Non-intrusive load monitoring (NILM) is a computational technique to allow appliance-level energy disaggregation for sustainable energy management. Most NILM models require considerable training data to capture sufficient appliance signatures for robust model fitting. However, local on-site training cannot satisfy that requirement due to limited data availability. It is thus conceivable to perform data collaboration among different stakeholders. Unfortunately, current collaborative learning approaches rely on deep learning, encryption, and differential privacy techniques associated with either expensive computation or inefficient communication. In this paper, we propose a cost-effective collaborative learning framework, Fed-GBM (Federated Gradient Boosting Machines), consisting of two-stage voting and node-level parallelism, to address the problems in comodelling for NILM. Through extensive experiments on real-world residential datasets, Fed-GBM shows remarkable performance on convergence, accuracy, computation and communication efficiency. The impact of hyper-parameters in Fed-GBM is also extensively studied to guide better practical use.
CCS CONCEPTS• Computing methodologies → Machine learning; • Hardware → Power and energy.