Thermal comfort prediction with limited local data utilizing heterogeneous external sources via federated learning

Thermal comfort prediction plays a critical role in intelligent building control, yet it remains challenging under data scarcity. While external data can potentially improve prediction performance, effectively utilizing heterogeneous distributed data under privacy constraints remains challenging because not all external knowledge is equally relevant to the target environment. To address this limitation, this study proposes Federated Learning with Evaluation Layer (FedEL), a host-centric framework that enables a target client to evaluate and selectively integrate model updates from multiple sources. By incorporating a local evaluation mechanism and a ranking-based weighting strategy, FedEL prioritizes updates that are most relevant to the host objective while maintaining robustness under limited data. Experiments conducted on a multi-source thermal comfort dataset demonstrate that FedEL improves personalized thermal comfort prediction under heterogeneous data conditions. The proposed framework achieves the greatest benefits when local data are limited, where effective utilization of external knowledge is particularly important, while remaining competitive with local training and transfer learning when sufficient local data are available. The proposed framework provides a practical mechanism for privacy-preserving knowledge sharing across distributed data sources, enabling data-limited participants to benefit from external information without compromising data ownership. This capability supports more effective and accessible thermal comfort modeling and offers strong potential for real-world intelligent building applications.