| A solar radio burst is the enhancement of radio emission during the release of solar magnetic energy.Solar radio bursts often bring electromagnetic disturbances,which affect the normal operation of aerospace and communication equipment.Moreover,with the continuous improvement of the resolution of solar radio observation equipment,the amount of radio data has increased significantly.However,most of these data are the quiet solar radio data,and it is not necessary to collect the data with high resolution in the whole time period during the quiet radio.Massive observation data brings huge data storage pressure.If solar radio burst events can be effectively predicted,people can not only take precautions in advance to minimize their impact on production and life,but also adjust the time resolution of radio observation equipment data collection according to the forecast results.This enables high-resolution observations of solar bursts,while reducing the resolution during quiet solar radio,which will greatly reduce the amount of data storage.At present,most radio prediction methods are based on numerical data,which require a lot of time and manpower to process.In response to this situation,this thesis attempts to propose a solar radio burst prediction method based on full-disk solar magnetograms,using the Convolutional Neural Networks(CNN)in deep learning to automatically cxtract the magnatic field information contained in the full-disk solar magnetograms.Through multiple trainings,the connection between magnetic field characteristics and solar radio bursts is established,so as to predict whether there will be solar radio bursts within 24 hours based on magnetograms.The experimental results of the model on 10 CV datasets demonstrate that the accuracy of the proposed model is 87.5%± 0.7%,and the True Skill Statistic(TSS)is 72.3%±2.6%.These results indicate the strong reliability and wide applicability of the forecast model proposed in this thesis.According to the prediction results of this method,the working status of the radio observation equipment when collecting data can be adjusted in time.Not only can the burst data be effectively collected,but also the data before and after the burst events can be obtained to achieve high resolution acquisition of complete data on radio burst processes.Furthermore,the proposed model is also used to predict solar type-Ⅱ and type-Ⅲ bursts,respectively.It is found that the prediction performance for type-Ⅲ bursts is better than that of type-Ⅱ bursts.The result is well explained from the differences of their formation mechanisms.To investigate the performance of the forecast model under the time-series segmented dataset,this thesis uses the observation data from 1996 to 2003 to train the model,and uses the data set from 2004 to 2010 to test it.The validity of the forecast model proposed in this thesis is proved by comparing with other researchers’ forecast results.Finally,for the type-Ⅱ burst events that have the greatest impact on disastrous weather among the solar radio burst events,this thesis uses the improved VGG16 network with deeper layers to carry out forecasting research.The image enhancement method is used to solve the problem of the imbalance between type-Ⅱ burst and non-type-Ⅱ burst samples.Then,aiming at the poor generalization ability of the model and the continuity of magnetic image features in this scenario,the rectangular convolution kernel and BN layer are integrated into the VGG16 network to improve the original network structure.The experimental results demonstrate that the accuracy of the improved VGG16 network is 92.1%,and the TSS is 83.6%,which are 2.0%and 3.3%higher than the original VGG16 network.This improved method achieves a relatively ideal forecasting effect,which helps to better mine the characteristics of type-Ⅱ radio burst events and make early warnings for disastrous space weather. |
PDF Full Text Request