Improved Downlink Rates for FDD Massive MIMO Systems Through Bayesian Neural Networks-Based Channel Prediction

Abstract

In frequency-division-duplex (FDD) massive MIMO systems, channel state information (CSI) feedback waiting phase does not get fully exploited since base station needs to wait for the CSI feedback before transmitting downlink data. The proportion of the CSI feedback waiting phase during the downlink transmission would be high as the MIMO system scales up, which sacrifices downlink rates of the FDD massive MIMO systems significantly. In this paper, we first present a channel prediction-aided FDD scheme to utilize the idle waiting time efficiently. Then we propose a novel channel prediction method based on Bayesian neural network (BNN), which can handle the uncertainty in a natural manner and learn regularization from data without painstaking manual pre-tuning of network hyperparameters. Numerical results show that our proposed channel prediction-aided FDD scheme can achieve remarkable performance gains in terms of either achievable downlink rates or bit error rate. Moreover, our proposed BNN-based channel predictor is much more effective and robust in contrast to the state-of-the-art channel prediction techniques such as autoregressive model and recurrent neural network.