Performance Evaluation of Federated Learning for Residential Energy Forecasting

Abstract

Short-term energy-consumption forecasting plays an important role in the planning of energy production, transportation and distribution. With the widespread adoption of decentralised self-generating energy systems in residential communities, short-term load forecasting is expected to be performed in a distributed manner to preserve privacy and ensure timely feedback to perform reconfiguration of the distribution network. In this context, edge computing is expected to be an enabling technology to ensure decentralized data collection, management, processing and delivery. At the same time, federated learning is an emerging paradigm that fits naturally in such an edge-computing environment, providing an AI-powered and privacy-preserving solution for time-series forecasting. In this paper, we present a performance evaluation of different federated-learning configurations resulting in different privacy levels to the forecast residential energy consumption with data collected by real smart meters. To this aim, different experiments are run using Flower (a popular federated learning framework) and real energy consumption data. Our results allow us to demonstrate the feasibility of such an approach and to study the trade-off between data privacy and the accuracy of the prediction, which characterizes the quality of service of the system for the final users.