Solar Flare Forecasting Using Hybrid Neural Networks

Abstract Solar flares are one of the most intense solar activities, the result of a sudden large-scale release of magnetic energy in the form of electromagnetic radiation and energetic particles. Intense solar flares can severely threaten communication and navigation systems, oil pipelines, and power grids on Earth. Therefore, it is crucial to establish highly accurate solar flare prediction models to enable humans to anticipate solar flare eruptions in advance, thereby reducing human and economic losses. In this paper, we utilized the solar active region (AR) magnetogram provided by the Solar Dynamics Observatory’s Helioseismic and Magnetic Imager and the associated feature parameters of the magnetic field; specifically, the feature vectors of the magnetic field’s spatial structure characteristics and the magnetic field feature parameters are fused to predict solar flares. We built two solar flare prediction models based on a combination of convolutional neural networks (CNN) and a temporal convolutional network (TCN), called CNN-TCN, and predicted whether a ≥C- or ≥M-class flare event would erupt in ARs in the next 24 hr, respectively. Then, after training and testing our model, we focused on the true skill statistic (TSS). Through the model superiority discussion, the model obtained high average TSS values, with the ≥C and ≥M models achieving TSS scores of 0.798 ± 0.032 and 0.850 ± 0.074, respectively, suggesting that our models have good forecasting performance. We speculate that some key features automatically extracted by our model may not have been previously identified, and these features could provide important clues for studying the mechanisms of flares.