An Asynchronous Model Parameter Update Method for Electric Power Horizontal Federated Learning Scenarios

In order to break the "sample data barriers" between power grid enterprises and external organizations under the premise of data security and compliance, this paper introduces a horizontal federated learning architecture with a central parameter server. Addressing the challenges of varying compute node performance and limited network bandwidth, which affect the training speed of power equipment defect detection models, an asynchronous model parameter updating method is proposed. This method innovatively combines an event-driven mechanism with gradient clipping. It accelerates high-power compute nodes' processing by reducing wait times during model aggregation and minimizes gradient upload size using an information-entropy-based approach. This effectively shortens model parameter transmission time under low bandwidth, enhancing the training speed of the defect detection model in the federated learning framework. Simulation results show that this method can reduce the model training time by 8% under diverse compute capabilities and unrestricted communication, and by 13% under compute heterogeneity and restricted communication conditions.