Edge-End Cooperative Network Resource Allocation With Time Synchronization Awareness for Federated Learning-Based Distributed Energy Regulation

Abstract

Distributed energy regulation is essential for carbon reduction and neutralization of smart park, and relies heavily on distributed energy regulation model training based on federated learning (FL). Network resource allocation is vital to reduction of model training loss and delay. However, time synchronization offset as well as adversarial competition in resource allocation cause low training precision and high training delay. Therefore, we investigate the edge-end cooperative network resource allocation problem for low-delay and high-precision FL, considering the long-term time synchronization offset constraint. We propose a penalty dueling deep Q network (PDDQN)-based edge-end cooperative network resource allocation algorithm with time synchronization awareness, named TARGET, to solve the formulated problem. TARGET achieves joint optimization of adversarial routing selection and device scheduling through the combination of PDDQN and DQN. TARGET provides model training precision guarantee by considering both long-term and short-term constraints of time synchronization offset. Simulation results indicate TARGET achieves superior performance in global loss function, time synchronization offset, and distributed energy regulation.