Machine Learning-Based Channel Prediction for Widely Distributed Massive MIMO with Real-World Data

Abstract

Widely distributed massive multiple input multiple output (WD-MIMO) systems are promising candidates for future mobile networks, given their improved energy efficiency, coverage and throughput. To spatially separate the users, WD-MIMO relies heavily on accurate and timely channel state information (CSI), which is hard to obtain in high mobility scenarios. To reduce the amount of pilot overhead necessary for obtaining CSI, we investigate linear and machine learning (ML)-based CSI prediction techniques and compare them in terms of achievable spectral efficiency (SE). The considered methods are constant continuation, Wiener prediction, dense, and long short term memory (LSTM) neural networks (NNs). Real-world data from a widely distributed massive MIMO channel measurement campaign with various base station (BS) antenna array aperture sizes is utilized for NN training and validation purposes. The capability of the considered CSI prediction methods to mitigate the effects of channel aging in realistic high-mobility scenarios is analyzed for different geometries of the massive MIMO BS antenna arrays. We can demonstrate a SE improvement of 2 bit/s/Hz for the LSTM NN compared to a Wiener predictor.