引用本文: 刘海军, 雷东兴, 袁静, 乐会军, 单维锋, 李良超, 王浩然, 李忠, 袁国铭. 2024. 基于注意力机制LSTM的电离层TEC预测. 地球物理学报, 67(2): 439-451, doi: 10.6038/cjg2022Q0603 通讯作者: 袁静, 女, 1981年生, 副教授, 主要从事基于计算机视觉和深度学习的电磁数据智能处理相关的研究.E-mail: [email protected]

中图分类号:

电离层总电子含量(Total Electron Content，TEC)的监测与预报是空间环境研究的重要内容，对卫星通讯和导航定位等有重要意义.TEC值影响因素较多，很难确定精确物理模型来对其进行预测.本文设计了基于注意力机制的LSTM模型(Att-LSTM)，采用过去24小时TEC观测数据对未来TEC进行预测.选择北半球东经100°上，每2.5°纬度选择一个位置，共计36个位置来验证本文提出模型的性能，并与主流的深度学习模型如DNN、RNN、LSTM进行对比实验.取得了如下成果：(1)在选定的36个地区未来2小时单点预测上，基于本文的Att-LSTM模型的TEC预测性能明显优于其他对比模型；(2)讨论了纬度对Att-LSTM预测未来2小时TEC值时性能的影响，发现在北纬0°到60°之间，Att-LSTM预测性能随着纬度的升高而略有降低，在北纬62.5°~87.5°之间，模型预测性能出现扰动，预测效果略差；(3)讨论了磁暴期和磁静期模型的预测性能，发现无论是磁暴期还是磁静期，本文模型预测性能均较好；(4)还讨论了对未来多时点预测效果，实验结果表明，本文所提出的模型对未来2、4个小时的预测拟合度R-Square均超过0.95，预测结果比较可靠，对未来6、8、10个小时预测拟合度最高为0.7934，预测拟合度R-Square下降迅速，预测结果不可靠.

注意力机制 长短期记忆神经网络 总电子含量 Abstract:

The monitoring and prediction of total electron content (TEC) in the ionosphere is an important part of space environment research, which is of great significance for satellite communications, navigation and positioning. TEC values are affected by many factors, and it is difficult to determine an accurate physical model to predict them. In this paper, the LSTM model based on attention mechanism (Att LSTM) is designed to predict the future TEC using the TEC observation data of the past 24 hours. In this paper, we choose one location every 2.5 latitudes at 100° east longitude of the Northern Hemisphere, a total of 36 locations, to verify the performance of our proposed model. And compare with the mainstream deep learning models such as DNN, RNN, LSTM. The following achievements have been made in this paper: (1) On the single point prediction of the next two hours in the selected 36 regions, the TEC prediction performance based on the Att LSTM model in this paper is obviously superior to other comparison models; (2) The paper discusses the influence of latitude on the performance of Att LSTM in predicting the TEC value in the next two hours. It is found that between 0° and 60° north latitude, the prediction performance of Att LSTM decreases slightly with the increase of latitude. Between 62.5° and 87.5° north latitude, the prediction performance of the model is disturbed, and the prediction effect is slightly poor; (3) This paper discusses the prediction performance of the models of magnetic storm period and magnetic quiescent period, and finds that the models in this paper have good prediction performance for the next 2 hours, no matter in the magnetic storm period or the magnetic quiescent period; (4) The paper also discusses the prediction effect of future multiple time points. The experimental results show that the prediction fitting degree R-Square of the proposed model for the next 2 and 4 hours exceeds 0.95, and the prediction results are relatively reliable. The prediction fitting degree for the next 6, 8 and 10 hours is the highest 0.7934. The prediction fitting degree R-Square decreases rapidly, and the prediction results are unreliable.

Key words: Attention mechanism Long short-term memory Ionosphere Total electron content Le X A, Wan W X, Liu L B, et al. 2010. Development of an ionospheric numerical assimilation nowcast and forecast system based on Gauss-Markov Kalman filter—An observation system simulation experiment taking example for China and its surrounding area. Chinese Journal of Geophysics (in Chinese), 53(4): 787-795, doi: 10.3969/j.issn.0001-5733.2010.04.003 . Figure 7.

Comparison of prediction performance of different models in A ₁ —A ₃₆ area in Table 1

Figure 8.

Distribution of absolute errors between predicted and true TEC values for the Att-LSTM model in the latitude region 0° to 87.5°N

Figure 9.

Effectiveness of the Att-LSTM model for predicting TEC in the future two hours

Figure 10.

Histograms of absolute error distribution during magnetostatic and magnetic storm periods

Figure 11.

Comparison of TEC prediction performance of Att-LSTM models during magnetostatic (a) and magnetic storm (b) periods (Att-LSTM represents the TEC prediction value of the Att-LSTM model, and CODE represents the original TEC observations)

Figure 12.

Effectiveness of the Att-LSTM model in predicting TEC for multiple future time periods in the A ₁₁ —A ₁₉ regions

Figure 13.

Absolute error distribution histograms of the Att-LSTM model for the true and observed TEC values for multiple future time periods in the A ₁₁ —A ₁₉ regions