Forecasting Major Flares Using Magnetograms and Knowledge-informed Features: A Comparative Study of Deep Learning Models with Generalization to Multiple Data Products

In this study, we construct two kinds of data sets from distinct time periods, both comprising line-of-sight magnetograms and knowledge-informed features. We develop eight models for forecasting ≥M-class flares within 24 hr, including the image-based convolutional neural network (CNN), CNN-BiLSTM, CNN-BiLSTM-Attention, and Vision Transformer models, as well as the knowledge-informed neural network, BiLSTM, BiLSTM-Attention, and iTransformer models. We analyze the importance of knowledge-informed features by assessing categorical and probabilistic performance using the true skill statistic (TSS) and the Brier skill score (BSS), respectively. This is the first time the iTransformer has been applied to flare forecasting. Subsequently, we compare the forecasting performance of the eight models. Then, we investigate the generalization ability of the models across three different data products. Finally, we fairly compare the forecasting performance of iTransformer with that of the currently advanced NASA/CCMC models. The major results are as follows. (1) The R_VALUE feature consistently shows the best performance in both categorical and probabilistic forecasting for the knowledge-informed models. (2) The iTransformer yields the highest forecasting performance, with TSS and BSS scores of 0.768 ± 0.072 and 0.513 ± 0.063, respectively. The knowledge-informed deep learning models consistently outperform image-based models. (3) The three image-based models demonstrate good generalization performance in categorical forecasting on Space-weather Helioseismic and Magnetic Imager (HMI) Active Region Patches (SHARP), HMI, and Full-Disk Magnetograph (FMG), while the two knowledge-informed models exhibit excellent generalization performance on SHARP and HMI. This is the first time that FMG magnetograms and knowledge-informed features are used for flare forecasting. Additionally, the five models also demonstrate strong generalization ability on SHARP across different time periods. (4) The iTransformer exhibits superior forecasting performance compared to NASA/CCMC.

Abstract licence: CC BY

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.