Study On Recommendation System For Coronal Mass Ejection Similar Events

CMEs(Coronal Mass Ejections)is one of the important events in Sun-Earth system as it can induce geomagnetic disturbance and associated space environment effect.It is of special significance to predict whether CMEs will reach the Earth and when they will arrive.In the current business forecast,because the effect of historical similar events should be fully considered in the artificial experience forecast to give the final prediction.Since machine learning in the real application of unexpected results have been achieved.In particular,combining a large number of data sets and computers has achieved fruitful results in various fields.Recommendation systems,as a kind of information filtering system in machine learning can predict for scores of items or user preferences in view of the huge amounts of information.The recommendation algorithm can recommend the history of the user needs to CME events.Calculating the distance between each of these events by reference to similar thoughts to recommend the historical events,makes references for forecasters to make accurate prediction.Centering on the prediction of CMEs,the main research contents of this paper are as follows:(1)A multiple association list of 215 different events with 18 features including CMEs characteristic,active region coordinates and solar wind parameters was established.Based on the CME list,we designed a prediction CME arrival time model based on recommendation algorithm principle.The cosine distance and Euclidean distance were compared according to the two calculation methods commonly used in the recommendation system.Each physical quantity feature will get the appropriate weights through training.Error analysis showed that Euclidean distance is better than cosine distance.The mean absolute error and root mean square error of the Euclidean distance are 11.78 hours and 13.75 hours respectively,which has reached the average level of other model in CME Scoreboard,verifying the feasibility and effectiveness of recommendation algorithm.This work provides a new attempt to use recommendation algorithms and is expected to be applied to space weather forecasting.(2)Based on machine learning method,the 10 parameters of CMEs were used to predict whether CMEs will reach the earth.The recommendation algorithm and logistic regression were used to carry out several experiments.The experimental processes was divided into two parts: training weight process and recommendation process.The first part of the data was used to train the weight of parameters.10% of the date as the test set were drawn to verify the model effect of the selected optimal weight.The second part of the data was used to make recommendations.The parameter weights obtained in the first part of the experiment are used to build a recommendation model to recommend similar CMEs in histories.Compared with the experimental results,it was found that the result of using the Logistic Regression module was more excellent,with the recall rate reaching 73%,which fully verifies that Logistic Regression was more intelligent than the single distance calculation,and can deeply learn the physical laws behind the data.However,compared with the recommendation algorithm,it can’t recommend CME events with similar history,and can’t provide historical events for the forecasters to refer to.(3)For all the CME events in history,the model was built by using the images taken by SOHO/LASCO C2.An experiment in which images were used to predict whether CMEs will arrive,the image processing way was divided into two experiments: the first experiment was to explode CMEs all images within 6 hours by decay manner of superposition.The multi-mechanism similarity loss based on GPW(General Pair Weighting)framework was used to construct model,carried out three sets of tests,the first group consisted of positive and negative data,the second group was just positive sample,and the third group was only negative sample.Three experiments the hit rate was 98.6%,73.9%,100%,which can verify the results of the experiment model are feasible.The second experiment was divided into two groups.In the first group,the picture obtained by difference between the brightest picture in each event and the former picture was input into the neural network model,and the Recall rate was 41.82%,and the F1 score was 19.66%.In the second group,a representative picture of each event was selected for cutting according to CPA and angular width and then input into the model.The Recall rate was 45.61% and F1 score was 32.50%,which was improved compared with the results of the first group of experiments.