A federated learning approach to personalized prosumer recommendations in peer-to-peer energy networks

Decentralized Peer-to-Peer (P2P) energy-sharing systems become increasingly prevalent and there is a growing need for robust, privacy-preserving mechanisms that allow prosumers to express and manage their individual energy-sharing preferences. This paper presents a federated framework for modeling and addressing prosumers’ preferences in such systems. Conventional centralized approaches often fail to provide the required personalization, adaptability, and preference sensitivity, while also introducing significant privacy risks. Our proposed framework employs a three-layer federated learning (FL) architecture as its core that integrates key components such as preference-based clustering, similarity-weighted aggregation, and representative model distillation to handle prosumer preferences heterogeneity across distributed prosumers and enable personalized supplier recommendations without requiring direct data exchange. Extensive experiments demonstrate that the proposed approach outperforms standard FL baselines, achieving up to 35% improvement in preference alignment and 10% increase in normalized discounted cumulative gain (NDCG), along with stable scalarized rewards. Computational analyses further validate the framework’s scalability, which demonstrates its potential for large-scale implementation.