Deep Neural Networks Applied to Solar Flare Prediction: Deep Flare Net (DeFN)

Abstract:

Solar flare prediction is one of our important tasks for space weather forecast. People have tried to reveal fundamental mechanisms of a flare and develop prediction methods. Currently, there are four methods: (i) empirical, (ii) statistical, (iii) numerical and (iv) machine-learning (ML) methods. Now it is a hot topic to apply ML techniques to flare predictions, and some models have succeeded in improving skill scores. The deep neural network (DNN) is a newly developed algorithm which shows the highest accuracy of prediction in general. In DNN models, Convolutional Neural Network (CNN) can automatically extract features from images and accelerated DNN applications, but it has a disadvantage of unexplainable.

Here, we introduce our solar flare prediction model using a DNN named Deep Flare Net (DeFN). This model can calculate the probability of flares occurring in the following 24 hr in each active region, which is used to determine the most likely maximum classes of flares via a two-class classification (≥M vs. <M class, or ≥C vs. <C class). From 3x10⁵ observation images taken by SDO during 2010–2015, we detected active regions and calculated 79 features for each region, to which we annotated labels of X-, M-, and C-class flares. We adopted the features used in Nishizuka et al. (2017) and added some features for operational prediction: coronal hot brightening at 131 Å (T≥10⁷ K) and the X-ray and 131 Å intensity data 1 and 2 hr before an image. For operational evaluation, we divided the database into two for training and testing: the data set in 2010–2014 for training, and the one in 2015 for testing. The DeFN model consists of deep multilayer neural networks formed by adapting skip connections and batch normalizations.

To statistically predict flares, the DeFN model was trained to optimize the skill score, i.e., the true skill statistic (TSS). As a result, we succeeded in predicting flares with TSS=0.80 for ≥M class flares and TSS=0.63 for ≥C class flares. Note that in usual DNN models, the prediction process is a black box. However, in the DeFN model, the features are manually selected, and it is possible to analyze which features are effective for prediction after evaluation. In this talk, we would like to discuss the feature ranking revealed by DNN model and also the dependence of the optimization methods by TSS and other skill scores.