基于机器学习的电离层总电子含量经验预报模型

发布时间：2018-03-29 21:36

本文选题：电离层总电子含量　切入点：电离层暴　出处：《中国科学院大学(中国科学院国家空间科学中心)》2017年硕士论文

【摘要】：电离层是近地空间环境的重要组成部分,当爆发性太阳活动引起的行星际扰动经过地球空间时,将剧烈扰动电离层,使电离层参数大幅度偏离静日变化范围。由于电离层总电子含量(TEC)与卫星信号的传播误差直接相关,对电离层TEC暴的物理原因研究和模拟预报十分具有应用价值。本文基于电离层TEC地图数据,讨论分析了中国地区电离层对行星际-地磁扰动的响应特征,并利用深度学习递归神经网络技术建立TEC的经验预报模型,主要内容和结果如下:(1)利用CODE TEC地图中国地区的数据,分析了2015年3月17-18日地磁暴事件中,中国地区电离层负暴的响应过程,并结合行星际太阳风参数,地磁活动指数和极盖电位降指数,探讨此次事件的物理过程。分析结果表明,中国110°E经度附近,中高纬在磁暴主相有短暂的正暴效应,各纬度在3月18日均呈现强负暴效应。本次事件行星际磁场南向分量持续较长时间,太阳风能量注入极区,极大扰动了热层环流,引起了中纬电离层抬升、大气中性成分向赤道输送和扰动发电机电场,导致初相中高纬的TEC正暴和恢复相各纬度的负暴。同时,低纬电离层垂直漂移数据显示,未发生穿透电场,这可能是低纬地区未发生明显正暴的原因。(2)首次将处理时间序列的深度学习递归神经网络应用于电离层参数预报,结合太阳活动指数、地磁活动指数和CODE TEC网格数据,并增加行星际太阳风参数作为预报输入,建立了单站电离层TEC预报模型,可提前24小时对北京站的TEC进行预测。同时,构建传统BP网络模型与递归神经网络模型进行对比。结果显示,RNN网络对宁静电离层的预测误差约为3.0 TECU,低于BP网络;将太阳风参数加入预报因子的模型对电离层正暴的预测准确率的提升7.5%。RNN网络对于2001,2015年31个强电离层暴预报的均方根误差比BP网络低0.2~0.5 TECU,太阳风参数使RNN模型对31个事件的平均误差降低了0.45~0.65TECU。深度递归神经网络比BP网络更适用于电离层TEC的短期预报,预报因子加入太阳风数据对电离层正暴的预报效果有明显改善。(3)首次建立全网格TEC数据输入的神经网络区域经验预报模型。模型分别使用地磁活动指数和行星际太阳风参数,结合太阳活动指数和35°N~45°N、100°E~120°E的TEC网格,作为输入因子,同时与多站点TEC和经纬度输入的模型进行比较。结果表明,全网格区域模型能够更高的捕捉TEC的空间扰动传播特性,预测精度较高。对中高纬格点的TEC预测误差仅为3.22 TECU,但对赤道异常峰控制的低纬地区,预测误差为4.25 TECU。区域模型对单格点TEC的预测精度也得到了改善。
[Abstract]:The ionosphere is an important component of the near-Earth space environment. When the interplanetary disturbance caused by explosive solar activity passes through Earth space, the ionosphere will be severely disturbed. The ionospheric parameters deviate greatly from the range of diurnal variation. Since the ionospheric total electron content (TEC) is directly related to the propagation error of the satellite signal, In this paper, based on the ionospheric TEC map data, the characteristics of the ionospheric response to the interplanetary geomagnetic disturbances in China are discussed and analyzed. The empirical prediction model of TEC is established by using the technique of deep learning recurrent neural network. The main contents and results are as follows: 1) using the data of CODE TEC map in China, the geomagnetic storm events of March 17-18, 2015 are analyzed. The response process of the ionospheric negative storm in China and the physical process of the event are discussed by combining the interplanetary solar wind parameters, the geomagnetic activity index and the polar cap potential drop index. The results show that China is near 110 掳E longitude. In the middle and high latitudes, there is a transient positive storm effect in the main phase of the magnetic storm, and strong negative storm effect is observed at all latitudes on March 18. The southward component of the interplanetary magnetic field lasts for a long time, and the solar wind energy is injected into the polar region, which greatly disturbs the thermospheric circulation. As a result of the ionospheric uplift in the middle latitudes, the transport of atmospheric neutral components to the equator and the disturbance of the electric field of the generator, the TEC positive storms in the first and high latitudes and the negative storms at all latitudes of the recovery phase are caused. At the same time, the vertical drift data of the low latitude ionosphere show that, There is no penetrating electric field, which may be the reason why there is no obvious positive storm in the low latitude region.) for the first time, the depth learning recurrent neural network for processing time series is applied to the prediction of ionospheric parameters, combined with the solar activity index. The geomagnetic activity index, CODE TEC grid data and the interplanetary solar wind parameters are added as the prediction input. A single station ionospheric TEC prediction model is established, which can predict the TEC of Beijing station 24 hours in advance. The results show that the prediction error of RNN for quiet ionosphere is about 3.0 TECU, which is lower than that of BP neural network. Improving the accuracy of forecasting Ionospheric positive Storm by adding Solar Wind parameters to Prediction factors the RMS root mean square error of 31 strong ionospheric storms in 2001 and 2015 is 0.20.5 TECU lower than that of BP neural network. The solar wind parameters make the RNN model correct 31 for 2001 and 2015. Deep recurrent neural network is more suitable for short-term prediction of ionospheric TEC than BP neural network. The effect of solar wind data on the prediction of ionospheric positive storms is significantly improved. (3) for the first time, a neural network regional empirical prediction model for the input of full-grid TEC data is established. The model uses geomagnetic activity indices and interplanetary solar wind parameters, respectively. Combined with the solar activity index and the TEC grid of 35 掳N ~ (45 掳) and 100 掳E ~ (120 掳E) as input factors, and compared with the multi-station TEC and longitude and latitude input models, the results show that the full-grid region model can capture higher spatial disturbance propagation characteristics of TEC. The prediction accuracy of TEC is 3.22 TECU for the middle and high latitude points, but 4.25 TECU for the low latitude region controlled by the equatorial anomaly peak. The prediction accuracy of the regional model for single lattice TEC is also improved.
【学位授予单位】：中国科学院大学(中国科学院国家空间科学中心)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P352

【相似文献】