Decreasing False-alarm Rates in CNN-based Solar Flare Prediction Using SDO/HMI Data

Abstract A hybrid two-stage machine-learning architecture that addresses the problem of excessive false positives (false alarms) in solar flare prediction systems is investigated. The first stage a convolutional neural network (CNN) model based on VGG-16 extracts features from temporal stack consecutive Solar Dynamics Observatory Helioseismic and Magnetic Imager magnetogram images to produce flaring probability. probability added feature vector derived magnetograms train an extremely randomized trees (ERT) second binary deterministic (flare/no-flare) 12 hr forecast window. To tune hyperparameters architecture, new evaluation metric introduced: “scaled True Skill Statistic.” It specifically large discrepancy between true positive rate highly unbalanced event training data sets. Through hyperparameter tuning maximize this metric, our drastically reduces by ?48% without significantly affecting (reduction ?12%), when compared with predictions first-stage CNN alone. This, turn, improves various traditional classification metrics sensitive positives, such as precision, F1, Heidke Score. end result more robust system could be combined current operational flare-forecasting methods. Additionally, using ERT-based feature-ranking mechanism, we show output ranked terms relevance.