Transformer-Based CSI Prediction for 3GPP MIMO Codebooks

Abstract

Accurate channel state information (CSI) acquisition is essential for multiple-input multiple-output (MIMO) systems. The Third Generation Partnership Project (3GPP) specifies a limited feedback framework in which the user equipment reports CSI to the base station. In medium-to-high mobility scenarios, CSI becomes outdated due to channel time variations and processing delays. To address this issue, 3GPP Release 18 introduces the enhanced Type-II (eType-II) Doppler code-book, which incorporates CSI prediction. This paper investigates a Transformer-based AI/ML approach for CSI prediction in the angle-delay domain, targeting precoding with the eType-II Doppler codebook. A 3GPP-compliant fast fading channel model is considered together with standardized CSI reference signal structures. Simulation results demonstrate that the proposed approach outperforms autoregressive and non-predictive schemes in terms of both intermediate metrics and spectral efficiency.