A GNN–DRL–ResNet-Based Dynamic Routing Algorithm for Low Earth Orbit Satellite Networks

Abstract

There are a lot of nodes in a Low Earth Orbit (LEO) communication network, and their resource limits might change quickly. Because of these features, conventional routing techniques are not a good fit for LEO satellite networks. An inductive learning architecture called Graph Neural Network (GNNs) is proposed in this research to tackle this issue. The number of topological nodes that require training can be drastically decreased with the help of the suggested architecture. Because of this cut, the nodes' computational difficulty is reduced. In addition, routing methods are optimized and learned continuously using deep reinforcement learning (DRL), which is made even more generalizable by building the DRL agent with GNN. Every Deep Q-Network (DQN) agent manages its own tasks in the suggested algorithm for optimizing LEO satellite route planning based on graph neural networks which does not extracts the spatial and temporal information of networks that’s why we planned to propose RESNET model to replace DQN. By using the GNN paradigm, it discovers the nodes' concealed states. To determine the best routes, the DRLmodel takes these hidden states into account. A comparison and simulation were run to assess the algorithm's efficiency. Finally, optimization technique is presented to choose the shortest route. The outcomes demonstrate that the suggested method mitigates average overall latency while simultaneously increasing total network throughput. When compared to the Deep Q Networks (DQNs) and Dijikstras Algorithm, the suggested approach achieves a 25% and 30% improvement in average throughput, respectively. Additionally, it can adjust to different topologies and lower average end-to-end latency by 44% and 22%, respectively.