Solar-Flare Forecasting with Multi-Wavelength Transformers

Dunia Alatoom, Nikolaos Nikolaou

Keywords: Vision Transformer, Solar Flare Forecasting, Space Weather, Time-Series

Domains: AI for Science

TL;DR: A SoTA transformer-based deep learning framework for active-region-level solar flare forecasting that combines spatial & short-term temporal modeling.

External Link: Preprint: https://essopenarchive.org/doi/full/10.22541/essoar.177213252.29321624/v1 (Final accepted version TBC)

Abstract:

Solar flare forecasting remains a challenging problem due to the complex spatiotemporal evolution of solar active regions and the strong class imbalance associated with high-impact flare 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 combines pretrained Vision Transformer representations with lightweight convolutional processing, explicit temporal differencing, and attention-based temporal aggregation to examine the value of compact temporal context for short-term flare prediction.

Forecasting is formulated as a binary classification task predicting the occurrence of ≥M-class flares within a 24-hour forecasting horizon. Evaluation is performed on the SDOBenchmark dataset using active-region-level aggregation and skill-based metrics commonly adopted in space-weather forecasting. The results suggest that integrating spatial representations with explicit short-term temporal modeling can provide stable forecasting skill despite severe class imbalance. Across multiple random seeds, the selected configuration achieves a mean True Skill Statistic (TSS) of 0.81±0.04 and a Heidke Skill Score (HSS) of 0.73±0.05, while maintaining high detection rates and controlled false-alarm behavior. Finally, we present a detailed analysis of the most common false-positive and false-negative prediction patterns.

The paper has been accepted to the American Geophysical Union Space Weather journal. A preprint is available at: https://essopenarchive.org/doi/full/10.22541/essoar.177213252.29321624/v1

Submission Number: 138