Predicting Solar Flares using CNN and LSTM on Two Solar Cycles of Active
Region Data
Abstract
One major challenge of solar flare prediction with machine learning
methods is the scarcity of large flares. This issue of low positive
sample size is even more severe for data observed in the relatively weak
Solar Cycle 24, for example, the SHARPs data product. This partly
hampers the successful application of deep learning methods, especially
those dealing with high-dimensional spatial and/or temporal data. By
joining SHARPs with Space-Weather MDI Active Region Patches (SMARPs), a
new data product derived from observations in Solar Cycle 23, we are
able to obtain a fused dataset with nearly tripled positive samples. We
evaluated two deep learning methods, LSTM and CNN, using the selected
parameter sequences and image snapshots in the fused dataset. Experiment
results show that the two models trained on the fused dataset achieve
better or equivalent test set performance than those trained on a single
solar cycle. In addition, we demonstrate the improvement of the
performance of the stacking ensemble that combines LSTM and CNN. We
provided interpretation to CNN using modern visual attribution methods
in computer vision. The results show that CNN is able to identify
flare-related signatures in magnetograms.