基于GNSS PWV和LSTM的AI气象模型PWV偏差校正方法

Bias Correction Method for PWV of AI Weather Models Based on GNSS PWV and LSTM

ES评分 0

DOI 10.14075/j.jgg.2025.12.414
刊名
Journal of Geodesy and Geodynamics
年,卷(期) 2026, 46(7)
作者
作者单位

1 桂林理工大学测绘地理信息学院, 桂林, 541006;
2 长安大学地质工程与测绘学院, 西安, 710054

摘要
基于2017-2019年中国地壳运动观测网络240个GNSS站的PWV数据, 结合风乌和盘古两类AI气象模型反演的PWV数据(FW PWV和PG PWV), 进行AI气象模型PWV偏差校正研究。首先, 以GNSS PWV为参考, 验证FW PWV和PG PWV精度。其次, 为提升两类AI气象模型反演PWV的精度, 以2017-2018年GNSS站的经度、纬度、高程、年积日、AI气象模型反演PWV作为输入特征, 以FW PWV或PG PWV与GNSS PWV的差值作为输出特征, 使用LSTM构建两类AI气象模型的PWV偏差校正模型, 对2019年两类AI气象模型反演的PWV进行校正。结果表明, FW PWV和PG PWV的平均bias和RMSE分别为-0.01 mm、2.88 mm与0.28 mm、3.04 mm。校正后FW PWVL和PG PWVL的RMSE分别为2.29 mm和2.38 mm, 精度改善率分别为20.48%和21.71%, 并通过2019年探空站PWV验证了校正结果的可靠性。偏差校正模型不仅降低了各月份的RMSE, FW PWVL和PG PWVL的月RMSE范围均集中在1.4~3.4 mm, 夏季平均RMSE改善率为22.30%。PG PWVL在0~1.5 km高程区间偏差绝对值分别降低0.45 mm、0.46 mm、0.32 mm, 在1.5~2.0 km高程区间RMSE改善率高达23.55%;FW PWVL在高程大于2 km区域表现最优, 偏差绝对值降低0.23 mm, RMSE改善率为29.50%。
Abstract
Based on PWV data from 240 GNSS stations of the Crustal Movement Observation Network of China from 2017 to 2019, combined with PWV data retrieved from two AI-based meteorological models, Fengwu and Pangu (hereafter referred to as FW PWV and PG PWV), this study investigates bias correction methods for AI-derived PWV. First, using GNSS PWV as a reference, the accuracy of FW PWV and PG PWV were validated. Subsequently, to improve the accuracy of the two AI-based models, a bias correction method was constructed using a long short-term memory (LSTM) network. The input features included longitude, latitude, elevation, day of year, and AI-derived PWV from 2017 to 2018 at GNSS stations, while the output feature was the difference between FW PWV or PG PWV and GNSS PWV. The method was then applied to correct the AI-derived PWV for 2019. The results show that the mean bias and RMSE of FW PWV and PG PWV are -0.01 mm, 2.88 mm and 0.28 mm, 3.04 mm, respectively. After correction, the RMSE of FW PWVL and PG PWVL decreased to 2.29 mm and 2.38 mm, with improvement rates of 20.48% and 21.71%, respectively. The reliability of the corrected results was further validated using radiosonde PWV data from 2019. The bias correction method effectively reduced RMSE across all months, with monthly RMSE values of FW PWVL and PG PWVL concentrated between 1.4 mm and 3.4 mm, and an average RMSE improvement rate of 22.30% in summer. For PG PWVL, the absolute bias in the elevation range of 0 to 1.5 km decreased by 0.45 mm, 0.46 mm, and 0.32 mm, respectively, while the RMSE improvement rate reached 23.55% in the 1.5 to 2.0 km range. FW PWVL performed best in regions above 2 km elevation, with an absolute bias reduction of 0.23 mm and an RMSE improvement rate of 29.50%.
关键词
大气可降水量; 盘古气象模型; 风乌气象模型; LSTM; GNSS
KeyWord
precipitable water vapor; Pangu-weather model; Fengwu-weather model; LSTM; GNSS
基金项目
页码 825-833
  • 参考文献
  • 相关文献
  • 引用本文

叶世润, 林买金, 杨文易, 刘艳平, 肖玉坤, 谢劭峰. 基于GNSS PWV和LSTM的AI气象模型PWV偏差校正方法 [J]. 大地测量与地球动力学. 2026; 46; (7). 825 - 833.

  • 文献评论

相关学者

相关机构