Personalized federated learning for cross-building energy knowledge sharing: Cost-effective strategies and model architectures

Sufficient building operational data serve as the key premise to enable the development of reliable data-driven technologies for building energy management. Considering that individual buildings may suffer from various data scarcity problems, it is highly promising to devise cross-building data and knowledge sharing methods as possible solutions. Assuming homogeneity among individual building datasets, federated learning can be readily implemented to break the data silos among buildings without violating privacy and security compliances. Nevertheless, individual buildings typically present different operation and energy patterns, making their data heterogeneous and challenging for standard federated learning. To solve such problems, this study explores the domain of personalized federated learning to effectively utilize multi-source building operational data for collaborative data analyses. Three customized personalization strategies have been proposed for building energy prediction tasks and a novel transformer layer has been devised to facilitate tabular data modeling with minimal computational costs. Comprehensive data experiments have been designed to quantify the value of personalized federated learning in 10 building communities with varying building types and scales. The research results show that federated learning can reduce building energy predictions errors by 5.0% to 22.0% over localized solutions, while personalization strategies can further boost the performance by 3.4% to 15.9% given different training data availabilities. The research outcomes are helpful in achieving efficient cross-building data integration and knowledge sharing for district-level building energy management.