Optimized Federated Reinforcement Learning for Latency and Energy Minimization in Dynamic Satellite-Ground Graph Networks

Satellite–ground integrated networks (SGIN) represent a novel architectural paradigm that enables seamless spatiotemporal coverage, high-throughput data dissemination, and distributed computation offloading. By reducing end-to-end latency and alleviating resource contention, SGIN support efficient and adaptive task scheduling across heterogeneous network entities. However, the dynamic and time-varying topology of SGIN results in complex, multi-modal graph structures, posing significant challenges for unified representation, modeling, and optimization over space and time. To address these issues, we construct a graph-based model of the SGIN to jointly minimize total network latency and improve energy efficiency. A graph neural network (GNN)-based graph learning approach is developed to capture the dynamic features of satellite nodes. We further introduce a federated reinforcement learning framework to jointly optimize CPU cycle frequency, quantization bit allocation, and transmission power. This framework enhances convergence speed while preserving data privacy. Additionally, a transaction verification mechanism is integrated to illustrate the task execution process and demonstrate graph permutation invariance between the generative model and optimization objectives. Extensive simulations validate that the proposed algorithm achiveves about 20% lower latency and 25% lower energy consumption than baselines.