Hybrid Neural Network-Based Fading Channel Prediction for Link Adaptation

We propose an artificial intelligence-based channel prediction scheme that can potentially facilitate link adaptation in customized communication systems. Link adaptation is a key process for wireless communication that requires accurate channel state information (CSI). However, the CSI may be outdated because of computational and propagation delays. In addition, a subframe with no CSI reference signal cannot provide CSI feedback. The proposed scheme solves these problems by predicting future channels. Although traditional stochastic methods suffer from marginal prediction accuracy or unacceptable computational complexity, neural networks allow time series prediction for channels even considering constraints for practical application. We introduce a hybrid architecture for improving the prediction accuracy of the neural network when extracting meaningful features. The proposed scheme uses a single hybrid network that can predict channels in different environments. Simulations were performed using a spatial channel model to evaluate the performance at the system-level, and the results indicated that the proposed scheme effectively increases the prediction accuracy for the channel quality indicator and spectral efficiency