Hybrid deep convolutional neural network with one-versus-one approach for solar flare prediction

We propose a novel hybrid Convolutional Neural Network (CNN) model with one-versus-one approach to forecast solar flare occurrence with the outputs of four classes (No-flare, C, M, and X) within 24 h. We train and test our model using the same data sets as in Zheng, Li & Wang, and then compare our results with previous models using the true skill statistic (TSS) as primary metric. The main results are as follows. (1) This is the first time that the CNN model in conjunction with one-versus-one approach is used in solar physics to make multiclass flare prediction. (2) In the four-class flare prediction, our model achieves quite high mean scores of TSS = 0.703, 0.489, 0.432, and 0.436 for No-flare, C, M, and X class, respectively, which are much better than or comparable to those of previous studies. In addition, our model obtains TSS scores of 0.703 ± 0.070 for ≥C-class and 0.739 ± 0.109 for ≥M-class predictions. (3) This is the first attempt to open the black-box CNN model to study the visualization of feature maps for interpreting the prediction model. Furthermore, the visualization results indicate that our model pays attention to the regions with strong gradient, strong intensity, high total intensity, and large range of the intensity in high-level feature maps. The median gradient and intensity, the total intensity, and the range of the intensity for high-level feature maps increase approximately with the increase of flare level.