Recurrent Neural Network-Based Frequency-Domain Channel Prediction for Wideband Communications

Abstract

Outdated channel state information (CSI) severely degrades the performance of adaptive transmission systems that adapt their transmissions to channel fading. In contrast with mitigation methods that sacrifice scarce wireless resources to compensate for such a performance loss, channel prediction provides an efficient solution. A few predictors for frequency-flat channels were by far proposed, whereas those suited to frequency-selective channels are seldom explored. In this paper, therefore, we propose to apply a recurrent neural network to build a frequency-domain channel predictor for wideband communications. As an application example, integrating a predictor into a multi-input multi-output orthogonal frequency- division multiplexing system to improve the correctness of antenna selection is provided. Performance assessment is carried out in multi-path fading channels defined by 3GPP Extended Vehicular A and Extended Typical Urban models. Results reveal that this predictor is effective to combat the outdated CSI with reasonable computational complexity. It outperforms the Kalman filter-based predictor notably and has intrinsic flexibility to enable multi-step prediction.