Deep Learning Techniques for Channel Prediction in Non-terrestrial Networks

Non-Terrestrial Networks (NTNs) are characterized by an inherently dynamic propagation environment, resulting from the orbital movement of communication satellites. Depending on the orbital altitude, propagation effects, such as Doppler shift and propagation losses, can be characterized by different statistics and temporal variations as a communication satellite orbits over a coverage area. Nonetheless, such trends result from the geometrical relationship between the orbiting satellite and the user equipment on ground. As such, they can be partly accounted for by means of channel prediction techniques. With the advent of deep learning, it has been shown that specific neural network architectures can serve the purpose of performing time series forecasts. This has also been recognized in the field of NTN, not only from the point of view of the scientific literature, but also from that of the cellular standardization in 3rd Generation Partnership Project (3GPP) standards, with channel prediction being identified as one of the main use cases for the application of artificial intelligence in the radio access network. Thus, motivated by these considerations, in this chapter we present deep learning approaches to channel prediction in NTNs. After discussing the characteristics of the NTN channel, we introduce the typical deep learning architectures and use cases for channel prediction in NTN. In particular, we focus our analysis on two applications: firstly, we employ channel prediction as a mean to reduce the amount of pilot transmissions required for symbol demodulation; secondly, we implement a predictor to counteract channel aging in a Cell-Free Multiple Input Multiple Output (CF-MIMO) NTN system. For both applications, we also discuss computational complexity aspects. We conclude the chapter with a discussion on generalization vs adaptability in channel prediction in the context of NTNs.