基于因果卷积与LSTM网络的电离层总电子含量预报

Prediction of Ionospheric Total Electron Content Based on Causal Convolutional and LSTM Network

电离层总电子含量(TEC)不仅是分析电离层形态的关键参数之一,同时为导航及定位等空间应用系统消除电离层附加时延提供重要支撑。由于电离层TEC的时空变化特征,本文融合因果卷积和长短时记忆网络,以太阳活动指数F_(10.7)、地磁活动指数Dst和电离层TEC历史数据作为特征输入,构建深度学习模型,实现提前24 h预报电离层TEC。进一步利用2005-2013年连续9年的CODE TEC数据,全面评估了模型在北京站(40°N,115°E)、武汉站(30.53°N,114.36°E)和海口站(20.02°N,110.38°E)的预报性能。结果显示不同太阳活动条件下三个站的TEC值与真实测量值的相关系数都大于0.87,均方根误差大都集中在0~1TECU以内,且模型预报精度与纬度、太阳、地磁活动程度、季节变化相关。与仅由长短时记忆网络构成的预报模型相比,本实验模型均方根误差降低了15%,为电离层TEC预报模型的实际应用提供了参考。

The Total Electron Content(TEC)of the ionosphere is not only one of the key parameters to analyze the shape of the ionosphere,but also provides an important support for the navigation,positioning and other space applications to eliminate the additional ionospheric delay.Due to the temporal and spatial variation characteristics of ionospheric TEC,an ionospheric TEC hybrid deep learning model based on Causal convolution and Long Short-Term Memory network is proposed in this paper.The solar activity index F_(10.7),the geomagnetic activity index Dst and the historical ionospheric TEC data are used as feature inputs to predict the TEC 24 hours in advance.Using CODE TEC data covering the low and high solar activities during 2005-2013,the performance of the model is comprehensively evaluated at Beijing station(40°N,115°E),Wuhan station(30.53°N,114.36°E)and Haikou station(20.02°N,110.38°E).The results show the correlation coefficients between the predicted TEC values of the three stations and the actual values under different solar activity conditions are greater than 0.87,and most root mean square errors concentrated within 1 TECU.The prediction accuracy of the model is related to latitude,solar activity,geomagnetic activity and seasonal variation.Compared with the prediction model composed of LSTM network,the root mean square error of the proposed model is reduced by 15%,which provides a valuable reference for the practical application of the ionospheric TEC prediction.