Machine learning techniques applied to solar flares forecasting

Abstract

Space weather encompasses the Solar-Terrestrial environment’s interactions, emphasizing phenomena in the solar environment, such as sunspots, coronal mass ejections, and solar flares. The latter is one of the most relevant solar activity phenomena and the study object in this paper. The main challenge involved is to monitor the occurrence of solar flares and identifying features that help predict this phenomenon for specific classes. When M- or X-class flares occur, they may impact the health of astronauts and services used daily, such as satellite positioning services, telecommunications, and electrical networks — systems essential for modern life. This paper evaluates three methods for solar flares automatic classification: Support Vector Machine, Random Forest, and Light Gradient Boosting Machine. In addition to predicting the three available methods, we highlight the importance of variables for the LightGBM and Random Forest methods and the study of methods for data balancing. We also propose a system for predicting M- and X-classes flares 24, 48, and 72 h in advance. The methodologies used in the estimation and model validation processes involved cross-validation with stratified sampling and holdout methods, considering the use of balanced and imbalanced data. For the 24 h prediction horizon, we obtain the True Skill Statistic equal to 0.58 combining the algorithms via majority vote. We also achieved true positive and true negative rates equal to 0.81 and 0.77, respectively. For the forecasting in the interval between 24 h and 48 h, we obtained TSS higher than 0.50. In turn, for the interval of 48 h to 72 h, we obtained TSS higher than 0.54.