Toward Accurate Solar Flare Prediction Using Machine Learning

The Sun has the power to sustain life on Earth. But it can also disrupt human activity. Solar flares generate electromagnetic radiation that emanates in all directions and eject billions of tons of magnetically charged plasma into space—possibly toward Earth. Radiation and plasma can damage orbiting satellites, disrupt communications, and shut down national power grids. Yet, there is no way to forecast solar flares reliably. This project explores using machine learning and ground-based Hydrogen-Alpha (Hα) imagery to define solar flare precursors (presumably faculae or “bright spots”) that can provide up to 5 days of warning of potential events. In 2014, the Global Oscillation Network Group (GONG) collected nearly one million Hα images of the Sun, and the National Oceanic and Atmospheric Administration (NOAA) recorded over 3000 solar flares. Correlating the characteristics of the faculae (size, shape, orientation, etc.) with a recorded flare may identify precursors to flaring events. But, the space-based NOAA records include hundreds of flares that the ground-based GONG sites cannot detect (due to clouds and atmospheric turbulence). Therefore, this project did not achieve its initial objective to find solar flare precursors. It did, however, reveal details to improve future work in this area. For follow-on work, I recommend including space-based data sources (with multiple wavelengths including ultraviolet and x-ray data), magnetic field fluctuations, and only superior Hα imagery (with little or no atmospheric impediments). This project laid the groundwork, but more work and data are needed to identify precursors and forecast solar flares.