复杂系统管理与控制国家重点实验室知识库

空间天气是太阳到地球之间的环境状态。太阳剧烈活动会引发灾害性空间 天气事件，严重影响无线电传播、卫星通信、导航定位等系统的正常运行。地磁 暴是典型的灾害性空间天气事件，主要来源之一是冕洞。当太阳活动增加时，冕 洞向外喷射的高速带电粒子流会形成高速太阳风，太阳风与行星际磁场相互耦 合产生巨大能量带入地磁场从而引发地磁暴。

地磁暴形成过程中，地球同步卫星 SDO（Solar Dynamics Observatory，太阳 动力学天文台）和位于日地 L1 点（拉格朗日 1 点）的 ACE（Advanced Composition Explorer，高级成分探测器）可分别探测到反映冕洞信息的日面极紫外图像与太 阳风参数时间序列数据。由于探测站位置及数据获取方式的差异，极紫外图像与 太阳风参数反映的太阳活动将分别在 3～4 天、1～3 小时后影响地磁场，这说明 图像比太阳风数据包含更早的磁暴信息。根据这些数据信息可对地磁暴的等级， 地磁指数进行预报。目前地磁指数预报主要使用太阳风参数数据利用简单统计 模型进行，需人工提取特征，且只适用于少量典型有磁暴样本，不能自动拟合数 据关系，导致模型预测效果差且预测提前时间只有几小时。本文利用多模态日面 信息，基于深度学习建立预报模型，进行多尺度地磁指数预报，主要内容如下：

建立基于太阳风数据的预报模型。针对已有工作使用太阳风数据特征构造 忽略物理意义、数据集划分忽略样本均衡性的问题，改进特征构造方式，进行特 征选择并建立年份数据集。实验结果表明模型性能在提前 3 小时预测时显著提 升，但随预测提前时间变长而下降。这说明太阳风参数包含的信息不足以进行长 时间提前预测，又考虑到因图像特征与地磁指数对应关系不固定、不唯一导致模 型直接使用图像预测效果差的问题，提出同时使用图像与太阳风数据的多模态 预报模型。

建立基于图像预测太阳风的多模态深度学习预报模型。首先采用图像预测 太阳风风速，再与其他太阳风参数共同进行地磁指数预报。由于太阳及观测卫星 是轨道运动，因此某一时刻的日面极紫外图像自左至右不同区域特征反映了不 同时间条件下的日面状态。论文将改进后的卷积神经网络与长短时记忆网络结 合，同时提取图像的空间与时间特征。实验结果表明该模型不仅提升了预测性能，还将预测提前时间延长至 45 小时。

建立基于冕洞因子的多模态深度学习预报模型。通过霍夫变换、坐标变换、 经纬映射等预处理量化冕洞位置信息，计算所有样本的冕洞因子值，再与其他太阳风数据共同进行地磁指数预报。通过对比分析不同模型的效果，结果表明基于 冕洞因子的模型进一步提升了模型性能，并将预测提前时间由 45 小时延长至 3 天。

为了验证模型的实际应用能力，基于本文最佳模型对目前最新数据进行实 时预测。实验结果表明文中模型实时预测亦取得了较好效果。

Space weather concerns the state from the solar surface to the earth. Severe solar activity leads to space weather disaster events, which have a large impact on radio communication, satellite operation, navigation and positioning systems. Geomagnetic storm is a typical space weather disaster event which one of the main sources of are coronal holes. Coronal holes eject a high-speed charged particle stream forming solar wind When solar activities increase. Huge energy generated by interaction between solar wind and interplanetary magnetic field disturbs geomagnetic field resulting in geomagnetic storms.

There are two kinds of data can be obtained during geomagnetic storm events. The geostationary satellite SDO (Solar Dynamics Observatory) takes extreme ultraviolet images of coronal holes and the ACE (Advanced Composition Explorer) satellite located at L1 point (Lagrange 1 point) between sun and earth detects time series data of solar wind parameters. Due to the differences in location and acquisition ways of data between two satellites, geomagnetic field is disturbed after 3 to 4 days and 1 to 3 hours respectively by the solar activity reflected by extreme ultraviolet images and time series data of solar wind parameters, which shows that extreme ultraviolet images contain earlier magnetic storm information than data of solar wind. A geomagnetic index describes the intensity of the geomagnetic storm. We can predict the geomagnetic index accurately through making full use of data mentioned above. At present, geomagnetic index prediction mainly uses data of solar wind parameters with simple statistical model， which requires to extract features manually. So it is only suitable for a small number of typical samples with geomagnetic storms and not able to fit the data relationship automatically, resulting in poor prediction performance and prediction horizon of only a few hours. In this paper,multi-scale geomagnetic index prediction is based on multi-modal solar information with deep learning method. The main works are as follows:

Establish the geomagnetic index prediction model based on solar wind data.Aiming at problems that constructing feature of solar wind data ignores physical meanings andthe dividing data set ignores sample balance in previous work, improve the feature construction method, carry out feature selection and establish data set of years.Experimental results show that the model performance is significantly improved when forecasting 3 hours ahead, but it decreases as the prediction horizon becomes longer indicating that the information contained in solar wind parameters is not sufficient for long prediction horizon.Moreover，considering the poor performance when geomagnetic index prediction only uses images because of the weak correspondence between image features and geomagnetic index,the forecast model based on multi-modal data using both images and solar wind data is proposed.

Establish the geomagnetic index prediction model based on multi-modal data with deep learning by using images to predict solar wind speed. Firstly, the solar wind speed is predicted through images, and then the geomagnetic index is predicted together with other solar wind parameters. Because the sun and observation satellites move in orbit, the characteristics of different regions of solar extreme ultraviolet images at a certain time from left to right reflect the state of the solar surface under different time conditions. This paper combines the improved convolution neural network with long-short term memory network to extract the spatial and temporal features of images at the same time. Experimental results show that the model not only improves the prediction performance, but also extends the prediction horizon to 45 hours.

Establish the geomagnetic index prediction model based on multi-modal data with deep learning by calculating coronal factors.The coronal hole position information is quantified by Hough transforming, coordinate transforming and longitude latitude mapping. Therefore, values of coronal hole factors can be calculated for all samples.The geomagnetic index is predicted making use of solar wind data and coronal hole factors simultaneously. By comparing and analyzing the performances of different models, results show that the model based on coronal hole factors improves model performance further and extends the prediction horizon from 45 hours to 3 days.

Verify the practical application ability of the model in this paper. Geomagnetic index is predicted based on latest data in real time with the best model in this paper.Experimental results show that the real-time prediction also performs well.