Deep Learning-Based Channel Prediction With Path Extraction

Abstract

Deep Learning (DL)-based channel prediction has emerged as a complementary solution to Channel State Information (CSI) obsolescence in the context of beyond 5G networks. For this purpose, in this letter, a CSI prediction method is developed that leverages the sparsity inherent in millimeter-wave and 5G systems. The proposed approach is capable of isolating significant paths of the channel and a single DL model is implemented to make a prediction for each path individually. Numerical results reveal that this method offers a robust solution, achieving up to 60% higher accuracy and up to 80% reduction in computational load compared to state-of-the-art techniques, while also reducing the size of DL models and required datasets.