极地研究

极地研究 ›› 2025, Vol. 37 ›› Issue (4): 786-799.DOI: 10.13679/j.jdyj.20240059

• 研究论文 • 上一篇    下一篇

基于深度学习的SuperDARN雷达极区电离层对流速度模型构建及预测

石胜胜1, 刘二小1, 刘建军2   

  1. 1杭州电子科技大学通信工程学院, 浙江 杭州 310018;
    2自然资源部极地科学重点实验室, 中国极地研究中心(中国极地研究所), 上海 200136
  • 收稿日期:2024-05-31 修回日期:2024-08-29 出版日期:2025-12-30 发布日期:2026-01-12
  • 通讯作者: 刘二小

SuperDARN polar ionospheric convection velocity model based on deep learning

SHI Shengsheng1, LIU Erxiao1, LIU Jianjun2   

  1. 1College of Communication Engineering, Hangzhou Dianzi University, Hangzhou 310018, China;
    2Key Laboratory of Polar Science, MNR, Polar Research Institute of China, Shanghai 200136, China
  • Received:2024-05-31 Revised:2024-08-29 Online:2025-12-30 Published:2026-01-12

PDF (PC)

141

摘要： 地球电离层等离子体对流是由太阳风与地球磁场相互作用驱动的磁层大尺度对流循环与对流电场在极区电离层的映射, 其形成机制与行星际磁场—地球磁场耦合系统息息相关。SuperDARN作为研究地球磁层中空间天气现象的雷达网络, 是研究中高纬电离层对流的重要手段, 也是获得极区电离层对流速度数据的重要来源。本文采用SuperDARN分布在北半球的20部高频相干散射雷达获取到的2015年和2016年拟合的二维电离层对流速度数据, 应用BP-Adaboost模型、FC-GRU模型以及ED-ConvLSTM时空序列模型构建了二维电离层对流速度2 min的全域预测模型。该模型的输入为行星际磁场三分量、太阳风速度、太阳风密度和地磁指数6个近地空间参数以及回波点对应的经纬度, 输出为二维对流速度。然后利用独立数据集, 基于预测值与实测值的均方根误差(ERMSE)、平均绝对误差(EMAE)对BP-Adaboost模型、FC-GRU模型以及ED-ConLSTM时空序列等3种模型的性能进行评估, 并引入决定系数(R2)对比3种模型的性能。预测结果表明, 基于集成学习的神经网络模型BP-Adaboost预测误差最小。其预测对流速度幅值的ERMSE为119.62 m·s–1、速度方向角为62.23°, EMAE为75.70 m·s–1、方向角为41.80°, R2为0.70、方向角为0.59。总之, BP-Adaboost模型预测对流速度与实际对流速度之间的ERMSE、EMAE、R2均优于FC-GRU和ED-ConvLSTM模型, 因此, BP-Adaboost模型对于电离层对流速度构建具有一定的优越性。

关键词: 极区电离层对流, 模型构建与预测, 深度学习, SuperDARN雷达

Abstract:

The convection of the Earth’s ionospheric plasma maps the large-scale convective circulation of the magnetosphere and the convective electric field in the polar ionosphere that is driven by the interaction between the solar wind and the Earth’s magnetic field; it is closely related to the coupling system between the interplanetary magnetic field and Earth’s magnetic field. SuperDARN is a radar network for the study of space weather phenomena in the Earth’s magnetosphere. It is an important source of polar ionospheric convection velocity data and is useful for studying ionospheric convection at mid- to high-latitudes. This paper employed the two-dimensional, fitted ionospheric convective velocity data obtained from 20 high-frequency coherent scattering radars in the SuperDARN network in the northern hemisphere in 2015 and 2016. Using the BP-Adaboost model, the FC-GRU model, and the ED-ConvLSTM spatiotemporal sequence model, a two-dimensional global prediction model of ionospheric convection velocity for 2 minutes was developed. Model input includes the three components of the interplanetary magnetic field, the six near-Earth space parameters of solar wind speed, solar wind density, and geomagnetic index, as well as the longitude and latitude corresponding to the echo point. Two-dimensional convection velocity is the model output. Independent datasets were used to evaluate the performance of the BP-Adaboost, FC-GRU, and ED-ConvLSTM spatiotemporal sequence models on the basis of the Root Mean Square Error (ERMSE) and Mean Absolute Error (EMAE) between predicted and measured values. The coefficient of determination R2 was introduced to compare the performance of the three models. The results show that BP-Adaboost, the neural network model based on ensemble learning, has the smallest prediction error: (1) the ERMSE for predictions of the magnitude and direction of convection velocity are 119.62 m·s–1 and 62.23°, respectively; (2) the EMAE for predictions of the magnitude and direction of convection velocity are 75.70 m·s–1 and 41.80°, respectively; (3) the R2 for predictions of magnitude and direction are 0.70 and 0.59, respectively. Relative to the FC-GRU and ED-ConvLSTM models, the performance of the BP-Adaboost model in predicting ionospheric convection velocity is superior in terms of ERMSE, EMAE, and R2 values for both convection velocity magnitude and direction.


Key words:

polar ionospheric convection, model construction and prediction, deep learning, SuperDARN radar