Trajectory-Aware Intelligent Satellite Handover for Power Systems via Graph Neural Networks and Deep Reinforcement Learning

Abstract

Facing the power communication needs in remote areas lacking public network coverage, Low Earth Orbit (LEO) satellite networks are considered an important support for power communication due to their advantages of wide coverage, low latency, and high bandwidth. However, the high speed of the satellites leads to frequent link switching, which poses a challenge to communication continuity. This paper proposes an intelligent switching algorithm aimed at maximizing link availability time, minimizing switching frequency, and improving communication reliability. This algorithm uses Graph Neural Networks (GNNs) to model the topological structure and link status of low-Earth orbit satellite networks, constructing a dynamic graph that includes features such as inter-satellite visibility relationships, link quality (e.g., signal-to-noise ratio), and load, and extracting high-dimensional spatiotemporal state embeddings from it. The generated graph embeddings are used as state inputs for the reinforcement learning agent, enabling the agent to make link switching policy decisions based on comprehensive network topology and service status information. The algorithm integrates power system operational indicators such as telemetry cycle, alarm frequency, and load slope, and differentiates between the latency and bandwidth requirements of different services. The design employs a multi-objective optimization method to comprehensively optimize multiple indicators such as link availability time, number of handovers, and communication reliability. This method possesses online adaptive capabilities, enabling it to respond in real time to dynamic changes in network topology and service scenarios.