CSI Prediction With a Single AI Model for Heterogeneous Doppler Frequencies and Multipaths
Mehdi Meliha, Pascal Chargé, Salah Eddine Bouzid, Yide Wang, Christophe Henry, Yejian Chen · 2025
Accurate channel prediction is increasingly important, especially in scenarios where a user equipment (UE) operates at varying speeds, impacted by high Doppler shifts and channel dynamics. While AI-based techniques offer promising solutions, training a single model to handle different propagation paths and a wide range of UE velocities remains challenging. In this work, we present a CSI prediction technique that takes advantage of the inherent sparsity found in millimeter-wave and 5G systems. While this approach improves performance in dynamic environments, models trained at specific velocities often struggle to be generalized at different speeds. To overcome this limitation, we propose a Doppler downshifting technique, which serves as a unified AI model to enhance generalizability across various Doppler values, thereby reducing both model complexity and data requirements.