太阳活动区EUV图像的生成式模型耀斑分级与预报

近年来，不断发射的空基观测台持续传送回海量日面图像及日地间气象数据，为采用人工智能技术对太阳活动进行预报预警提供了数据基础。但是，极端天气爆发少，样本量较少；中等程度爆发稍多，样本量较多；常规无爆发天气常见，样本较为集中，样本不均衡状况严重影响机器学习方法在空间天气领域的广泛应用。本文面向多源多通道多尺度日面图像信息，构建了来自SOHO和SDO的1996-2015年日面活动区图像数据集；针对数据分布的不平衡，对太阳活动区图像作耀斑分级与预报。在对比分析元学习算法的基础上，设计了结合分类头设计和卷积核初始化的生成式模型；在使网络轻量化的基础上，能够将M和X级耀斑预报的检测率指标相较于普通的深度学习模型和无监督度量式模型分别提升10%和7%。

Generative Model-based of Flare Hierarchic Recognition and Forecast of Extreme Ultraviolet Images in Solar Active Region

In the past 20 years, massive solar images and space meteorological data that have been transmitted back continuously and constantly with increasing space-based observatories launched, provide a promising material basis for the application of artificial intelligence technology to forecast and early warning solar activities. However, due to the less the extreme solar eruption and therefore the smaller relevant sample size, a slightly more the moderate solar activities outbreak and a little more the sample data set size, and the common routine non-outbreak space-weather always occurring and thus the samples become concentrated, thereby these condition and phenomenon result in sample imbalance and unlabeled data and so on which seriously affects the wide application of machine learning methods in the field of space weather forecasting and early warning. To handle the imbalance disturbance of sample data set for flare hierarchic recognition and forecast, this paper designs artificial intelligence algorithms for extreme ultraviolet images of solar active regions. Firstly, a dataset of extreme ultraviolet images of solar active regions from SOHO and SDO from 1996 to 2015 was constructed. Then the generated models combined classification head and initialization of convolution kernel are well-designed, and better index of accuracy for M and X flare are experimentally achieved and proved. Simultaneously, in terms of lightweight networks for deep learning, some comparison and analysis of multi algorithms on Meta learning were also discussed, this proposed method achieves finally 10% and 7% increments in POD accuracy compared with ordinary deep learning based method and unsupervised metric learning method, respectively.