3:30 PM - 4:00 PM
[SCG55-11] Solar Flare Prediction and Space Weather Forecasting using Machine Learning
★Invited Papers
Keywords:Solar flare, Machine Learning, Prediction methods, Space Weather Forecasting
In recent years, prediction and analysis methods of solar flares that apply machine learning techniques have been intensely studied, as the amount of solar observation data has increased. In space weather forecasting, the prediction of solar flares has been a long-standing issue, and the application of machine learning techniques to solar images has made it possible to achieve higher prediction accuracy than that achieved by manual forecasting. Conventional solar physics research has focused on elucidating the elementary processes of solar flare mechanisms. Based on this knowledge and experience, our prediction model enabled more accurate prediction by extracting and learning the characteristics of sunspot magnetic fields and X-ray brightening, which appear as precursor phenomena before a solar flare, from a large amount of data. The prediction model is now in operation and is used in daily space weather forecasting meetings.
Our solar flare prediction model, named Deep Flare Net, consists of three parts. (1) Automatic detection of active regions around sunspots from solar corona and photospheric magnetic field data observed by X-ray, ultraviolet, and visible light, (2) extraction of 79 physical features from each region, and (3) input of these features to deep neural networks to predict the largest flare that will occur within 24 hours. The X-ray flux of a solar flare is expressed in logarithm, similar to the magnitude of an earthquake, and the flare size is classified as X, M, or C from the largest to the smallest, similar to the seismic intensity of an earthquake. Sunspots are formed by magnetic flux tubes emerging from the interior and accumulate magnetic distortion energy around magnetic neutral lines during their growth process. It has been found that capturing physical quantities representing magnetic distortion, trigger mechanisms, and pre-flare brightening is important for more accurate prediction.
True Skill Statistics (TSS) is used as the evaluation scale for flare prediction, and TSS is independent of the flare event rate of the training data, allowing for a fair comparison in the paper. Deep Flare Net has succeeded in achieving a prediction accuracy of 0.80 for greater than M-class flare predictions, which is better than 0.50 for manual predictions. Furthermore, international benchmark tests have been conducted in recent years, and the selection of a standard evaluation scale has once again become an issue. In space weather forecasting, we are trying to improve forecast accuracy by using AI forecasts and numerical simulation forecasts in different cases. In this talk, we will introduce our model and discuss the challenges of collaboration between solar observation data and data science.
Our solar flare prediction model, named Deep Flare Net, consists of three parts. (1) Automatic detection of active regions around sunspots from solar corona and photospheric magnetic field data observed by X-ray, ultraviolet, and visible light, (2) extraction of 79 physical features from each region, and (3) input of these features to deep neural networks to predict the largest flare that will occur within 24 hours. The X-ray flux of a solar flare is expressed in logarithm, similar to the magnitude of an earthquake, and the flare size is classified as X, M, or C from the largest to the smallest, similar to the seismic intensity of an earthquake. Sunspots are formed by magnetic flux tubes emerging from the interior and accumulate magnetic distortion energy around magnetic neutral lines during their growth process. It has been found that capturing physical quantities representing magnetic distortion, trigger mechanisms, and pre-flare brightening is important for more accurate prediction.
True Skill Statistics (TSS) is used as the evaluation scale for flare prediction, and TSS is independent of the flare event rate of the training data, allowing for a fair comparison in the paper. Deep Flare Net has succeeded in achieving a prediction accuracy of 0.80 for greater than M-class flare predictions, which is better than 0.50 for manual predictions. Furthermore, international benchmark tests have been conducted in recent years, and the selection of a standard evaluation scale has once again become an issue. In space weather forecasting, we are trying to improve forecast accuracy by using AI forecasts and numerical simulation forecasts in different cases. In this talk, we will introduce our model and discuss the challenges of collaboration between solar observation data and data science.