Multi‐Wavelength Transformer‐Based 24‐Hour Solar Flare Forecasting at the Active‐Region Level

Abstract

Solar flare forecasting remains challenging due to the complex spatiotemporal evolution of solar active regions (ARs) and the severe class imbalance associated with high‐impact events. In this work, we investigate a transformer‐based framework for active‐region–level solar flare forecasting using short sequences of multi‐wavelength observations from the Solar Dynamics Observatory. The proposed approach integrates pretrained Vision Transformer representations with lightweight convolutional processing, explicit temporal differencing, and attention‐based temporal aggregation to examine the role of compact temporal context in short‐term flare prediction. Forecasting is formulated as a binary classification task targeting the occurrence of M‐class flares within a 24‐hr prediction horizon. Evaluation is conducted on the SDOBenchmark data set using active‐region–level aggregation and skill‐based metrics commonly adopted in space‐weather forecasting. The results indicate that combining spatial representations with explicit short‐term temporal modeling can yield stable forecasting skill under strong class imbalance. Across multiple random seeds, the selected configuration attains a mean True Skill Statistic of 0.81 0.04 and a Heidke Skill Score of 0.73 0.05, alongside high detection rates and controlled false‐alarm behavior.