Solar flare forecasting model based on automatic feature extraction

Abstract

Solar flares are outbursts in the solar atmosphere resulting from sudden release of magnetic energy. The associated high energy particles and radiation threaten the safety of astronauts, reduce the lifetime of satellites, disturb the radio communications and degrade the precision of Global Positioning System. The radiation reaches the Earth about 8 min, and high energy particles take about 30 min to reach the earth after a solar flare. So solar flare forecasting is critical for providing enough time to respond to the space weather effects. Up to now, many statistical and machine learning methods are used to build a solar flare forecasting model. A machine learning based solar flare forecasting model normally requires solar physicists to design a feature extractor which can transform the observational images of active regions into physical features, and then the relationships between the features and the solar flares are discovered by the machine leaning algorithm. The priori knowledge of the solar physicists is added into the solar flare forecasting model by designing the feature extractor. For most of the machine learning methods, the hard part is what kinds of features should be extracted from the raw data. Considerable solar physicists spend a lot of time extracting the physical parameters from observational data of active regions. Deep learning method, which removes this manual step, can automatically discover useful patterns from the raw data and build a forecasting model. Instead of designing the feature extractor by solar physicists, we learn a solar flare forecasting model from magnetogram pixels by using deep learning method. We use Caffe, which is a deep learning framework developed by the Berkeley Vision and Learning Center, to build a convolutional neural network for solar flare forecasting. In order to compare the performance of proposed forecasting model with that of the forecasting model built by using traditional machine learning method, we build the other solar flare forecasting model based on the same dataset. In the traditional forecasting model, physical parameters designed by the solar physicist are extracted from the magnetogram of active regions, and then these parameters are fed to the forecasting model. We build a traditional forecasting model by multilayer neural networks. For convenience, the solar flare forecasting model built by using deep learning method is called deep model, and the solar flare forecasting model built by using the traditional multilayer neutral networks is called traditional model. Using the same testing data, the performances of the deep model and the traditional model are evaluated and compared. We find that the performance of the deep model is little better than that of the traditional model. The results confirm that the deep model can automatically learn solar flare forecasting features from magnetograms of active regions. This is our first time to automatically learn the forecasting patterns for solar flares from raw data instead of designing the physical patterns by solar physicists. The effectiveness of the deep learning method for the solar flare forecasting is validated. In the future, the deep learning method can be used to automatically discover the solar flare forecasting patterns from the vector magnetograms or the extreme ultraviolet images of active regions.