基于深度学习的SuperDARN雷达极区电离层对流电势模型构建及预测  

SuperDARN polar ionospheric convection potential model based on deep learning

作　　者：邓天云 刘二小 徐晨 DENG TianYun;LIU ErXiao;XU Chen(College of Communication Engineering,Hangzhou Dianzi University,Hangzhou 310018,China)

机构地区：[1]杭州电子科技大学通信工程学院,杭州310018

出　　处：《地球物理学报》2022年第3期819-829,共11页Chinese Journal of Geophysics

基　　金：国家重点研发计划(2018YFC1407300,2018YFC1407304);国家自然科学基金(41704154,41431072,41674169)资助。

摘　　要：电离层等离子体对流是表征电离层电动力学的重要现象,对空间天气有着重要的指示作用.SuperDARN雷达网是研究中高纬电离层对流的重要手段,也是获得对流电势数据的重要来源.本文基于SuperDARN雷达12月份的对流电势数据,采用BP模型、FC-LSTM模型以及ED-ConvLSTM时空序列模型构建了高纬电离层等离子体对流电势30min预测模型,然后采用独立的数据集,基于预测值与实测值的结构相似度(SSIM)、均方根误差(RMSE)以及线性相关系数(LC)三个统计指标对模型的性能进行了评估,同时分析对比了三种模型预测的越极盖电势(CPCP)和越极盖电场(CPEF)与实测值的统计分布情况.结果表明,BP模型和FC-LSTM模型由于没有充分挖掘到空间上的信息,因此整体误差较大,前者SSIM、LC以及RMSE分别为0.80、0.89、4.38 kV,后者为0.76、0.86、4.96 kV,而ED-ConvLSTM模型则分别为0.83、0.91、3.96 kV,因其能充分捕捉到空间上的信息,三个指标明显优于前两种模型,同时在CPCP分布和CPEF分布的相似度上,ED-ConvLSTM模型性能也表现得最好.本文结果证明了时空序列模型ED-ConvLSTM在对流电势预测上的有效性.As a crucial phenomenon characterizing the ionospheric electrodynamics,ionospheric plasma convection plays an essential role in indicating the weather of the space.SuperDARN radar network is an important tool in the study of the ionospheric potential convection in middle and high latitudes,and also the major source of convection data.Based on the convection potential data of SuperDARN in December,this paper adopted BP model,FC-LSTM model and ED-ConvLSTM spatiotemporal serial model to build the 30-minute prediction model of the ionospheric high-latitude plasma convection potential,and used independent datasets to evaluate the performance of the models based on the three statistical parameters of structural similarity(SSIM),root mean square error(RMSE),and linear correlation coefficient(LC)between the prediction value and the measured value.The Cross Polar Cap Potential(CPCP),Cross Polar Electric Field(CPEF)and the statistical distribution of the measured and prediction values in the three models have been analyzed and compared.The results indicate that,BP model and FC-LSTM model have larger deviations due to the failure of fully capturing the spatial information.The SSIM,LC and RMSE from the BP model are respectively 0.80,0.89,4.38 kV,and those from the FC-LSTM model are respectively 0.76,0.86,4.96 kV,while those from the ED-ConvLSTM model are respectively 0.83,0.91,3.96 kV which are notably better than the first two models owing to its ability to capture the spatial information.Meanwhile,in terms of the similarity between the CPCP distribution and CPEF distribution,ED-ConvLSTM model also performs better,which proved the effectiveness of the ED-ConvLSTM spatiotemporal serial model in predicting ionospheric convection potential.

关 键 词：深度学习 时空序列模型 超级双子极光雷达网 电离层对流电势 

分 类 号：P352[天文地球—空间物理学]