Deep Learning-Based Channel Prediction With Outdated CSI for Underwater Acoustic Communications

Abstract

Channel state information (CSI) is essential for improving the performance of underwater acoustic (UWA) communications, which support applications such as ocean monitoring and resource exploration. However, owing to the substantial propagation delay of acoustic signals, the available CSI often becomes outdated, leading to temporal misalignment between the available and actual channel states. This letter presents a deep learning (DL)–based approach to predict the most recent CSI from outdated observations, enabling UWA transmissions to adapt to instantaneous channel variations. Specifically, a one-dimensional convolutional neural network is employed for time-domain cross-time CSI prediction, and a path switching (PS) mechanism is designed with the DL framework to capture the non-continuous temporal evolution of dominant multipath components, while also serving as a stochastic regularizer. Experimental results on real channel data demonstrate that the proposed method outperforms existing baselines in prediction error while maintaining low computational complexity. The source code of this study is available on https://github.com/cwangusc/CsiONet/