AI-Based Wireless Channel Prediction - Generalized Models for Adaptive Measurements and Predictions

Abstract

In wireless communication systems, channels can be time-varying, and their conditions can change due to factors such as mobility, interference, and environmentalconditions, which brings challenges in maintaining communication reliability. Hence, acquiring channel state information (CSI) is a critical step in physical layer of wireless communication. However, when accurate CSI or precoder fed back by a User Equipment (UE) is used by a network for MIMO precoding, the received CSI/precoder can get outdated quickly, consequently resulting in a loss of user and system throughput due to the utilization of outdated precoders. In a feature introduced in 3GPP Rel-18, a UE can be configured to measure multiple instances of the channel, use them to predict a number of channel/precoder instances in the future and report the predictions to the network. The performance of such a scheme depends highly on the ability to predict the channel with high accuracy. In this study, we explored AI-based methods for channel prediction to address this challenge. By utilizing a past window size of measured channel sequences, the proposed method forecasts future channel conditions. Out of all the AI models tested, the Transformer Encoder-Only model demonstrated superior performance, its counterparts and a classical non-AI based autoregressive (AR) model. We also found that, training the model with a diverse dataset with a mix of UE velocities, embedding across the time dimension within the Transformer EncoderOnly model, and constructing the model within the complex domain yields enhanced generalization capability. Furthermore, we developed a generalized model capable of using varying number of channel measurements to predict varying number of channel instances in the future.