Research On Magnetic Type Recognition Of Sunspot Group And Solar Flare Prediction Based On Deep Learning

Solar active region is the interference source of the solar earth space environment,and solar flare eruption is a very violent solar activity,which can lead to very extreme space weather such as geomagnetic explosion.Severe space weather disturbances can cause damage to satellites,power facilities,short-wave radio communication systems on the ground and other systems,resulting in huge economic losses.Accurate prediction of space weather events,such as solar flares,can give humans enough time to prepare and take protective measures in advance to weaken the impact of severe space weather on human activities.Solar flares erupt mainly in the atmosphere above sunspots.Solar flares are related to the area,number,Mackintosh type and magnetic type of sunspot groups.Because the internal physical mechanism of solar flare eruption is still not clear at present,hidden laws can only be mined from a large number of solar observation data.Traditional processing methods cannot make full use of big data,while deep learning is good at mining hidden laws from big data.Therefore,the deep learning method is of great significance and bright prospect in the research of magnetic type identification of sunspot groups and solar flare outburst prediction.In this paper,a deep learning model is constructed to identify the magnetic types of sunspot groups,and a simplified Mt.Wilson magnetic classification is identified based on the information of continuous spectrum and magnetic field of solar active regions.In this model,the magnetic types of sunspot groups are extracted from the neural network with the structure of CNN and Res Net,and the integrated learning algorithm Light GBM is used to infer the magnetic types of sunspot groups by fusion features.The model parameters with stable performance were selected through5-fold cross-validation to minimize overfitting.The F1 scores of Alpha class,Beta class and Beta-X class were 0.9675,0.9327 and 0.8302,respectively.The reduced workload of manual classification is beneficial to the prediction of space weather events such as solar flares.Based on the magnetic field map and magnetic characteristics of the solar active region,this paper builds a model to predict whether there will be a solar flare of class C or above class M in the next 24 hours or 48 hours.According to two different common evaluation indexes,two models are constructed to optimize their respective evaluation indexes,and the results are better than other similar studies.The F1 score of F1＿FFM in 48-hour prediction of class C solar flares can reach0.656.TSS＿FFM model can reach 0.72 in 24 hours of m-class solar flares forecast on the common index TSS.This study shows that deep learning has great potential in space weather prediction.The experimental results show that the deep learning can be used in the prediction of the type recognition and the solar flask,and the good data segmentation scheme and the cross-verification can realize the fitting of the model in time,and the different indexes can allow the model to be optimized in different directions,which can meet different requirements.