A real-time solar flare forecasting system with deep learning methods

In this study, we develop five deep learning models, a Convolutional Neural Network (CNN) model, a CNN model with Squeeze-and-Excitation Attention(CNN-SE), a CNN model with Convolutional Block Attention Module (CNN-CBAM), a CNN model with Efficient Channel Attention (CNN-ECA), and a Vision Transformer (ViT) model, for predicting whether >= C or >= M-class solar flares occurring within 24 hours. We build a real-time forecasting system using these five models, which can achieve classification and probability forecasting. The 10-fold cross-validation sets are generated in chronological order using the full-disk magnetograms provided by the Solar Dynamics Observatory/Helioseismic and Magnetic Imager at 00:00 UT from May 1, 2010, to March 31, 2023. Then after training, validation, and testing our models, we compare the results with the true skill statistic (TSS) and Brier Skill Score (BSS) as assessment metrics. The major results are as follows: (1) There are no statistically significant differences in TSS and BSS performance between models with attention mechanisms and the CNN model. (2) For >= C-class flare prediction, the Recall of the ViT model reaches 0.833, significantly better than that of the CNN model. For >= M-class flare prediction, the Recall of the CNN-ECA and ViT models are 0.799 and 0.855, respectively, which are significantly higher than those of the CNN model. (3) We develop a full-disk solar flare prediction system that has been running since May 1, 2023. By December 31, all five models achieve a TSS of 0.984 for predicting >= C-class flares, with the CNN-SE model demonstrating a BSS of 0.939. For >= M-class flares, the CNN-SE model achieves a TSS of 0.304, while the BSS values for the CNN and CNN-SE models are 0.019 and 0.018, respectively. Additionally, the prediction performance for >= M-class flares on the testing set without No-flare class samples, is similar to that of real-time predictions, validating the good generation performance of the model in real-time forecasting.