Forecasting solar flares with a transformer network

Abstract

Space weather phenomena, including solar flares and coronal mass ejections, have significant influence on Earth. These events can cause satellite orbital decay due to heat-induced atmospheric expansion, disruption of GPS navigation and telecommunications systems, damage to satellites, and widespread power blackouts. The potential of flares and associated events to damage technology and disrupt human activities motivates prediction development. We use Transformer networks to predict whether an active region (AR) will release a flare of a specific class within the next 24 h. Two cases are considered: ≥C-class and ≥M-class. For each prediction case, separate models are developed. We train the Transformer to use time-series data to classify 24- or 48-h sequences of data. The sequences consist of 18 physical parameters that characterize an AR from the Space-weather HMI Active Region Patches data product. Flare event information is obtained from the Geostationary Operational Environmental Satellite flare catalog. Our model outperforms a prior study that similarly used only 24 h of data for the ≥C-class case and performs slightly worse for the ≥M-class case. When compared to studies that used a larger time window or additional data such as flare history, results are comparable. Using less data is conducive to platforms with limited storage, on which we plan to eventually deploy this algorithm.