Enhancing Solar Flare Prediction with Innovative Data-Driven Labels

Jinsu Hong, Anli Ji, Chetraj Pandey, Berkay Aydin

Abstract: Solar flare prediction presents a significant challenge in space weather forecasting. Currently, existing solar flare prediction tools heavily rely on the GOES classification system. These tools commonly use the maximum X-ray flux measurement within a specific prediction window, often set at 24 hours, as a basis for labeling instances. However, the background X-ray flux experiences considerable fluctuations throughout the solar cycle, leading to misleading outcomes during solar minimum and an increase in false alarms. To address this issue, we propose a new set of solar flare intensity labels computed from GOES X-ray flux, with the aim of enhancing the accuracy of flare prediction methods. Our approach involves innovative labeling methods that take into account relative increases and cumulative measurements across prediction windows. In this paper, we introduce the concept of the ‘relative X-ray flux increase’ and provide an explanation of how this metadata is derived. Additionally, we present new cumulative indices and data-driven categorical labels specifically designed for active region-based and full-disk flare prediction models. We also assess the feasibility of integrating our new labels into established solar flare prediction models. Our findings demonstrate that these data-driven labels can be valuable supplements to existing techniques and their integration has the potential to enhance the effectiveness of solar flare prediction tools.