![[2602.16264] Prediction of Major Solar Flares Using Interpretable Class-dependent Reward Framework with Active Region Magnetograms and Domain Knowledge](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.16264] Prediction of Major Solar Flares Using Interpretable Class-dependent Reward Framework with Active Region Magnetograms and Domain Knowledge
Summary
This article presents a novel supervised classification framework for predicting major solar flares using class-dependent rewards and deep learning models, achieving superior predictive capabilities.
Why It Matters
Understanding solar flares is crucial for mitigating their impacts on technology and space exploration. This research advances predictive models, potentially improving forecasting accuracy and response strategies for solar events, which can affect satellite operations and power grids.
Key Takeaways
- Introduces a class-dependent reward framework for flare prediction.
- CDR-Transformer outperforms traditional deep learning models when using knowledge-informed features.
- Sensitivity analysis shows that the predictive performance is stable across different reward choices.
Computer Science > Machine Learning arXiv:2602.16264 (cs) [Submitted on 18 Feb 2026] Title:Prediction of Major Solar Flares Using Interpretable Class-dependent Reward Framework with Active Region Magnetograms and Domain Knowledge Authors:Zixian Wu, Xuebao Li, Yanfang Zheng, Rui Wang, Shunhuang Zhang, Jinfang Wei, Yongshang Lv, Liang Dong, Zamri Zainal Abidin, Noraisyah Mohamed Shah, Hongwei Ye, Pengchao Yan, Xuefeng Li, Xiaojia Ji, Xusheng Huang, Xiaotian Wang, Honglei Jin View a PDF of the paper titled Prediction of Major Solar Flares Using Interpretable Class-dependent Reward Framework with Active Region Magnetograms and Domain Knowledge, by Zixian Wu and 16 other authors View PDF HTML (experimental) Abstract:In this work, we develop, for the first time, a supervised classification framework with class-dependent rewards (CDR) to predict $\geq$MM flares within 24 hr. We construct multiple datasets, covering knowledge-informed features and line-of sight (LOS) magnetograms. We also apply three deep learning models (CNN, CNN-BiLSTM, and Transformer) and three CDR counterparts (CDR-CNN, CDR-CNN-BiLSTM, and CDR-Transformer). First, we analyze the importance of LOS magnetic field parameters with the Transformer, then compare its performance using LOS-only, vector-only, and combined magnetic field parameters. Second, we compare flare prediction performance based on CDR models versus deep learning counterparts. Third, we perform sensitivity analysis on reward engineering for CDR ...
