Solar Flare Class Prediction Using Machine Learning

Abstract

Solar flares are intense bursts over the surface of the Sun which interrupts the satellite signals, power grid, and communication systems, hence the accurate prediction is essential for mitigating the potential impacts caused by these kinds of flares. This paper outlines a machine learning approach to solar flare prediction from analysis of magnetic field evolution of the Sun. From NASA's SHARP dataset of 24-hour time-series magnetic parameters May 2010–May 2018, supplemented with eight flare-history features from the GOES X-ray catalog (May 2010–May 2018), we trained a model with dimensionality reduction, feature selection, and supervised learning. Principal Component Analysis (PCA) preserved temporal magnetic field patterns with dimensionality reduction, and feature selection methods chose the 10 most informative features from a starting set of 25. This article equipped the Random Forest Classifier to accurately predict the flares because of its robustness against the high-dimensional data. The Random Forest achieved 90.91% accuracy and 87% macro-average precision in flare prediction from C1 to X9 classes, beating baselines. These findings suggest the promise of the application of PCA, unsupervised feature selection, and ensemble learning for online solar flare prediction, offering a scalable and robust platform for early-warning space weather systems.