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

The repository includes the datasets and experimental code used in our research, along with descriptions of these files. In this study, we construct two kinds of datasets from distinct time periods, both comprising line-of-sight magnetograms and knowledge-informed features. The first kind of dataset is based on SHARP LOS magnetograms and knowledge-informed features from May 1, 2010, to February 13, 2022. The zip files named “10_Fold_SHARP_CV_Image.zip” and “10_Fold_SHARP_CV_Knowledge_informed” provide the ten-fold cross-validation set, normalized and split using the AR-based cross-validation method, for SHARP LOS magnetograms and knowledge-informed features, respectively. The second kind of dataset consists of magnetograms and knowledge-informed features from SHARP, HMI, and FMG corresponding to the same ARs, spanning from April 1, 2023, to July 21, 2024. The zip files named “23_24_SHARP_HMI_FMG_Image.zip” and “23_24_SHARP_HMI_FMG_Knowledge_informed” provide magnetograms and knowledge-informed features from SHARP, HMI, and FMG, respectively. The zip file named "CCMC_Compare.zip" provides the real-time solar flare forecast results from various models collected on the NASA/CCMC official website (https://ccmc.gsfc.nasa.gov/scoreboards/flare/) from April 1, 2023, to July 21, 2024, as well as the SHARP testing dataset we used, which is highly consistent with their time and AR numbers. The file named “Correlation_Alignment.py” provides the cross-correlation method we used to align magnetograms and extract the same regions as the FMG from each complete HMI magnetogram. The file named “Knowledge-informed_Features_Calculate.py” provides the method for calculating the 31 knowledge-informed features from the LOS magnetograms. The file named “CCMC_Metric_Calculate.py” provides the code for calculating categorial and probabilistic forecast metrics based on the NASA/CCMC solar flare forecast results. The zip file named "Visual.zip" provides visualizations for the CNN, CNN-BiLSTM, and ViT models, showing the regions of interest in the magnetograms that the models focus on. In this paper, we design seven currently popular deep learning models, including four image-based models (e.g., CNN, CNN-BiLSTM, CNN-BiLSTM-Attention, and ViT) and three knowledge-informed models (e.g., BiLSTM, BiLSTM-Attention and iTransformer), for predicting ≥M-class solar flares within 24 hr. Additionally, we use a traditional NN model based on knowledge-informed features as the baseline model for comparison with other deep learning models. The zip file named "Model.zip" contains the eight solar flare prediction models used in this study, including the model architectures, saved model parameters, and the evaluation metric results from testing.