基于LSTM-Informer模型的卫星钟差短期预报  

Short-term prediction of satellite clock error based on LSTM-Informer model

作　　者：王建敏[1] 孙廷松 孙建宇 李胜旗 WANG Jianmin;SUN Tingsong;SUN Jianyu;LI Shengqi(School of Geomatics,Liaoning Technical University,Fuxin,Liaoning 123000,China;Shaanxi Binchang Coal Industry Limited Liability Company,Xi'an 710075,China)

机构地区：[1]辽宁工程技术大学测绘与地理科学学院,辽宁阜新123000 [2]陕西彬长煤业有限责任公司,西安710075

出　　处：《测绘科学》2024年第8期16-24,共9页Science of Surveying and Mapping

基　　金：国家自然科学基金面上项目(41474020)。

摘　　要：针对非线性卫星钟差序列在预报过程中精度不高的问题,该文利用长短时记忆神经网络(LSTM)具有良好的数据回归特点与自注意力(Informer)神经网络模型在长时间序列预报中降低运算复杂度特性,提出了一种基于LSTM回归数据在Informer网络预报的组合模型。贝叶斯优化(BO)具有调参迭代次数少特点,通过BO优化LSTM模型和Informer模型超参数,可以有效解决两种模型参数设置问题。同时,将LSTM-Informer、BO-LSTM、BO-Informer、单一LSTM模型的预报精度进行对比。实验结果表明,LSTM-Informer模型精度相较LSTM、BO-LSTM、BO-Informer模型分别提升16.6%~93.1%、9.2%~92.5%、1.1%~89.2%;由此验证LSTM-Informer组合模型对于一次差分的非线性钟差序列进行预报的可行性。Aiming at the low precision of nonlinear satellite clock series in the forecasting process, the long-term memory neural network(LSTM)with good data regression characteristics and Informer neural network model were used to reduce the complexity characteristics of computing in long-term sequence forecast in this paper. A combination model based on the LSTM regression data on the Informer network forecast. Bayeste optimization(BO)has the characteristics of small number of messengers. Through BO optimization LSTM model and Informer model super parameter, it could effectively solve the problem of two model parameter settings. At the same time, the prediction accuracy of LSTM-Informer model, BO-LSTM model, BO-Informer model and single LSTM model were compared. Compared with LSTM model, BO-LSTM model and BO-Informer model, the accuracy of LSTM-Informer model was improved by 16.6%~93.1%,9.2%~92.5%,1.1%~89.2% respectively, which verified the feasibility of forecasting the LSTM-Informer combination model to a differential non-linear clock difference sequence.

关 键 词：钟差预报 长短时记忆 贝叶斯 自注意力机制 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] P228.4[自动化与计算机技术—控制科学与工程] TN967.1[天文地球—大地测量学与测量工程]