基于多源数据和广义回归神经网络的ZWD预报模型  

A predicting ZWD model based on multi-source data and generalized regression neural network

作　　者：黎峻宇 姚宜斌[3] 刘立龙 张豹[3] 黄良珂 曹利颖 LI Junyu;YAO Yibin;LIU Lilong;ZHANG Bao;HUANG Liangke;CAO Liying(College of Geomatics and Geoinformation,Guilin University of Technology,Guilin 541006,China;Guangxi Key Laboratory of Spatial Information and Geomatics,Guilin 541006,China;School of Geodesy and Geomatics,Wuhan University,Wuhan 430079,China)

机构地区：[1]桂林理工大学测绘地理信息学院,广西桂林541006 [2]广西空间信息与测绘重点实验室,广西桂林541006 [3]武汉大学测绘学院,湖北武汉430079

出　　处：《测绘学报》2023年第9期1492-1503,共12页Acta Geodaetica et Cartographica Sinica

基　　金：广西科技计划项目(2020GXNSFBA297145,GuikeAD23026177);国家自然科学基金(42064002,42074035);广西空间信息与测绘重点实验室基金(21-238-21-05);桂林理工大学科研启动基金(GUTQDJJ6616032)。

摘　　要：对流层湿延迟是GNSS误差源较难改正的部分。主流的天顶湿延迟(ZWD)经验模型大多基于单源数据,如探空数据或再分析资料的一种,且通过预设模型函数表征ZWD在不同尺度上的变化,难以准确描述ZWD在不同尺度上的非线性复杂变化,精度有待进一步提高。针对此问题,基于多源数据和具有强大非线性逼近能力的广义回归神经网络(GRNN),构建了一种ZWD预报模型。首先,采用GRNN模型优化和降采样两种不同数据源的格网ZWD,并将其与探空ZWD合并,获取高质量的ZWD数据集;然后,根据ZWD受时空影响较大的特点和机器学习的特点,设计了GRNN训练模型的输入变量和输出变量;最后,采用后验寻优的方法确定模型参数,进而获得最优的预报模型。经探空ZWD验证,相比国际典范经验模型GPT2w,模型的精度提高了25.3%(以RMS计);相对同方法单源(探空)数据模型,精度改善了11.1%;且模型的预报精度具有良好的时空稳定性。此外,计算效率和PPP应用试验结果表明,模型的计算效率可满足GNSS实时应用的需求,且应用于PPP的改进效果优于GPT2w。本文所提方法无须预设模型函数即可获得较高的ZWD预报精度,为ZWD模型化提供了一种思路。Tropospheric wet delay is a more difficult part of GNSS error sources to be corrected.Most of the approved empirical models of zenith wet delay(ZWD)are based on single-source data(i.e.radiosonde data or reanalysis data),and the variation patterns of ZWD on different scales are characterized by preset model functions,so it is difficult to accurately describe the nonlinearly complex variations of ZWD on different scales,and the accuracy needs to be further improved.To address this issue,a predicting ZWD model is constructed based on multi-source data with higher spatiotemporal resolution and a generalized regression neural network(GRNN)with strong nonlinear approximation capability.Firstly,grid ZWD of two different data sources is optimized and downsampled by a GRNN model,and merged with radiosonde ZWD to obtain high-quality ZWD dataset.Then,the input and the output vectors of the GRNN training model is designed according to the characteristics that ZWD is greatly affected by time and space and the characteristics of machine learning.Finally,a posteriori optimization method is used to determine the model parameters,and then the optimal forecasting model is obtained.Validated by the radiosonde ZWD,in comparison with the approved empirical GPT2w model and the single-source(i.e.radiosonde)data model with the same method,the accuracy of the proposed model is improved by 25.3%and 11.1%respectively in terms of RMS.And the accuracy of the proposed model has good spatiotemporal stability.In addition,the computational efficiency and PPP application experimental results show that the computational efficiency of the proposed model can meet the needs of GNSS real-time applications,and the improvement effect of PPP is better than that of GPT2w.The proposed model obtains high ZWD forecasting accuracy without setting the model function,which provides an idea for ZWD modeling.

关 键 词：对流层湿延迟 多源数据 GRNN 非线性逼近 预报 

分 类 号：P228[天文地球—大地测量学与测量工程]