基于长短时记忆神经网络的中国地区电离层TEC预测

电离层总电子含量(Total Electron Content，TEC)作为描述电离层形态、结构及变化的重要参量，一直是近地空间环境中重要的研究对象之一.本文利用太阳活动与地磁活动参量，结合欧洲定轨中心(Center for Orbit Determination in Europe，CODE)的TEC数据，给出了一种基于长短时记忆(Long Short-Term Memory，LSTM)神经网络的电离层TEC短期预测模型，并将其应用于2015年中国单站和区域电离层TEC提前1 h的预测中.单站TEC预测结果显示：LSTM神经网络模型预测的TEC与CODE-TEC的均方根误差为2.572 TECU (1 TECU=10 ¹⁶el/m²)，比国际参考电离层(International Reference Ionosphere，IRI)2016模型、反向传播(Back Propagation，BP)神经网络模型预测的TEC与CODE-TEC的均方根误差小5.183 TECU和0.667 TECU；在电离层扰动期与宁静期，LSTM神经网络模型预测的TEC与CODE-TEC的均方根误差分布在1.653~3.532 TECU，均方根误差明显小于IRI-2016模型、BP神经网络模型与CODE-TEC之间的均方根误差.中国区域预测结果显示：LSTM神经网络模型预测值与CODE-TEC值的均方根误差为2.721 TECU，比BP神经网络模型小0.716 TECU，其误差绝对值小于5 TECU的比例为92.83%，比BP神经网络模型的比例高5.77%，并且LSTM神经网络模型能更好地预测赤道异常区TEC的变化特征；同时，LSTM神经网络模型的预测值与CODE-TEC值具有较好的相关性，其相关系数达到0.989.整体而言，LSTM神经网络模型不仅能够准确反映中国地区电离层TEC时空变化特征，而且预测精度明显优于传统BP神经网络模型.

Prediction of ionospheric TEC over China based on long and short-term memory neural network

As an important parameter to describe the shape, structure, and changes of the ionosphere, the Total Electron Content (TEC) has always been an important topic of near-Earth space environmental research. In this study, the TEC data of the Center for Orbit Determination in Europe (CODE) and the parameters of solar activity and geomagnetic activity are combined to develop a regional TEC short-term forecast model based on the Long Short-Term Memory (LSTM) neural network. Meanwhile, the model is used to predict the single station and regional ionospheric TEC over China in 2015. The results for some single stations illustrate that the Root Mean Square Error (RMSE) of TEC predicted by the LSTM model is 2.572 total electron content unit (TECU, 1 TECU=1016el/m2), which is 5.183 TECU and 0.667 TECU lower than those predicted by the International Reference Ionosphere (IRI) 2016 model and the Back Propagation (BP) neural network model, respectively. During the magnetic disturbed and quiet days in 2015, the RMSE between the values predicted by the LSTM neural network and the CODE-TEC values are mainly from 1.653 TECU to 3.532 TECU, and the prediction accuracy of the LSTM neural network model is better than those of the IRI-2016 model and the BP neural network model. The results over China show that the RMSE between the predicted values of the LSTM neural network model and the CODE-TEC values is 2.721 TECU, which is 0.716 TECU smaller than that of the BP neural network model. The ratio of the absolute error values predicted by the LSTM neural network model less than 5 TECU is 92.83%, which is 5.77% higher than that of the BP neural network model. Furthermore, LSTM neural network model can well predict the change characteristics of TEC in the equatorial ionospheric anomaly region. The predicted values of the LSTM neural network model have a good correlation with the CODE-TEC values, and its correlation coefficient reaches 0.989. On the whole, the LSTM neural network model can not only accurately reflect the temporal and spatial characteristics of the ionospheric TEC in China, but also has a higher forecasting accuracy compared with the BP neural network model.