基于深度学习的全球电离层TEC预测

Global ionospheric TEC prediction based on deep learning

电离层总电子含量(total electron content,TEC)是卫星时代以来最重要的电离层参数,具有重要的理论意义和应用价值.文中提出了一种基于深度学习方法的全球电离层TEC预测模型,采用编码器-解码器结构配合卷积优化的长短时记忆(long short-term memory,LSTM)网络来实现全球TEC的空间和时间预测.模型空间经纬度分辨率为5°×2.5°,时间精度为1 h.地磁活动平静时的预测结果表明,模型提前1天预测的TEC全局均方根误差(root mean-square error,RMSE)小于1.5 TECU,提前7天以内预测的RMSE小于2 TECU.在弱磁暴时期,模型预测的RMSE为2.5 TECU左右.不同地磁活动指数以及不同纬度情况下的对比结果发现,随着预测时间以及地磁活动剧烈程度的增加,模型预测的RMSE会逐渐变大,中高纬度地区模型的预测精度优于低纬赤道地区.

Total electron content(TEC)is the most important ionospheric parameter since the satellite era,and has important theoretical significance and application value.This paper proposes a global ionospheric TEC prediction model based on deep learning methods.We use an encoder-decoder structure to match the convolution-optimized long and short-term memory network(ConvLSTM)to achieve global TEC spatial and temporal prediction.The spatial latitude and longitude resolution of this model is 5°×2.5°,and the time accuracy is one hour.The prediction results when the geomagnetic activity is calm indicate that the model advances the global root mean-square error(RMSE)predicted for one day is less than 1.5 TECU,and the predicted root mean square error within one week in advance is less than 2 TECU.During the period of weak magnetic storm,the prediction error of this model is about 2.5 TECU.By comparing the results of different geomagnetic activity indexes and different latitudes,we found that with the increase of the forecast time and the intensity of geomagnetic activity,the error of this model will gradually increase,and the model has better predictions in the middle and high latitudes.