Short-Term Prediction of Doubly-Dispersive Channels for Pulse-Shaped OTFS using 2D-ConvLSTM

Abstract

In this paper, we investigate the ability of recurrent neural networks to perform channel predictions for orthogonal time frequency and space modulation (OTFS). Due to 2D orthogonal precoding, OTFS promises high time-frequency (TF) diversity which turns out to enable robust communication even in high mobility scenarios. To exploit high diversity gain, knowledge of accurate channel state information (CSI) is essential. In OTFS, the CSI can directly be estimated in the delay-Doppler (DD) domain. Vehicular channels however are considered to be doubly-dispersive and therefore require a channel estimation on a per frame basis. This motivates the investigation of short-term channel prediction. We propose a scheme to estimate the channel coefficients collected on vehicular trajectory and predict them into the future using 2D-convolutional long short-term memory network (2D-ConvLSTM). First numerical results show that a prediction of the channel coefficients is possible.