Federated learning-based non-intrusive load monitoring adaptive to real-world heterogeneities

Non-intrusive load monitoring (NILM) is a key way to cost-effectively acquire appliance-level information in advanced metering infrastructure (AMI). Recently, federated learning has enabled NILM to learn from decentralized meter data while preserving privacy. However, as real-world heterogeneities in electricity consumption data, local models, and AMI facilities cannot be eliminated in advance, federated learning-based NILM (FL-NILM) may underperform or even fail. Therefore, we propose a FL-NILM method adaptive to these heterogeneities. To fully leverage diverse electricity consumption data, dynamic clustering is integrated into cloud aggregation to hierarchically mitigate the global-local bias in knowledge required for NILM. Meanwhile, adaptive model initialization is applied in local training to balance biased global knowledge with local accumulated knowledge, enhancing the learning of heterogeneous data. To further handle heterogeneous local NILM models, homogeneous proxy models are used for global-local iteration through knowledge distillation. In addition, a weighted aggregation mechanism with a cache pool is designed for adapting to asynchronous iteration caused by heterogeneous AMI facilities. Experiments on public datasets show that the proposed method outperforms existing methods in both synchronous and asynchronous settings. The proposed method's advantages in computing and communication complexity are also discussed.