Description
Solar flares are large eruptions of electromagnetic radiation from the Sun that can affect the Earth's atmosphere and our technologies (e.g., radio communications). Flares are identified by the arrival of their energetic photons at Earth, meaning that their space-weather effects occur at the same time we become aware that a flare is in progress - this makes it essential for us to forecast them in advance. This work aims to predict solar flares within a 24-hour window using a Deep Learning model. We use 3D vectormagnetic images obtained from the Solar Dynamics Observatory (SDO) Space-weather HMI Active Region Patch (SHARP) data series, specifically the solar-radial component of the magnetic field. By using whole active region full-resolution images as input we want to improve our understanding of the physics leading up to flares and thus also improve forecasting accuracy. We use radial-field images from 2013 to 2023, inclusive, at a cadence of 24 hours along with the corresponding Geostationary Operational Environmental Satellites (GOES) X-ray flare events in the next 24 hours to create the image and flare-outcome label pairs. Filtering is performed to limit our data set to images containing single NOAA-numbered active region within ±75° longitude. With HARP separated data sets for training and testing, we implement a Fully Convolutional Network (FCN) for the binary classification of flare events with GOES X-ray flare class above C1. We present a statistical evaluation of the model’s predictive performance using various classification metrics, assessing its ability to distinguish between flare and non-flare events.
