PX-25-C004 - Federated Learning-Based Energy Consumption Prediction for Office Building VRF Systems: A Feasibility Study for Privacy Protection ✓ Most Recent

Variable Refrigerant Flow (VRF) systems are widely adopted in modern office buildings due to their high energy efficiency and flexible control, accounting for nearly 50% of China’s central air-conditioning market. However, traditional energy consumption prediction models typically rely on centralized data sharing of building and equipment information, posing risks of privacy breaches and hindering cross-institutional collaboration. This study proposes a federated learning (FL) framework to achieve high-precision energy consumption prediction while preserving data privacy, utilizing localized model training and genetic algorithm-optimized global parameter aggregation. Experiments were conducted using real-world operational data from 518 VRF systems in Guangdong, China (covering 3,232 months), employing Artificial Neural Networks (ANN), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM) for model training. Genetic algorithms were introduced to optimize the aggregation weights of sub-models in federated learning. Results demonstrate that the global federated learning model achieves prediction accuracy comparable to centralized baseline models while effectively safeguarding user privacy. This study validates the potential of federated learning in constructing reliable energy consumption prediction models in distributed environments, offering novel insights for privacypreserving energy management platforms in smart buildings.