基于人工神经网络的卫星钟差长期预报  

Long-Term Prediction of Satellite Clock Offset Based on Artificial Neural Network

作　　者：雷雨 徐劲松 蔡宏兵 

机构地区：[1]西安邮电大学计算机学院,陕西 西安 [2]江苏师范大学江苏圣理工学院,江苏 徐州

出　　处：《测绘科学技术》2022年第3期131-139,共9页Geomatics Science and Technology

摘　　要：卫星钟差长期预报可以为卫星自主导航提供必要的先验信息。针对二次多项式模型和灰色模型对钟差长期预报误差较大的问题,建立了适用于卫星钟差预报的人工神经网络(artificial neural network, ANN)模型。首先,利用一阶差分算子对卫星钟差序列作平稳化处理,获得钟差一阶差分序列;其次,使用差分序列构造网络所需的训练样本,并采用超限学习机(extreme learning machine, ELM)学习算法对网络进行训练;最后,使用网络模型对一阶差分序列进行逐步递推,再利用一阶累加算子对差分序列预报值进行还原,获得钟差长期预报值。以国际GNSS服务组织(Intentional GNSS Service, IGS)发布的事后GPS精密星历作为分析数据进行1~60 d的预报试验,结果表明,ANN模型的短期和长期预报精度和预报稳定性均明显优于二次多项式模型和灰色模型,其中,1~60 d预报精度相比于二次多项式模型和灰色模型分别提高87.51%和81.09%,预报稳定度比二次多项式模型和灰色模型分别提高88.94%和82.9%。Satellite clock offset is required for satellite autonomous navigation as priori information. The traditional models for long-term prediction of clock offset, e.g., the quadratic polynomial (QP) and GM(1,1) grey models, are not fully satisfactory. In current work, an artificial neural network (ANN) model is developed to accurately predict long-term satellite clock offset. A raw time-series of satellite clock offset is firstly differenced epoch-by-epoch by first-order differenced generation operation (1-DGO) to obtain a stationary series. Secondly, the differenced series is used as date basis to construct training dataset, and then the extreme learning machine (ELM) algorithm is employed for network training. Finally, multi-step extrapolation for differenced series is performed by the ANN model in a recursive way. Final GPS satellite clock products provided by the International GNSS Service (IGS) are taken as examples to carry out a 60 d-ahead prediction experiment. The results show that the accuracy and stability of the both short-term and long-term predictions generated by the developed ANN model are noticeably than those by the QP and GM(1,1) models. Compared with the QP and GM(1,1) models, the accuracy of the predictions up to 60 days is increased by 87.51% and 81.09%, respectively, and the stability of the 60-day-ahead predictions is increased by 88.94% and 82.9%, respectively.

关 键 词：卫星钟差 预报 人工神经网络 超限学习机 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]