Prediction of X-ray solar flare based on active region evolution Santi Sulistiani, Tiar Dani
Research Center for Space, National Research and Innovation Agency (BRIN)
Abstract
Accurate, timely warnings of solar flares are vital for safeguarding satellites, power grids, and HF communications. We develop a supervised machine-learning classifier that predicts, 24-hours in advance, whether an active region will produce a C-, M-, X-class flare or remain quiet, by ingesting two-day time-series of heliographic longitude, McIntosh and magnetic classifications, area and spot count. Trained on January 1998-June 2018 events and validated on an independent July 2018-March 2025 set, the model reaches 66.5% overall accuracy (precision = 0.70, recall = 0.66, F1 = 0.68). Skill is highest for the dominant no-flare class (F1 = 0.82) but drops for rarer M (F1 = 0.26) and X (F1 = 0.13) flares, reflecting severe class imbalance. These results benchmark the current limits of feature-based flare forecasting and motivate future work on balanced training strategies and physics-informed predictors to improve detection of high-impact events.
Keywords: solar flare- prediction- space weather- machine learning
