Interpretable Deep Learning for Solar Flare Prediction

We propose to incorporate three interpretable methods, namely SHAP (SHapley Additive exPlanations), PDP (partial dependence plots) and Anchors, into a deep learning-based model, called SolarFlareNet, for operational flare forecasting. SolarFlareNet takes as input a sample of SHARP (Space-weather HMI Active Region Patches) magnetic parameters and predicts as output whether a solar flare would occur within the next 24 hours. We analyze flare events that occurred from May 2010 to December 2022 using the Geostationary Operational Environmental Satellite's X-ray flare catalogs and construct a database of flares with identified active regions in the catalogs. This database, together with the SHARP magnetic parameters, is used to train and test the SolarFlareNet model. Our experimental results describe the use of the three proposed methods (SHAP, PDP, and Anchors) to interpret the SolarFlareNet model and demonstrate the effectiveness of the methods.