Lightweight TCN-Based Spatial-Temporal Channel Extrapolation for RIS-Aided Communication

Reconfigurable intelligent surface (RIS) has become a promising technology to enhance the coverage for the millimeter-wave communications. In this paper, we consider the acquisition of time-varying cascaded channels. Generally, this requires huge pilot overhead due to the high dimensional channel structure and the short channel coherence time. To address the above problem, we design a spatial-temporal channel extrapolation scheme by employing deep learning algorithms. Specifically, the temporal convolutional network (TCN) is utilized to explore the dynamic nature of the cascaded channels and to realize the channel prediction over the time domain. Besides, the neural ordinary differential equations are adopted to implement the spatial extrapolation. Moreover, to avoid massive computation caused by a large number of parameters, we achieve a lightweight network through the structured probabilistic pruning. Simulation results show the effectiveness of the designed deep learning-based channel extrapolation scheme.