Cluster Head Federates for Distributed Energy Resources

Abstract

Accurate forecasting of energy demand is important for planning energy generation and distribution. For energy demand forecasting with distributed energy resources (DER), the central server typically collects data from individual nodes to train deep-learning based models resulting in more accurate predictions than models trained at the individual node level. This however result in concerns about data privacy. Federated learning provides an alternative to train the models collaboratively at distributed nodes without sharing the data.We present a novel Smart Cluster Head Client based Federated Learning (SCHC-FL) approach in which cluster-heads act as smart clients for contributing in federated learning. The open power system data (OPSD) with 54 nodes have been used to create a scenario of energy demand forecasting in distributed energy resources. We present the comparison of the proposed SCHC-FL approach with the existing federated learning and non-federated learning based approaches. Our proposed approach results in up to six times quicker training than other federated learning based approaches with comparable accuracy on convergence.