Abstract
Accurate energy forecasting for industrial IoTequipped modern built environments is vital for sustainable energy resource management. However, using data from multiple buildings raises privacy concerns. This paper proposes a privacypreserving federated learning framework with two approaches: FTran, which uses federated learning with a transformer-based model, and SeFTran, which adds secure aggregation. In SeFTran, we included secure aggregation where model weights are split into additive shares using additive secret sharing, encrypted via Advanced Encryption Standard (AES) algorithm, with keys exchanged via the Diffie-Hellman (DH) protocol and aggregated without exposing weights. This ensures client data remains private from both the server and other clients. We evaluated our models against several baselines, including ANN, LSTM, BiLSTM, and Stacked LSTM models. Results show that proposed FTran and SeFTran achieve high prediction performance with RMSE scores of 1.2346 and 1.2441, and MAE scores of 0.9972 and 0.9968, respectively, while SeFTran providing strong privacy protection and performance close to unencrypted FTran. This demonstrates the potential of the proposed secure federated learning framework for smart building systems.