comfort, improve energy efficiency, and preserve data privacy, making TFL suitable for scalable and real-world smart building applications.