Adaptive Asynchronous Federated Learning for Digital Twin Driven Smart Grid
Zhuoqun Zhang, Haipeng Peng, Lixiang Li, Shuang Bao
Abstract
The smart grid represents a revolutionary advancement, ushering power systems into an era of enhanced intelligence. Unlike traditional power networks, smart grids demand superior real-time performance, security, and accuracy. However, their development faces significant challenges, including delays in device status updates, malicious station attacks, and a lack of user trust, which impede service quality improvements. To address these issues, this paper proposes a Privacy-Preserving Smart Grid distributed Collaborative computing system (PPSG) that integrates blockchain and asynchronous federated learning technologies. A digital twin framework tailored for smart grids is designed, enabling real-time simulation and reflection of electrical devices’ states and behaviors, thereby enhancing service responsiveness. Additionally, to tackle non-independent and identically distributed (Non-IID) data and outdated local models, a dual dynamic aggregation factor asynchronous federated learning scheme is introduced, improving service accuracy. The "Proof of Contribution" blockchain consensus algorithm is employed to assess contributions to computational tasks and utilize stochastic processes to mitigate election fraud, thereby strengthening security. Extensive comparative experiments on Non-IID datasets and heterogeneous devices demonstrate PPSG’s superior learning performance, efficiency, and reliability. Furthermore, experiments using real power grid datasets validate its practical applicability, scalability in large-scale node environments, and feasibility for real-world deployment.