Resource Optimization of Space-Air-Ground Integrated Network Based on Digital Twin and Federated Multi-agent Reinforcement Learning

With the rapid advancement of sixth-generation wireless communication systems (6G), traditional fixed-infrastructure-based mobile edge computing (MEC) struggles to meet performance demands such as ultra-high reliability, ultra-low latency, and high data rates in scenarios with sparse communication facilities or sudden disasters. The space-air-ground integrated network (SAGIN), leveraging its extensive coverage capabilities and high communication capacity, offers a novel solution featuring global seamless coverage, low latency, and efficient energy utilization. This effectively addresses the limitations of MEC in large-scale, dynamic environments. Specifically, by integrating heterogeneous devices such as satellites, unmanned aerial vehicles (UAVs), and ground stations, SAGIN supports coverage across extremely vast areas, effectively compensating for the shortcomings of traditional terrestrial networks in remote regions and disaster scenarios. However, SAGIN’s high dynamism, device heterogeneity, and distributed deployment introduce complex resource scheduling and task allocation challenges. Against this backdrop, digital twin (DT) technology provides critical support for network optimization and management by delivering precise real-time mapping of physical systems. We designed and implemented a digital twin-based integrated air-ground-space network simulation platform (DT-SAG) to achieve collaborative modeling and intelligent scheduling of satellites, UAVs, and ground stations. To address the collaborative control and resource optimization challenges in SAGIN, we combined the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm with a federated learning (FL) framework. The MADDPG algorithm resolves the instability issues of traditional algorithms in complex multi-agent environments through centralized training and distributed execution. Meanwhile, FL reduces communication overhead, ensures data privacy, and enhances training efficiency by performing local training with global parameter aggregation. By integrating these techniques, the proposed FLMADDPG algorithm enables efficient collaboration among agents in SAGIN while avoiding the data privacy and communication bottlenecks associated with centralized learning. Experimental results demonstrate that the proposed method effectively enhances network coverage efficiency, energy management, and communication performance, showcasing the immense potential of integrating digital twins with federated learning.