Research On Deep Learning Based Solar Flare Prediction

The eruption of solar flares will bring a series of disastrous space weather events,such as changing the orbit of satellites and damaging radio communications.However,the exact physical mechanism of solar flares is unknown.Therefore,the study of solar flare prediction has important practical value and scientific significance.In recent years,most of the solar flare prediction models extract handcrafted features from the magnetogram of the active region and use traditional machine learning methods to forecast solar flares.Handcrafted features are often complicated in design and cannot express the rich magnetic field characteristics of the active region,which leads to performance bottleneck of flare prediction.In this thesis,we introduce the existing flare prediction methods in detail and analyze the shortcomings of these methods.On this basis,we propose a flare predcition method based on deep learning.Firstly,aiming at the problem of low efficiency and low accuracy caused by handcrafted features,we propose a novel method to extract features from the magnetograms of the active region based on convolutional neural network.Then,in order to overcome the shortcomings of the exisiting sequential model,we propose a method to express the temporal information of the active region based on long short-term memory network to further improve our flare prediction model.Experiments are conducted with the SHARPs data series published by SDO/HMI.The experimental results show that our method can effectively improve the performance of flare prediction.