Scholarlink

<p>Satellite clock bias (SCB) is a crucial factor that impacts the accuracy of real-time precise point positioning (RT-PPP). However, in practical scenarios, considering how to ensure positioning performance under real-time service interruptions is often necessary, making developing a high-precision SCB prediction model essential. Aiming at the complex time–frequency coupling characteristics and substantial noise interference exhibited by SCB of the third-generation BeiDou Navigation Satellite System, this paper proposes the MGAMNet, a deep learning model integrating multi-granularity perception with hierarchical hybrid mechanisms. First, a multi-granularity patch (MGP) module is designed to construct multi-granular subsequences for enhancing the modeling of multi-frequency disturbance characteristics in SCB periodic terms and local semantic structures, combined with a multi-granularity mixing (MGM) module to realize cross-granularity information fusion. Attention mechanisms are then utilized to model long-term and short-term dependencies. Experimental results demonstrate that MGAMNet significantly outperforms four mainstream models—LSTM, GRU, LSTM-Attention, and Informer—in 1&#xa0;h, 12&#xa0;h, and 24&#xa0;h prediction tasks, achieving the best performance across five evaluation metrics. In short-term prediction (1&#xa0;h), the RMSE of MGAMNet is reduced to 0.06&#xa0;ns, with an average performance improvement of over 45.2% (with improvements of up to 55.2%) compared to the aforementioned models. For 12&#xa0;h medium-term prediction, the RMSE is further controlled at 0.46 ns, with accuracy enhancements ranging from 44.8 to 66.1% relative to each comparative model. In the more challenging 24&#xa0;h long-term prediction task, MGAMNet still maintains a stable advantage, with an RMSE of 0.87&#xa0;ns and a reduction in error by 42.4–66.2%.</p>

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

MGAMNet: a multi-granularity aware and hierarchically mixed network for BDS-3 satellite clock bias prediction

  • 期刊论文

  • 发表日期：

    2025年10月27日

查看全文
  • Zhanpeng Cao,
  • Zhimin Sha,
  • Dong Lv,
  • Pengzhi Wei,
  • Bowen Xiong,
  • Shirong Ye

摘要

Satellite clock bias (SCB) is a crucial factor that impacts the accuracy of real-time precise point positioning (RT-PPP). However, in practical scenarios, considering how to ensure positioning performance under real-time service interruptions is often necessary, making developing a high-precision SCB prediction model essential. Aiming at the complex time–frequency coupling characteristics and substantial noise interference exhibited by SCB of the third-generation BeiDou Navigation Satellite System, this paper proposes the MGAMNet, a deep learning model integrating multi-granularity perception with hierarchical hybrid mechanisms. First, a multi-granularity patch (MGP) module is designed to construct multi-granular subsequences for enhancing the modeling of multi-frequency disturbance characteristics in SCB periodic terms and local semantic structures, combined with a multi-granularity mixing (MGM) module to realize cross-granularity information fusion. Attention mechanisms are then utilized to model long-term and short-term dependencies. Experimental results demonstrate that MGAMNet significantly outperforms four mainstream models—LSTM, GRU, LSTM-Attention, and Informer—in 1 h, 12 h, and 24 h prediction tasks, achieving the best performance across five evaluation metrics. In short-term prediction (1 h), the RMSE of MGAMNet is reduced to 0.06 ns, with an average performance improvement of over 45.2% (with improvements of up to 55.2%) compared to the aforementioned models. For 12 h medium-term prediction, the RMSE is further controlled at 0.46 ns, with accuracy enhancements ranging from 44.8 to 66.1% relative to each comparative model. In the more challenging 24 h long-term prediction task, MGAMNet still maintains a stable advantage, with an RMSE of 0.87 ns and a reduction in error by 42.4–66.2%.