极区电离层对流速度的浅层神经网络建模与分析  

Modeling the polar ionospheric convection velocity vectors using shadow neural networks

作　　者：王平[1] 李洁[1] 韩冰[1] 胡泽骏[2] 高新波 刘建军[2] 胡红桥[2] WANG Ping;LI Jie;HAN Bing;HU ZeJun;GAO XinBo;LIU JianJun;HU HongQiao(School of Electronic Engineering,Xidian University,Xi′an 710071,China;SOA Key Laboratory for Polar Science,Polar Research Institute of China,Shanghai 200136,China;The Chongqing Key Laboratory of Image Cognition,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]西安电子科技大学电子工程学院,西安710071 [2]中国极地研究中心自然资源部极地科学重点实验室,上海200136 [3]重庆邮电大学重庆市图像认知重点实验室,重庆400065

出　　处：《地球物理学报》2022年第4期1197-1213,共17页Chinese Journal of Geophysics

基　　金：国家自然科学基金项目(61572384,62076190,41874195和41831072);国家重点研发计划重点专项课题(2018YFC1407303和2018YFC1407304);国际子午圈关键科学问题国际合作预先研究(A131901W14)联合资助.

摘　　要：电离层对流是太阳风与地球磁场相互作用下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,与行星际磁场-地球磁场耦合系统息息相关.本文基于SuperDARN(Super Dual Aurora Radar Network)分布在北半球的23部高频相干散射雷达获取到的二维电离层对流速度对其进行建模研究.模型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六个空间物理参数,模型输出为二维对流速度.模型选择两种广泛应用于空间物理建模的浅层神经网络即广义回归神经网络(General Regression Neural Network,GRNN)和误差反向传播(Back Propagation,BP)神经网络.实验结果显示,GRNN模型和BP模型的速度幅值均方根误差分别为174.96 m·s^(-1)和234.21 m·s^(-1),速度方向角均方根误差分别达到62.30°和88.07°,相比于对流速度最大值2000 m·s^(-1)和360°的方向角范围来说,其误差是可以接受的.外推性实验结果显示,在第24个太阳周期时,GRNN模型和BP模型的速度幅值均方根误差分别为305.35 m·s^(-1)和738.15 m·s^(-1),速度方向角均方根误差分别为82.01°和90.56°.实验结果表明,GRNN在时间外推性上的效果优于BP神经网络,更适用于预测对流速度.我们发现在四种典型空间环境条件下,利用GRNN模型预测的瞬时对流速度来构建的全域对流模式与现有统计模型构建的对流模式相似,从而验证预测的对流速度可以用于分析瞬时极区电离层对流.The ionospheric convection is primarily driven by interactions between the solar wind and magnetosphere which represents the coupled solar wind-magnetosphere system.The global ionospheric convection map contributes to ionospheric dynamics analysis.The convection velocity vectors observed by 23 HF coherent backscatter radars in the Northern Hemisphere of SuperDARN(Super Dual Aurora Radar Network)are used in predict modeling of convection velocities.The inputs of predict model are 6 dimensional vectors(IMF B_(x),B_(y),B_(z),V_(p),N_(p),AE),and the outputs are magnitudes and azimuths of predicted convection velocities.General Regression Neural Network(GRNN)and Back Propagation(BP)neural network are employed to predict convection velocities.For GRNN and BP,the root mean square errors(RMSEs)of predicted velocity magnitudes are 174.96 m·s^(-1)and 234.21 m·s^(-1)respectively.The RMSEs of predicted velocity azimuths for these two models are 62.30°and 88.07°.Due to the maximization of velocity magnitude and azimuth are 2000 m·s^(-1)and 360°,the RMSEs of GRNN and BP are acceptable.When testing on the dataset of next solar cycle,the RMSEs of GRNN are 305.35 m·s^(-1)and 82.01°.The RMSEs of BP are 738.15 m·s^(-1)and 90.56°.It is obvious that the RMSEs of GRNN maintains lower than BP which means that the predict performance of GRNN is better than that of BP.After fitting predicted convection velocities of GRNN,the global convection maps are similar to maps from empirical models.These results show that the convection patterns based on predicted velocities from neural networks could be applied into polar convection pattern analysis.Moreover,predicted velocities by GRNN and BP could be used to convection velocity map completion.

关 键 词：神经网络 电离层对流速度 预测建模 全域电离层对流模式 高频相干散射雷达 

分 类 号：P352[天文地球—空间物理学]