Distributed context-aware Transformer enables dynamic energy consumption prediction for smart building networks

The integration of communication networks and artificial intelligence enables the effective collection of data over smart building networks, facilitating more accurate predictions of Building Energy Consumption (BEC). However, existing schemes for BEC prediction suffer from limited dynamic adaptability, risks of privacy leakage, and the inability to accurately capture actual energy consumption patterns. To improve prediction accuracy while ensuring privacy and dynamic adaptability, we propose a novel BEC prediction design that incorporates dynamic threshold participation and privacy-preserving mechanisms. Specifically, we design a three-tier network architecture integrated with threshold participation tokens to support dynamic access and dropout of building entities during the BEC model construction process. Furthermore, we develop a Context-Aware Transformer (CAT) network integrated into Federated Learning (FL) to enhance feature sensitivity and facilitate the sharing of knowledge derived from Internet of Things (IoT) data and BEC features. Finally, we evaluate the performance of our design using real-world data, and the results demonstrate that our design achieves superior performance in distributed BEC prediction.