Research On Load Balancing Routing Algorithms For Leo Satellite Constellation Broadband Networks

Low Earth Orbit(LEO)satellite broadband networks,as a crucial component of integrated space-ground networks,provide high-speed network access services to global users.Their advantage lies in extensive coverage,enabling users to access broadband networks at any time and any location.Due to the bursty nature of network traffic and the uniform distribution of low Earth orbit satellites,areas with high population density and substantial network traffic often experience link congestion.In such situations,the inability to fully utilize idle links may adversely affect data transmission efficiency.Therefore,researching efficient and reliable load-balancing routing algorithms for LEO satellite networks is of paramount importance.This thesis addresses the issue of load distribution imbalance in LEO satellite networks by investigating load-balancing routing algorithms,aiming to enhance network service performance.The main contributions of this thesis encompass the following three aspects:First,we establish the theoretical models required for LEO satellite constellation broadband networks.By analyzing the satellite network structure,we construct topology and inter-satellite visibility moels.The communication model takes into account the characteristics of the channel.The user model describes ground user distribution,host density,and request behavior.The cache queue model evaluates the processing capacity and congestion level of satellite nodes.The delay model incorporates queueing delay into link costs.These key models provide the theoretical basis for subsequent research on LEO satellite constellation broadband networks.Second,this thesis transforms the load balancing routing problem in LEO satellite constellation broadband networks into a partially observable Markov decision process.We propose a deep reinforcement learning-based load balancing routing algorithm(MA-RLR)that adopts the classic Actor-Critic framework from the reinforcement learning field to iteratively learn and optimize routing decisions for each intelligent agent in a distributed architecture.MA-RLR employs a packet-level routing scheme,offering flexibility to adjust routing policies in real-time in response to network state fluctuations.However,the need for neural network inference in every forwarding process may lead to additional computational overhead and decision latency,affecting the average maximum throughput of satellites.To address this issue,we further design a flow-based routing algorithm(MA-RLFR)that reduces the number of neural network inferences through a policy-sharing mechanism,thereby lowering the average decision latency for satellites.Due to the numerous and unpredictable factors affecting policy applicability,we introduce a delay threshold as a determining factor in the algorithm to enhance the practical value of MA-RLFR and the possibility of deploying it in real satellite network scenarios.Finally,we evaluate and compare the performance of the proposed algorithms with OSPF and ELB algorithms on a simulation test platform.The results show that the proposed algorithms outperform the other two algorithms in terms of average end-to-end delay,throughput,packet loss rate,and load balancing index.In the aspect of load balancing index,MA-RLR algorithm improves by 49.9% and 7.8% compared to OSPF and ELB algorithms,respectively.In terms of average satellite decision latency,the MA-RLFR algorithm achieves a 60.1% improvement over the MA-RLR algorithm.