基于机器学习的ERP中长期预报方法研究

Medium- and Long-Term ERP Prediction Based on Machine Learning

ES评分 0

DOI 10.14075/j.jgg.2025.07.248
刊名
Journal of Geodesy and Geodynamics
年,卷(期) 2026, 46(5)
作者
作者单位

1 中国科学院上海天文台, 上海, 200030;
2 中国科学院大学天文与空间科学学院, 北京, 100049;
3 北京卫星导航定位中心, 北京, 100094

摘要
目前, 地球自转参数(ERP)的中长期预测(预测时长超过30 d)精度仅能达到毫角秒量级, 与天体测量学所要求的微角秒量级以及空间大地测量学所需的毫米量级存在显著差距, 难以满足实际需求。本文系统评估了4种深度学习模型(BP、RNN、LSTM和Transformer)在ERP短中长期预报中的性能, 重点关注长期预测。实验结果表明, 在180 d预测中, Transformer模型极移预测精度较传统方法提升43.8%, 日长变化(LOD)预测误差也降至0.27 ms; UT1-TAI预测中, BP模型精度最佳, RMSE为7.83 ms。该研究可为解决北斗系统自主运行中的ERP中长期预报难题提供参考。
Abstract
Currently, the accuracy of medium- and long-term predictions of Earth rotation parameters (ERP) (prediction duration exceeding 30 days) can only reach the milliarcsecond level. This level of accuracy is significantly lower than the microarcsecond level required by astrometry and the millimeter level needed by space geodesy. This paper systematically evaluated the performance of four deep learning models (BP, RNN, LSTM, and Transformer) in short-, medium-, and long-term ERP forecasts, with a focus on long-term predictions. The experimental results show that in 180-day predictions, the precision of polar motion forecasts by the Transformer model is 43.8% higher than that of traditional methods, and the prediction error of length of day (LOD) is reduced to 0.27 ms. In UT1-TAI predictions, the BP model achieved the best accuracy with an RMSE of 7.83 ms. This study can provide reference for solving the medium- and long-term ERP forecasting challenges in the autonomous operation of the Beidou system.
关键词
地球自转参数(ERP); 深度学习; 长期预报; 极移; UT1-UTC
KeyWord
Earth rotation parameters (ERP); deep learning; long-term prediction; polar motion; UT1-UTC
基金项目
页码 610-617
  • 参考文献
  • 相关文献
  • 引用本文

张思晨, 周善石, 郭睿, 李晓杰. 基于机器学习的ERP中长期预报方法研究 [J]. 大地测量与地球动力学. 2026; 46; (5). 610 - 617.

  • 文献评论

相关学者

相关机构