Federated machine learning for indoor thermal comfort: Real-time zone-based personalization and HVAC control
Journal of Building Engineering, cilt.114, 2025 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 114
- Basım Tarihi: 2025
- Doi Numarası: 10.1016/j.jobe.2025.114476
- Dergi Adı: Journal of Building Engineering
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC
- Anahtar Kelimeler: Federated learning, HVAC, Personalization, Smart buildings, Thermal comfort
- Süleyman Demirel Üniversitesi Adresli: Evet
Özet
This study addresses the critical challenge of predicting and managing indoor thermal comfort in smart buildings under conditions of high inter-individual variability and stringent data privacy requirements. Accurate, real-time comfort prediction is essential for improving occupant satisfaction, reducing unnecessary HVAC energy use, and supporting climate-resilient building operations. To solve this problem, the study proposes a personalized federated learning framework for thermal comfort prediction and HVAC control in smart buildings. The system enables real-time model adaptation at the client level by integrating subjective feedback of thermal sensation vote, clothing insulation, and metabolic rate, with environmental inputs of air temperature, relative humidity, air velocity, and outdoor conditions. A gradient-boosted regressor was selected as the base model and trained using the ASHRAE Global Thermal Comfort Database II, while real-time personalization was simulated using the Chinese Thermal Comfort Dataset across five clients. Results show substantial improvements in predictive accuracy and reduced mean absolute error through incremental client-side updates. Simulated HVAC actions achieved moderate accuracy while revealing overcooling tendencies. Deployment-oriented diagnostics further indicated improved calibration and agreement and, under an asymmetric hysteretic policy, increased neutral-zone correctness with reduced overcooling. The proposed framework supports decentralized, privacy-preserving comfort modeling and adaptive HVAC control for multi-user environments.