Transformer-Based Time-Domain Precoding Extrapolation for Massive MIMO

Precoding in massive multiple-input-multiple-output (mMIMO) systems relies on accurate estimation of downlink channel state information (CSI). However, when the number of antennas is large, obtaining CSI data incurs significant pilot and feedback overhead. In this paper, we propose a transformer based precoding network (TPN) that infers the future precoding matrices by exploring the time-frequency characteristic of historical wireless channels. Next, we design a suitable precoding matrix switching scheme based on channel correlation to further reduce the pilot overhead. Moreover, we leverage the network pruning technique to reduce the computational complexity of the proposed TPN. Simulations demonstrate that the sum-rate can be improved by 15% compared with the traditional zero-order holding method.