Privacy-preserving knowledge sharing for few-shot building energy prediction: A federated learning approach

The data-driven method is a promising way to predict the energy consumption of buildings, however suffering from the data shortage problem in various scenarios. Even though transfer learning can improve the few-shot prediction performance by utilizing other buildings' data, the centralized approach poses potential privacy risks. To tackle this issue, the paper proposes a privacy-preserving knowledge sharing framework to facilitate the few-shot building energy prediction based on federated learning. First, a private data aggregation scheme is established to encrypt the sensitive data with shared random masks and guarantee the privacy of the data preprocessing and model optimization. Then, to alleviate the intrinsic data heterogeneity, a dynamical clustering federated learning algorithm is proposed to implement the intra-cluster and inter-cluster knowledge sharing along with the iterative clustering process for participating buildings. Finally, the network-based transfer learning approach is incorporated into the distributed framework to establish the customized model based on trained cluster models and further boost the prediction performance for each building. Extensive experiments on the Building Data Genome Project 2 (BDGP2) dataset indicate that the federated approach witnesses a desirable prediction performance while preserving the privacy of building occupants.