Knowledge-Informed Deep Neural Networks for Solar Flare Forecasting

Recently, although various deep learning techniques have been applied to building space weather prediction models, a large amount of relevant prior knowledge of solar eruptions and magnetic properties is ignored during the model development. By integrating prior knowledge in flare production into the convolutional neural network (CNN) structures, we have developed a knowledge-informed deep neural network model aiming at forecasting solar flares. The line-of-sight magnetograms of Space-weather HMI Active Region Patches (SHARP) from May 2010 to December 2018 are selected. We have surveyed the relationships between solar flares and both the active region (AR) area and magnetic type. When integrating prior knowledge into the CNNs, three aspects are considered: (a) keeping the magnetic structure unchanged (data preprocessing) while filling SHARP magnetograms into squares, (b) grouping the data samples into two subsets according to different flare productivities (sample grouping), and (c) adding AR area as an extra input parameter to the CNN (extra input parameter implementation). Pure CNN model, Fusion model 1, and Fusion model 2 are built to forecast M-class or above flares in the next 48 hr, which involve data preprocessing, data preprocessing and sample grouping, and all the three aspects, respectively. Fusion model 2 that augments the most prior knowledge has the best performance. Our results imply that prior knowledge can play an important role in building deep learning flare forecasting models. In the future, adopting knowledge-informed deep neural networks will be an effective way to further improve the forecasting performance for other space weather events.