融合太阳磁场和黑子序列特征的太阳耀斑深度学习预报模型研究

Solar activities are the source of the Sun-Earth space environment changes, and the social development and national security could be affected by the space weather hazards. Therefore, the prediction of solar activity events has important application value. The traditional forecast models aim at revealing the statistical relationship between static physical parameters and solar flares, and their shallow learning methods face big challenges in depicting massive data. In recent years, the great success of deep learning in big data processing has brought new ideas for solar flare forecast research. Based on deep learning technology, the project intends to develop a deep learning solar flare prediction model that incorporates the characteristics of multi-source solar observation sequences. Aiming at the photosphere vector magnetic field observation sequence and the sunspot observation sequence of the solar active area, we propose a method of automatic extraction of spatiotemporal features based on convolution neural network and design a multi-source data feature fusion sharing and expression mechanism, then build a long short-term memory neural network based time series prediction model. This model is expected to automatically tap the temporal and spatial dependence of the solar active regions implied in the massive observational data, and then to discover the evolution rules among them, so as to improve the forecasting ability of solar eruptions and provide better space weather forecasting services.

太阳活动是日地空间环境变化的源头，由太阳爆发引起的空间天气灾害会对高科技依赖的人类社会和国家安全产生重要的影响。因此，对太阳活动事件的预报具有重要的应用价值。传统预报模型多采用浅层学习方法，主要以揭示静态物理参量与太阳耀斑之间的统计关系为目的，在分析海量太阳观测数据时面临极大的挑战。近年来，深度学习在大数据处理中的巨大成功为太阳耀斑预报研究带来新的思路。本项目拟展开融合多源太阳观测序列特征的深度学习太阳耀斑预报模型研究。课题面向太阳活动区光球矢量磁场观测序列和太阳黑子观测序列图像，提出一种基于卷积神经网络的时空特征自动提取方法，并设计多源数据特征融合共享表达机制，进而构建基于长短期记忆递归神经网络的时序预测模型。模型有望能自动挖掘海量观测数据中隐含的太阳活动区在时间和空间上的依赖关系，进而发现活动区的演化规律及其与耀斑的关系，从而提高太阳爆发活动的预报能力并更好地为空间天气预报服务。

太阳爆发活动引发的灾害将会影响社会和国家安全，因此对太阳活动事件的预报具有重要的意义。本项目的主要研究成果包括：1）进行了太阳活动数据搜集和整理，标注了约一万张图片；2）提出了基于深度学习的太阳活动区自动检测模型，准确率达到94%；3）提出了基于三维卷积神经网络的太阳活动预报模型、基于时序卷积网络的太阳活动预报模型，部分太阳爆发事件预测精度达到90%；4）研究并提出了多视图学习深度模型、基于深度卷积长短时记忆网络和集成深度学习框架的序列模式识别等模型。项目资助发表学术论文12篇，其中SCI检索论文9篇，EI检索国际会议论文3篇；项目组培养了3名博士研究生，2名硕士研究生。