Towards A Peer-To-Peer Residential Short-Term Load Forecasting With Federated Learning

The inclusion of intermittent and renewable energy sources has increased the importance of demand forecasting in the power systems. Smart meters play a critical role in modern load forecasting due to the high granularity of the measurement data. Federated Learning can enable accurate residential load forecasting in a distributed manner. In this regard, to compensate for the variability of households, clustering them in groups with similar patterns can lead to more accurate forecasts. Usually, clustering requires a central server that has access to the entire dataset, which collides with the decentralized nature of federated learning. In order to complement federated learning, this study proposes a decentralized peer-to-peer strategy that employs agent-based modeling. We evaluate it in comparison to a typical centralized k-means clustering. To create clusters, we compare Euclidian and Dynamic time warping distances. We employ these clusters to build short-term load forecasting models using federated learning. Our results reveal the possibility of using peer-to-peer (P2P) clustering along with simple Euclidean distances and Federated learning (FL) to obtain highly performant load forecasting models in a fully decentralized manner.