Research And Implementation Of The Prediction Algorithm For The Arrival Time To Earth Of CMEs Based On Deep Learning

Coronal mass ejections(CMEs)are solar eruption phenomena that have a huge impact on the solar-terrestrial space environment.When arriving at Earth,it may disturb the geomagnetic field and cause geomagnetic storms.Fast and accurately predicting the arrival time of CME events that cause geomagnetic disturbances to Earth is vital to reduce the harm caused by CMEs.Almost all traditional methods are only useful for geoeffective CMEs when predicting the arrival time to Earth,but current methods cannot accurately determine the geoeffectiveness of CME events.The definition of geoeffective CMEs is that CMEs eventually arrive at Earth and cause geomagnetic disturbances.In fact,most CME events are not geoeffective,which will lead to that the traditional prediction methods are not universal and the prediction is not comprehensive.In this thesis,we propose a deep learning algorithm based on satellite time series observation images,which combines the geoeffectiveness and arrival time prediction of the CME events in deep learning for the first time.Specifically,the arrival time prediction to Earths of CMEs is divided into two stages.Our algorithm first determines the geoeffectiveness of CME events,and for the geoeffective CME events,it further predicts their arrival time.Firstly,we construct a large-scale CME data set suitable for the research and evaluation of the algorithm in this thesis.Secondly,on the geoeffectiveness prediction of CMEs,a deep residual network embedded with the attention mechanism is proposed,which can effectively extract key regional features;the feature map fusion algorithm of indefinite length is proposed to solve the situation that the number of images in each CME event is not fixed which bases on the point importance of the feature map and can effectively fuse the feature maps extracted from each image;knowledge distillation method is used to compress and accelerate the model,which assists lightweight student model training by ensemble the knowledge of multiple teacher models.Finally,for the arrival time prediction of geoeffective CMEs,a data expansion method based on sample splitting is proposed for the case that there are few geoeffective CME events,so that the data size can be increased by 1 0x;a deep residual regression network based on group convolution is proposed to balance the performance and speed of the prediction;a cost-sensitive regression loss function is proposed to allow the model to focus more attention on difficult samples.A prototype system supporting user interaction is built based on the proposed algorithm,and we conduct a large number of experiments to evaluate the performance of the algorithm and system.The results show that the algorithm of this thesis has a better performance than the traditional deep learning methods and the system has good real-time and robustness.