基于机器学习方法对23太阳活动周质子事件的研究

在耀斑伴随日冕物质抛射（CME）事件编目数据的基础上，进行太阳质子事件（SPE）匹配，构建研究数据集.利用Apriori算法挖掘SPE与耀斑级别、耀斑发生日面位置以及CME角宽度和速度的关联关系.结果表明:X级耀斑、全晕CME、高速（>1000km·-1） CME和日面西半球耀斑是最可能伴随质子事件的4种特征，其诱发质子事件概率依次为0.366，0.355，0.30，0.155.角宽度低于120°或速度低于400km·-1的CME产生质子事件的概率为0.高速CME产生质子事件的概率是低速（400～1000km·-1） CME的8.6倍，X级耀斑产生质子事件的概率是M级耀斑的6.2倍，日面西部耀斑产生质子事件的概率是日面东部耀斑概率的3.9倍，全晕CME产生太阳质子事件的概率是非全晕（120°～360°） CME的3.8倍.对太阳质子事件样本进行过采样处理，利用随机森林等5种典型有监督学习算法，构建了基于第23太阳活动周耀斑和CME特征的质子事件预测模型.结果表明，该预报模型的质子事件预测准确率、精确率和召回率均控制在91%以上. 

Research on Solar Proton Event in the 23rd Solar Cycle Using the Machine Learning Methods

Solar Proton Event (SPE) can pose crucial risks to the spacecraft. It is meaningful to analyze and build the relationships between SPE and the associated Coronal Mass Ejection (CME) and solar flares. In this study, the SPE in the 23rd solar cycle is investigated by using machine learning methods. Datasets were constructed based on CME and the solar proton events lists from 1997 to 2006 from the CDA web database. Apriori algorithm are used to survey the correlations between SPEs and the characteristics of flares and CME. The results show that X class flares, full halo CME, high speed (greater than 1000km·-1) CME, and western flares are the four characteristics that most likely to be associated with SPE. The corresponding probabilities are 0.366, 0.355, 0.30 and 0.155. The SPE probabilities at the condition of more than one (CME or flare) features occurring simultaneously were exhibited as well. Using the over sampled CME and flares features, five SPE prediction models are built through five different supervised machine learning algorithms, thus Logistic Regression, Support Vector Classification, K-nearest neighbor, Random Forest and Gradient Boosting Decision Tree. The models all present pretty good prediction accuracy (>0.94), precision (>0.96) and recall rate (>0.91).