Author: Beomsoo Ko, Hwanjin Kim, and Junil Choi
Conference and Year: ICTC 2021
In this paper, we focus on channel prediction for massive MIMO wideband systems. In general, precise channel state information (CSI) is essential to fully utilize the benefits of the large array of antennas. However, the CSI is likely to be out-dated due to the feedback delay between a base station and a user equipment. Hence, the channel prediction is required to enhance the performance of the system.
The wideband channel is transformed in to multiple parallel orthogonal-frequency-division-multiplexing (OFDM) channels. In the conventional machine learning-based channel prediction studies, a neural network (NN) is generated for each sub-carrier channel (sub-channel), which results in large training overhead. Hence, we propose a channel predictor, which only generates a single NN for every sub-channel to decrease the training overhead.
Since the OFDM sub-channels are highly correlated, a single NN trained with a single OFDM sub-channel is sufficient to represent the channel statistics of every sub-channels. In the simulation, we showed that the normalized mean square error (NMSE) performance of the proposed channel prediction has almost the same performance with the conventional channel predictor while reducing the training overhead significantly.
