Scholarlink

In rapidly time-varying 5G communication environments, uplink-estimated channel state information (CSI) at the base station (BS) often becomes outdated due to channel aging, leading to degraded system performance. To address this issue, we propose DA-ENC, a novel Dual Attention-based (DA) Encoder architecture that utilizes historical CSI to accurately predict future downlink CSI. DA-ENC features a lightweight encoder-only structure enhanced by a DA mechanism and custom modules designed to capture the spatio-temporal and statistical characteristics of CSI. Extensive experiments demonstrate that DA-ENC consistently achieves the lowest normalized mean squared error (NMSE) across various user mobility patterns and signal-to-noise ratio (SNR) levels, exhibiting strong robustness to rapid channel fluctuations and noise. Under few-shot training with only 10% of the data, DA-ENC maintains excellent generalization. In zero-shot cross-scenario evaluations, it outperforms all baselines without requiring fine-tuning. Compared to the Transformer model, DA-ENC uses fewer parameters and achieves significantly lower inference latency (6.91 ms), making it highly suitable for real-time CSI prediction in next-generation wireless systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DA-ENC: A Dual Attention-Based Encoder Architecture for Channel Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Minghui Chen,
  • Zhiyong Luo

摘要

In rapidly time-varying 5G communication environments, uplink-estimated channel state information (CSI) at the base station (BS) often becomes outdated due to channel aging, leading to degraded system performance. To address this issue, we propose DA-ENC, a novel Dual Attention-based (DA) Encoder architecture that utilizes historical CSI to accurately predict future downlink CSI. DA-ENC features a lightweight encoder-only structure enhanced by a DA mechanism and custom modules designed to capture the spatio-temporal and statistical characteristics of CSI. Extensive experiments demonstrate that DA-ENC consistently achieves the lowest normalized mean squared error (NMSE) across various user mobility patterns and signal-to-noise ratio (SNR) levels, exhibiting strong robustness to rapid channel fluctuations and noise. Under few-shot training with only 10% of the data, DA-ENC maintains excellent generalization. In zero-shot cross-scenario evaluations, it outperforms all baselines without requiring fine-tuning. Compared to the Transformer model, DA-ENC uses fewer parameters and achieves significantly lower inference latency (6.91 ms), making it highly suitable for real-time CSI prediction in next-generation wireless systems.