Detecting Giant Solar Flares Based on Sunspot Parameters Using Bayesian Networks

摘要

This paper presents the use of Bayesian Networks (BN) in a new area, the detection of solar flares. The paper describes how to learn a Bayesian Network (BN) using a set of variables representing sunspots parameters such that the BN can detect and classify solar flares. Giant solar flares happen in the Sun's atmosphere quite frequently and as a consequence they can affect Earth. The work described here shows the relationship between the learned networks and the causality expected by solar physicists. The data used for learning and cross validation experiments show that the network substructures are easy to learn and robust enough to predict solar flares. The systems presented here are capable of detecting the flares within 72 hours, while the current method used today does the same work within 24 hours in advance only. It is also shown that sunspot parameters change over time, so different networks can be learned and perhaps combined in order to build a robust forecast system.