In-orbit channel prediction: deep learning architectures for 6G non-terrestrial networks

Abstract Wireless communications rely on the transmission of pilot sequences to estimate and compensate the effects of the propagation channel on the received signals. This is particularly important in non-geostationary non-terrestrial networks (NTNs) due to the inherent mobility of one end of the communication system, resulting in severe impairments to orthogonal frequency division multiplexing (OFDM) transmissions. To reduce the pilot overhead, regular and pilot-less OFDM slots are alternated, with a channel prediction algorithm being implemented at the receiver to equalize received data in absence of pilots based on the most recent channel estimates. Three prediction algorithm architectures based on convolutional neural networks are presented, comparing their prediction mean squared error and computational complexity. Based on these results, the best performing model is selected and evaluated in a complete NTN link. With the chosen model, a user throughput gain of 8.33% can be achieved under various propagation conditions, including fading models and Doppler spreads unseen during training. The considered neural network only requires 138k multiply-and-accumulate units, hinting at the possibility of real-time inference under strict power constraints.