Application of deep reinforcement learning to solar major flare forecasting

In this study, we have applied deep reinforcement learning to solar major flare forecast. For this, we use full-disk magnetograms at 00:00 UT from Solar and Heliospheric Observatory/Michelson Doppler Imager (1996 August - 2010 December) and Solar Dynamics Observatory/Helioseismic and Magnetic Imager (2011 January - 2019 December), and Geostationary Operational Environmental Satellite X-ray flare data. The solar cycle 23 and the solar cycle 24 data are used for training and test, respectively. Our deep learning flare forecast model based on the Convolutional Neural Network (CNN) predicts Yes or No" of daily flare occurrence for M- and X-class. We adopt a deep Q-learning network (DQN), a method of deep reinforcement learning, for model training. We test the DQN model performance using various reward guidance and compare them with the other models based on different methods, in view of various skill scores such as true skill statics(TSS) and Applemans skill score(ApSS). Our results show that the reinforcement learning could improve flare model performance under the guidance of proper rewards.