CNN and LSTM Neural Networks for spectral efficiency improvements in LEO Satellite networks

Providing stable, efficient, and reliable communication methods is important in modern technological development and implementation. LEO satellite networks are one of the communication methods that have risen in popularity due to their ability to cover remote areas. Optimization of LEO satellite usage can, however, be a demanding task. In this paper, various neural network (NN) architectures are tested to see which ones can provide SNR prediction that can improve spectral efficiency in simulated channels with various shadowing levels, Doppler frequency shifts, and expected SNRs. The simulator of shadowed Rice fading was used to create a total of 90 different channels, and a total of 4 different NN architectures were implemented. Two different convolutional and two different long-short-term-memory (LSTM) NN architectures were used to predict the future SNR values in the channels. These predictions were combined with an estimated margin to select the best adaptive coding and modulation employment following the Digital Video Broadcasting Satellite Second Generation (DVB-S2X) satellite protocol. The simulation results showed that the proposed networks provided a stable improvement over the baseline outdated information approach, and the simplest NN was outperformed by more complex ones in most cases (maximum difference of 0.15 b/s/Hz). The paper shows how different NN improvements occur in different channels and indicates that utilizing more resources to support a more complex NN implementation could be beneficial, based on the available computational resources and the need for spectral efficiency improvement.