Distributed Optimal Scheduling of Integrated Energy Systems Based on Federated Reinforcement Learning

This paper addresses the economic optimization scheduling problem of integrated energy systems while taking the data privacy concerns into account. While traditional optimization methods can address scheduling problems in specific scenarios, they often struggle to adapt rapidly to the stochastic variations in loads and renewable energy sources. Data-driven deep reinforcement learning offers significant advantages in tackling this issue, but it involves handling a substantial amount of sensitive data. To address this challenge, this paper proposes a method that combines collaborative learning among multiple intelligent agents with privacy-preserving techniques to maximize system efficiency and protect user privacy. Initially, each intelligent agent undergoes local training, followed by the uploading of model parameters to a global server for privacy protection and federated averaging. Global average parameters are then distributed to individual agents for local model updates, with this process iterated multiple times until satisfactory performance is achieved. Experimental results demonstrate that this method achieves secure transmission and processing of user data through privacy protection measures, while achieving significant success in low-carbon economic dispatch. With the support of differential privacy mechanisms, the integrated energy system achieves an excellent balance between privacy protection and solution performance.