The Application Of Machine Learning In The Prediction Of Solar Eruptions

With the development of space exploration and space activities,the relationship between human and space is getting closer.However,the environment in which space activities take place is not invariable.When the Sun erupts violently,the space environment may be severely disturbed,which will seriously affect the technical systems of aerospace,communication,navigation and power transmission.The solar activity is the most important source of space environment disturbance.Accurate warning and forecasting of solar eruptions can give the space system more time to avoid and mitigate the effects of space weather disasters.Presently,the mechanism of solar eruptions is not fully understood.Analyzing morphological and evolutionary characteristics of ARs is an important means to predict solar eruptions.With the rapid accumulation of solar activity data,the traditional methods of manually extracting solar activity characteristics is no longer able to meet the needs of solar eruption predictions.It is an inevitable trend to apply machine learning method to forecast solar flares.In this paper,we used the convolutional neural network(CNN)to realize the automatic identification of AR magnetic types based on the Mount Wilson classification scheme,and built a preliminary solar flare prediction model.The main research contents and results are as follows:1.Based on the SDO/HMI SHARP continuum images and magnetograms taken during the time interval 2010-2017,we presented an automatic procedure for recognizing magnetic types in sunspot groups.A series of model training is carried out by utilizing the CNN method.In general,CNN has a productive performance in the identification of the AR magnetic types.The overall accuracy is over 95%,the recognition accuracy for Alpha type reaches 98.0%,the accuracy for Beta type maintains 89.5%,and the accuracy for Beta-x type is 96.0%.More details of the magnetogram can be obtained by increasing the number of convolution layers in CNN,and this is an effective way to improve the recognition accuracy of Beta magnetic type and Beta-x magnetic type.2.Based on the SOHO/MDI magnetogram taken during the time interval 1996-2010 and SDO/HMI SHARP magnetogram taken during the time interval 2010-2017,the CNN model was construct for forecasting solar flares.The prediction accuracy for the occurrence of a ? M class flare in ARs in the next 48 hr reaches 72.9%,and the prediction accuracy for no ? M class flares in the next 48 hr is 70.6%.