Deep Reinforcement Learning Based Load Balancing Routing for LEO Satellite Network

Abstract

LEO satellite network (LEO-SN) constitutes an indispensable part of the space-air-ground integrated network (SAGIN). However, the characteristics of unbalanced traffic load, intricate orbits, high mobility, frequent link handovers, diverse storage and communication capabilities of nodes gradually lead to the ineffectiveness of the central routing planning methods. In order to enhance the flexibility and agility of the planning process, this paper presents an intelligent decentralized load balancing routing algorithm using deep reinforcement learning for the LEO-SN, which takes into consideration the queuing delay, storage space, communication bandwidth and propagation delay of merely one-hop satellite node. Simulation and analysis demonstrate that the proposed load balancing routing algorithm could converge quickly for both training and test sets, and possesses better performance than the comparison methods in terms of both the packet loss rate and transmission latency.