Massive MIMO Channel Prediction in Real Propagation Environments Using Tensor Decomposition and Autoregressive Models

Massive multiple-input multiple-output (MIMO) has been widely deployed in practice due to its capability of improving spectral efficiency. In mobility scenarios where users move at relatively high velocities (e.g., 40 km/h), however, massive MIMO faces a severe problem called “channel aging”. This problem leads to performance degradation as outdated and inaccurate channel state information (CSI) is used in massive MIMO beamforming. In the paper, we study time-variant massive MIMO channel characteristics in real propagation environments, and propose channel prediction schemes using tensor decomposition and autoregressive (AR) models to combat channel aging. Specifically, a multi-dimensional massive MIMO channel can be regarded as a higher-order tensor and modeled by the canonical polyadic decomposition (CPD). By applying the CPD, channel components that vary relatively slowly in time are extracted, and their time variations are captured and predicted using AR models. Based on channel data measured at the 2.6 GHz frequency band, we evaluate the prediction accuracy of the CPD-AR scheme and the corresponding performance using zero-forcing (ZF) beamforming. For comparison, a scheme using the multi-dimensional discrete Fourier transform (DFT) and AR prediction is also presented. Simulation results show that both schemes can reduce CSI error by 3–10 dB as compared to the aged CSI, resulting in significant improvements of ZF sum-rates. The research indicates that massive MIMO channels are to a large extent predictable in both line-of-sight (LOS) and non-line-of-sight (NLOS) scenarios, and the proposed prediction schemes can effectively alleviate channel aging when users move in relatively high speeds.