Machine Learning-Based Channel Prediction with Reduced Training Overhead for Massive MIMO-OFDM Systems

Channel prediction addresses outdated channel state information by forecasting future channels based on past channel estimates. We propose a machine learning (ML)-based approach using neural networks to learn complex temporal statistics. Unlike conventional offline-trained predictors that suffer from unfamiliar environments, our online re-training framework adapts to varying channel conditions by re-training the networks from scratch. To minimize the re-training time for practical implementation, we introduce an aggregated learning (AL) approach for massive multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) systems. AL splits and aggregates training data in array or frequency domains of MIMO-OFDM channels, significantly reducing data collection time. Numerical results show that AL not only decreases training time overhead but also improves prediction performance across various scenarios.