Research On Dynamic Routing Algorithm For LEO Satellite Networks Based On Graph Neural Networks

LEO satellite networks have become an important part of communication network architecture due to their advantages of extensive coverage,independence from geographical and natural conditions,long communication distance,and enormous communication capacity.The rapid movement of satellites in their respective orbits causes rapid changes in topology,frequent disconnection and reconnection of intersatellite links in LEO satellite networks.At the same time,because to the satellite’s unique operating environment,its small size places severe limits on equipment such as batteries.As a result,one of the technical challenges for LEO satellite networks is developing effective and appropriate routing algorithms for issues such as time-variant network topology and limited battery power.In addition,graph neural networks have been one of the most attractive study areas in the field of deep learning in recent years,as they can fully exploit the topology of graph data to extract spatial information and generate lowdimensional representations.Furthermore,the high generalization capability of graph neural networks makes it appropriate for studying LEO satellite networks.In this thesis,the following research works are accomplished based on LEO satellite networks.1.A dynamic routing algorithm named MPNN-SNR based on message passing neural networks is proposed for LEO satellite networks to meet the characteristics of frequent topology changes and time-variant traffic load and other status.The algorithm uses message passing neural networks to learn the representations of LEO satellite network states,which are then fed into to the deep Q learning module to output the optimal routing policy.This thesis combines graph neural networks and deep reinforcement learning to tackle the routing problem in LEO satellite networks.The simulation results show that the algorithm can effectively increase the LEO satellite network’s performance in terms of delay,throughput,and packet drop rate,as well as to adapt to topology changes in the LEO satellite network.2.An energy-efficient routing algorithm named GATE-SNR based on graph attention is proposed for LEO satellite networks to deal with the issues of limited battery capacity and low energy utilization.First,a detailed description of the satellite energy model is given.Then,a novel graph attention model termed GATE is proposed to address the problem that the existing graph attention models have ignored the edge features,and a deep reinforcement learning algorithm named D3 QN is used to derive the optimal routing policy.Meanwhile,to improve the training efficiency,a valid action filter is proposed to filter out the invalid actions.Finally,the simulation results show that the algorithm can improve the performance of the LEO satellite network in terms of delay,packet drop rate and energy utilization,as well as has stronger generalization ability.3.Sat SIM,an LEO satellite networks simulation system built on Python and several open-source libraries,is designed to perform real-time and accurate simulations of LEO satellite network scenarios based on real satellite operational data,as well as to facilitate interaction with graph neural networks and deep reinforcement learning modules.