摘要
针对基于物理参数的GNSS时间序列预测需要输入预测时刻的物理特征,这些未来时刻的特征取值通常很难获取的问题,该文引入深度学习模型TimesNet,以消除对未来时刻的物理特征输入的依赖,并选取10个连续运行基准站(CORS)站点的周解大地高时序开展预测实验。实验结果表明,TimesNet在不同站点的均方根误差(RMSE)保持在2~9 mm,在同一站点3—12个月内预测精度波动幅度小于3 mm,说明该模型在长时序预测时具有更强的鲁棒性;此外,数据的离散特性会对预测精度和模型参数的敏感性产生显著影响。总体而言,TimesNet模型在周解大地高时序数据的预测中表现出较为理想的效果。
Aiming at the problem that physical parameter-based GNSS time series prediction requires the input of physical features at the prediction moment,and the feature values taken at these future moments are usually difficult to obtain,the deep learning model TimesNet was introduced to eliminate the dependence on the input of physical features at future moments in this paper,and the prediction experiments were carried out by selecting the weekly solved geodesic high time series of 1o CORS sites.Experimental results showed that the RMSE of TimesNet was maintained at 2~9 mm at different sites,and the fluctuation of the prediction accuracy was less than 3 mm within 3-12 months at the same site,which indicated that the model had stronger robustness in the prediction of long time series;in addition,the discrete characteristics of the data will have a significant impact on the prediction accuracy and the sensitivity of the model parameters.Overall,the TimesNet model showed a more satisfactory effect in the prediction of high time-series data of circumscribed geodesy.
作者
时文斌
蒋涛
罗力
章传银
王伟
SHI Wenbin;JIANG Tao;LUO Li;ZHANG Chuanyin;WANG Wei(Chinese Academy of Surveying and Mapping,Beijing 100036,China;Beijing Fangshan Satellite Laser Ranging National Observation and Research Station,Beijing 100036,China;Liaoning Natural Resources Affairs Service Center,Shenyang 110086,China)
出处
《测绘科学》
北大核心
2025年第6期45-54,共10页
Science of Surveying and Mapping
基金
国家重点研发计划资助项目(2021YFB3900200,2021YFB390023)
中国测绘科学研究院基本科研业务费项目(AR2405)。
