一种组合模型的电离层总电子含量预报方法

Prediction of ionospheric total electron content by combined model

针对电离层总电子含量(TEC)的非线性、非平稳等多种因素影响会导致全球导航定位服务数据的高噪声问题,提出一种小波分解、长短期记忆(LSTM)网络模型、埃尔曼(Elman)神经网络模型组合的方法:采用国际全球卫星导航系统服务组织(IGS)中心提供的不同纬度、不同时间段的TEC格网点数据,利用db4小波分解对前20 d的TEC样本序列进行分解得到高频信息与低频信息;再分别利用LSTM模型和Elman模型对高频信息和低频信息进行预报;然后将2种模型的预报值进行重构;最后利用滑动窗口预测连续多个2 d数据进行分析研究。实验结果表明,组合模型在春、夏、秋、冬不同季节的电离层预报的均方根误差分别为0.85、0.68、0.84和0.84个总电子含量单位(TECu),平均绝对值残差分别为0.66、0.55、0.60和0.69个TECu,平均相对精度分别为97.1%、97.1%、96.7%、95.9%,与2种单一模型相比可有大幅度提升。

Aiming at the problem that it is liable to high noise in global navigation positioning service data for the nonlinear and non-stationary effects of ionospheric total electron content(TEC),the paper proposed a combination method of wavelet decomposition,long-short term memory(LSTM)network model and Elman neural network model:based on the TEC grid data of different latitudes and time periods provided by International GNSS(global navigation satellite system)Service(IGS)Center,db4 wavelet decomposition was used to decompose the TEC sample sequence of the first 20 days to obtain high-frequency information and low-frequency information;and LSTM model and Elman model were used to predict the high-frequency information and low-frequency information,respectively;finally,the sliding window was used to predict the data of several consecutive 2 days for the analysis.Experimental result showed that the root mean square error of ionospheric prediction of the combined model in different seasons of spring,summer,autumn and winter would be 0.85,0.68,0.84 and 0.84 total electron content units(TECu),respectively,the mean absolute difference be 0.66,0.55,0.60 and 0.69 TECu,respectively,and the average relative accuracy be 97.1%,97.1%,96.7%and 95.9%,respectively,which could be a great improvement compared with the two single models.