We consider the flare prediction problem that distinguishes flare-imminent active regions that produce an M- or X-class flare in the future 24 hours, from quiet active regions that do not produce any flare within $\pm 24$ hours. Using line-of-sight magnetograms and parameters of active regions in two data products covering Solar Cycle 23 and 24, we train and evaluate two deep learning algorithms—CNN and LSTM—and their stacking ensembles. The decisions of CNN are explained using visual attribution methods. We have the following three main findings. (1) LSTM trained on data from two solar cycles achieves significantly higher True Skill Scores (TSS) than that trained on data from a single solar cycle with a confidence level of at least 0.95. (2) On data from Solar Cycle 23, a stacking ensemble that combines predictions from LSTM and CNN using the TSS criterion achieves significantly higher TSS than the “select-best” strategy with a confidence level of at least 0.95. (3) A visual attribution method called Integrated Gradients is able to attribute the CNN’s predictions of flares to the emerging magnetic flux in the active region. It also reveals a limitation of CNN as a flare prediction method using line-of-sight magnetograms: it treats the polarity artifact of line-of-sight magnetograms as positive evidence of flares.

Many current research efforts undertake the solar flare classification task using the Space-weather HMI Active Region Patch (SHARP) parameters as the predictors. The SHARP parameters are scalar quantities based on spatial average or integration of physical quantities derived from the vector magnetic field, which loses information of the two-dimensional spatial distribution of the field and related quantities. In this paper, we construct two new sets of spatial features to expand the feature set used for the flare classification task. The first set uses the idea of topological data analysis to summarize the geometric information of the distributions of various SHARP parameters. The second set utilizes tools coming from spatial statistics to analyze the vertical magnetic field component Br and summarize its spatial variations and clustering patterns. All features are constructed within regions near the polarity inversion lines (PILs) and classification performances using the new features are compared against the SHARP parameters (also along PIL). We found that using the new features can improve the skill of the flare classification model and new features tend to have higher feature importance, especially the spatial statistics features. This potentially suggests that even using a single channel Br, instead of all SHARP parameters, one can still derive strongly predictive features for flare classification.