Communication Efficiency of Asynchronous Federated Learning with Proportional Fairness in Smart Substations

• Yuewei Tian,
• Yujia Wang,
• Lisa Guo,
• Yang Su,
• Mengru Zuo,
• Lintao Shi,
• Zhonggui Ma,

Vol. 20, No. 3, March 31, 2026
10.3837/tiis.2026.03.010, Download Paper (Free):

• Asynchronous Federated Learning
• communication efficiency
• Industrial IoT
• Intelligent Monitoring Systems
• Substations

Abstract

This paper investigates the communication efficiency of asynchronous federated learning (AFL) with proportional fairness in smart substations. To address the challenges of asynchronous client-server interactions and imbalanced data distribution, we propose a proportional fairness-based AFL method in which a dual-weight adjustment strategy is introduced. Specifically, intelligent monitoring terminals in smart substations serve as distributed clients that collaborate asynchronously with a central server to jointly train a global monitoring model. The dual-weight adjustment strategy dynamically allocates client contributions to global model updates through two key parameters: sample weight reflecting local data volume and delay weight quantifying network latency. The delay weight is derived from a novel hybrid model combining transmission delay calculated via Shannon's theorem and training delay based on CPU cycle analysis, enabling adaptive staleness compensation. Theoretical analysis demonstrates that this method prioritizes clients with larger datasets and lower communication latency, thereby optimizing both communication efficiency and model performance. Simulation results demonstrate that the proposed method outperforms conventional AFL methods by achieving 53.3% faster convergence and 11.5% higher accuracy.

Statistics

Cite this article