Massive MIMO Channel Prediction: Machine Learning Versus Kalman Filtering

Abstract

This paper addresses a channel prediction problem for massive multiple-input multiple-output (MIMO) systems. Previous channel prediction methods exploited theoretical channel models, which deviate from realistic channels. In this paper, we develop and compare a machine learning (ML)-based channel predictor and a vector Kalman filter (VKF)-based channel predictor using the spatial channel model (SCM), which is widely used in the 3GPP standard. The VKF-based channel predictor is first developed based on the autoregressive (AR) model. Then, the ML-based channel predictor is developed using linear minimum mean square error (LMMSE) pre-processed channel data. Numerical results show that both channel predictors have substantial gain over the outdated channel in terms of channel prediction accuracy and data rates. The total computational complexity of the ML-based predictor is higher than that of the VKF-based predictor, but once trained, the ML-based predictor has lower complexity than the VKF-based predictor.