Federated Learning for Building Energy Prediction and Fault Detection Across Climate Zones: A Proof-of-Concept Study

The escalation of global building energy consumption, driven substantially by heating, ventilation, and air conditioning (HVAC) systems, necessitates robust, privacy-preserving analytics for energy prediction and fault detection.However, building meter data is fragmented across owners and vendors, preventing centralization of raw data for model training.While federated learning offers a solution, systematic comparisons under real-world cross-climate heterogeneity remain limited.This paper presents a proof-of-concept federated learning framework evaluated on a representative subset of five diverse buildings from the Building Data Genome Project 2 dataset.We integrate Federated Averaging (FedAvg), FedProx, climate-zone Clustered Federated Learning, and Personalized Federated Learning for next-hour HVAC load prediction, alongside a federated autoencoder for anomaly detection.Comparative analysis reveals that flat FedAvg is consistently the weakest prediction strategy [coefficient of variation of the root mean square error (CV (RMSE) = 9.99%)], while Clustered and Personalized Federated Learning (FL) recover most of the accuracy gap, with Personalized FL achieving the best overall performance (CV (RMSE) = 7.98%, R² = 0.861).For fault detection using synthetically injected fault signatures (sensor drift, stuck-meter, and spike patterns), the federated autoencoder achieves a superior F1-score (0.794 vs. 0.748) compared to a local-only detector.Notably, both detectors exhibit near-identical receiver operating characteristic-area under the curve (~0.972), indicating that threshold calibration rather than representational capacity drives the observed operating differences.This preliminary study underscores the efficacy of federated learning as a privacy-preserving tool, establishing a baseline for future large-scale validation across full building portfolios.