Solar Flare Forecasting Using Machine Learning and SDO/HMI Data: A Multiple Machine Learning Model and Data Curation Technique Comparison Study

Abstract Forecasting solar flares is essential to mitigate their harmful effects on our technologies. This paper describes an in-depth study of solar flare forecasting using machine learning (ML) models. These ML models utilize active region patch parameters extracted from the Solar Dynamics Observatory Helioseismic and Magnetic Imager (HMI) data; specifically, Space-weather HMI Active Region Patch (SHARP) parameters. Six models are considered: k -nearest neighbors, linear discriminant analysis, logistic regression, quadratic discriminant analysis, random forest classifier, and support vector machine. A detailed analysis of hyperparameter selection for each of these models is included in the study. The importance of each SHARP parameter in model performance is also examined. A major component of this work examines the incorporation of active region flaring history into the forecasting process. This examination indicates that this incorporation substantively improves forecasting accuracy. Another major contribution of the work here is a consideration of the effects of multiple data curation methods on the predictive performance of the models. The results presented here indicate that using high-quality data can markedly improve predictive accuracy. Furthermore, this work demonstrates that using SHARPs that contain more than one active region limits the performance of these models. The performance of models with varying lead times is also explored, providing insights into the choice of the best prediction windows.