Solar Flare Prediction Based On 3D Convolutional Neural Network

Solar flare is an important form of solar activities.Solar flares occur in the solar atmosphere,which will release huge energy and high-intensity electromagnetic radiation rapidly.The huge energy and high-intensity electromagnetic radiation will reach the space near earth in a short time,which will threaten short wave communication,satellite navigation and space flight,and seriously affect the production and life of human beings.Therefore,it is of great significance to forecast the solar flares,and it is an important work in the space weather forecast.So far,the specific physical mechanism of solar flares has not been very clear,and the prediction models of solar flares base on physics have not been established.Most of the existing prediction models depend on the statistical relationship between solar flares records and historical observation data of solar.In recent years,with the development of computer technology,the machine learning technology represented by convolutional neural networks have been used to establish prediction models of solar flares for many times.Many studies show that dynamic changes of solar activity parameters(mainly represented by magnetic field intensity)with time in solar active regions are closely related to the eruption of solar flares.However,traditional convolution neural networks are based on 2D convolution technology.And the 2D convolution kernels can only do the convolution calculations in 2D images,so the traditional 2D convolution neural networks cannot effectively capture the characteristics of the dynamic changes of solar observation data,which leads to two problems: the accuracy of the prediction models of solar flares is not higher enough and the training of forecasting models of solar flares relies on too much data.The main works of this paper are as follows:(1)the continuous solar observation data(i.e.the whole day magnetogram)are processed to fully reflect the dynamic changes of solar activity parameters with time in the solar active area;(2)two kinds of prediction models of solar flares based on 3D convolution neural networks are proposed in this paper: the binary classification models of solar flares prediction based on 3D convolution neural networks and four classification models of solar flares prediction based on 3D convolution neural networks.Both models adopt 3D convolution technology to fully capture the dynamic change characteristics of the solar magnetic picture taken by SDO/HMI(Solar Dynamics Observatory/ Helioseismic and Magnetic Imager),and finally forecast the solar flares.The experimental results show that the prediction models of solar flares based on3 D convolution neural networks have higher prediction performance than the traditional 2D convolution neural networks in both binary classifications and four classifications.In binary classification,the demand for the data set size for the prediction models of solar flares based on 3D convolution neural networks are much smaller than the prediction models of solar flares based on 3D convolution neural networks.