Transformer-Based CSI Prediction for Pinching Antenna Systems in Mobile Communication System

Abstract

Pinching antenna systems (PASS) have attracted growing interest as a new class of flexible antennas due to their ability to dynamically construct line-of-sight paths. These characteristics make them highly promising for complex mobile communication environments, particularly in high-speed railway (HSR) or highway tunnel scenarios where communication performance often degrades. In high-mobility scenarios, accurate channel state information (CSI) acquisition is critical but severely hindered by channel aging and rapid temporal variations. Under limited observation windows and geometric information, the inherently ill-conditioned characteristics of PASS channels further complicate channel acquisition. To improve CSI accuracy without the prohibitive pilot overhead of frequent estimation, we develop a Transformer-based channel prediction network tailored for PASS in high-mobility scenarios. Furthermore, we develop an enhanced network architecture that exploits implicit user position information during the training phase to achieve higher prediction accuracy, while maintaining a low-complexity inference phase that operates without requiring real-time position data. Numerical results demonstrate that the proposed frameworks achieve robust CSI prediction using only received pilot signals. Incorporating positioning as an auxiliary task further mitigates channel aging and improves prediction performance, thereby supporting the practical deployment of PASS in high-mobility scenarios.