基于GPT3和改进注意力LSTM-MHA的区域ZTD预测模型

A Regional ZTD Prediction Model Based on GPT3 and Improved Attention LSTM-MHA

ES评分 0

DOI 10.14075/j.jgg.2025.07.250
刊名
Journal of Geodesy and Geodynamics
年,卷(期) 2026, 46(4)
作者
作者单位

1 湘潭大学自动化与电子信息学院, 湘潭, 411105;
2 湘潭大学数学与计算科学学院, 湘潭, 411105;
3 成都国星通信有限公司, 成都, 611731

摘要
针对常用的ZTD预测模型中经验模型精度较低, 气象参数模型精度受限于气象参数的问题, 提出一种结合长短期记忆网络的改进多头注意力机制(LSTM-MHA)的区域ZTD预测模型, 基于第三代全球气压和气温模型(GPT3)数据与站点经纬度、高程、年积日、小时进行建模。选取欧洲地区26个连续运行参考站(CORS)2023年的ZTD数据作为参考值, 与GPT3模型、LSTM模型、CNN模型进行对比实验。结果显示, LSTM-MHA模型结果平均偏差接近于0, 均方根误差和平均绝对误差分别为13.43 mm和12.4 mm, 精度相比于GPT3、CNN、LSTM模型分别提高60.8%、55.2%、53.3%。通过评估LSTM-MHA模型在不同距离和不同预测时长下的性能发现, 所提方法在稀疏站点群和长时预测中均表现稳定, 适合区域多站点预测任务。
Abstract
Aiming at the problems of low accuracy of empirical models and meteorological parameter model accuracy limited by meteorological parameters in commonly used ZTD prediction models, we propose a regional ZTD prediction model based on improved multi-head attention mechanism combined with long short-term memory network (LSTM-MHA). The model is constructed using data from the third-generation global pressure and temperature model (GPT3), along with station latitude, longitude, elevation, day of year, and hour of day. Using ZTD data from 26 continuously operating reference stations (CORS) in Europe in 2023 as reference values, comparative experiments were conducted with the GPT3 model, LSTM model, and CNN model. The results show that the LSTM-MHA model achieves a mean bias close to 0, with root mean square error and mean absolute error of 13.43 mm and 12.4 mm, respectively. The accuracy of LSTM-MHA model has been improved by 60.8%, 55.2%, and 53.3% compared to the GPT3, CNN, and LSTM models. Evaluation of the LSTM-MHA model performance under different distances and prediction lengths reveals that the proposed method performs stably in both sparse station networks and long-term forecasting scenarios, demonstrating its suitability for regional multi-station prediction tasks.
关键词
对流层延迟; 多头注意力机制; 长短期记忆网络; 区域建模; 气象参数模型
KeyWord
tropospheric delay; multi-head attention mechanism; long short-term memory network; regional modeling; meteorological parameter model
基金项目
页码 473-480
  • 参考文献
  • 相关文献
  • 引用本文

蔡成林, 赵婕, 关文绘, 秦懿. 基于GPT3和改进注意力LSTM-MHA的区域ZTD预测模型 [J]. 大地测量与地球动力学. 2026; 46; (4). 473 - 480.

  • 文献评论

相关学者

相关机构