Physics-Informed Global Model Development via Federated Learning for Indoor Temperature Prediction Toward Advanced Building Control

Model-based predictive control provides an effective approach for improving building energy efficiency by optimizing heating, ventilation, and air-conditioning operation while maintaining indoor thermal conditions. However, its performance strongly depends on the availability of an accurate and generalizable prediction model of building thermal dynamics. Existing studies have focused on individual buildings, where each building is modeled and optimized independently. Recent studies have demonstrated the potential of federated learning (FL) for a range of building applications. However, most existing approaches remain predominantly data driven. This study proposes a federated physics-informed neural network for generalized indoor temperature prediction across multiple building zones. The global RC model was developed and then incorporated into the neural-network loss function as a physics-based constraint, while FL was employed to develop a shared global prediction model without pooling the raw datasets of individual clients. The developed global model demonstrated good single-step-ahead prediction performance across the four target zones, with root-mean squared error (RMSE) values ranging from 0.36 °C to 0.59 °C, mean absolute error values from 0.25 °C to 0.57 °C. For multi-step-ahead prediction, the average RMSE was approximately 0.68 °C during the estimation period and 0.99 °C during validation. Moreover, when the zone excluded from the federated training process and evaluated as an unseen client the proposed model achieved RMSE values of 0.74 °C and 0.86 °C for one-step-ahead and multi-step-ahead predictions, respectively, compared with 0.86 °C and 0.98 °C for the conventional data-driven FL model.