测绘学报 ›› 2024, Vol. 53 ›› Issue (1): 65-78.doi: 10.11947/j.AGCS.2024.20230082

• 大地测量学与导航 • 上一篇    下一篇

基于BiLSTM模型的BDS-3短期钟差预报精度研究

潘雄1, 黄伟凯1, 赵万卓1, 张思莹1, 张龙杰1, 金丽宏2   

  1. 1. 武汉纺织大学计算机与人工智能学院, 湖北 武汉 430200;
    2. 武汉纺织大学数理科学学院, 湖北 武汉 430200
  • 收稿日期:2023-04-07 修回日期:2023-11-27 发布日期:2024-02-06
  • 通讯作者: 金丽宏 E-mail:2022018@wtu.edu.cn
  • 作者简介:潘雄(1973-),男,博士,教授,研究方向为深度学习、卫星导航定位。E-mail:pxjlh@163.com
  • 基金资助:
    国家自然科学基金(42174010;41874009);湖北省自然科学基金(2023AFB435)

Research on short-term prediction accuracy of BDS-3 clock bias based on BiLSTM model

PAN Xiong1, HUANG Weikai1, ZHAO Wanzhuo1, ZHANG Siying1, ZHANG Longjie1, JIN Lihong2   

  1. 1. School of Computer Science and Artificial Intelligence, Wuhan Textile University, Wuhan 430200, China;
    2. School of Mathematical and Physical Sciences, Wuhan Textile University, Wuhan 430200, China
  • Received:2023-04-07 Revised:2023-11-27 Published:2024-02-06
  • Supported by:
    The National Natural Science Foundation of China(Nos. 42174010; 41874009); The Hubei Province Natural Science Foundation(No. 2023AFB435)

摘要: 提出了一种改进的北斗钟差预测模型,将传统的单向长短期记忆神经网络(LSTM)扩展为双向长短期记忆网络(BiLSTM),引入了3种自适应匹配超参数的算法提高钟差数据短期预报的精度。首先,对LSTM进行优化,建立BiLSTM模型,介绍了超参数的3种选择方案(粒子群搜索(PSO)、麻雀搜索(SSA)和贝叶斯搜索(BOA)),并给出了相应的适用范围。然后,详细介绍基于超参数优化BiLSTM模型的钟差预报的步骤。最后,利用GFZ卫星钟差数据,从不同轨道类型、5 min采样间隔、15 min采样间隔等方面进行了1、6和12 h的单天和多天预报对比试验,并进行了相应模型的时间复杂度分析。试验结果表明,采用超参数方案优化后的BiLSTM模型在进行1、6和12 h预报时,相较于二次多项式模型、灰色模型、长短期记忆神经网络的模型和BiLSTM模型,平均精度可分别提升86.21%、83.32%、69.99%和55.17%。在3种优化方案中,使用PSO算法对IGSO类型卫星的优化效果较好;使用BOA算法对MEO类型卫星的钟差优化效果较好;使用SSA算法在整体上优化效果最好。虽然经过超参数优化后的BiLSTM模型训练时间相对常用模型较长,但预报速度较快,总体上能够满足实时预报时间要求。

关键词: 钟差预报, BiLSTM, 超参数优化, 神经网络

Abstract: This paper proposes an enhanced model for BeiDou clock bias prediction, which extends the traditional LSTM to BiLSTM and introduces three adaptive hyperparameter matching algorithms (PSO, SSA, BOA) to improve short-term forecast accuracy. Firstly, LSTM is optimized to establish BiLSTM model, and three hyperparameter options have corresponding application scopes. Secondly, detailed steps for prediction with the hyperparameter-optimized BiLSTM model are outlined. Finally, the comparative experiments which consider the factor of the orbit types and sample intervals, are conducted for 1, 6, and 12 hours with the clock products provided by GFZ. The results show that the hyperparameter-optimized BiLSTM outperforms QP, GM, LSTM, and traditional BiLSTM models with an average improvement of 86.21%, 83.32%, 69.99%, and 55.17%. As for the three optimization schemes, SSA exhibits the best overall optimization, and PSO and BOA are more suitable for the IGSO and MEO satellites, respectively. Although the hyperparameter-optimized BiLSTM model takes a long time to train, its rapid forecasting speed can be guaranteed for the requirement of real-time applications.

Key words: clock bias prediction, BiLSTM, hyperparameter optimization, neural network

中图分类号: