Meta-Learning for Graph Neural Network-Based Power Allocation in LEO Satellite Communications

In this work, we develop a power allocation policy for low-earth orbit (LEO) satellite communications with uneven traffic distribution. The inter-beam interference is one of the bottlenecks for achieving high data rates. To manage interference, we represent the power allocation policy by a graph neural network (GNN) that is scalable to the number of beams of the satellite. The domain knowledge in satellite communications is exploited to design the structure and hyper-parameters of the GNN. To improve the generalization ability of the GNN in scenarios with different traffic distributions, we train the GNN in an unsupervised manner by using meta-learning. The idea is to train the GNN with different traffic distributions and then test it in unseen scenarios. Our simulation results show that the data rate achieved by the GNN-based power allocation policy is 30% higher than existing power allocation policies and is close to the upper bound. The GNN outperforms the fully-connected neural network in terms of scalability. Finally, we demonstrate that with meta-learning, the GNN does not need samples in unseen scenarios to achieve high data rates.