SHORT-TERM SOLAR FLARE LEVEL PREDICTION USING A BAYESIAN NETWORK APPROACH

Abstract

A Bayesian network approach for short-term solar flare level prediction has been proposed based on three sequences of photospheric magnetic field parameters extracted from Solar and Heliospheric Observatory/Michelson Doppler Imager longitudinal magnetograms. The magnetic measures, the maximum horizontal gradient, the length of neutral line, and the number of singular points do not have determinate relationships with solar flares, so the solar flare level prediction is considered as an uncertainty reasoning process modeled by the Bayesian network. The qualitative network structure which describes conditional independent relationships among magnetic field parameters and the quantitative conditional probability tables which determine the probabilistic values for each variable are learned from the data set. Seven sequential features—the maximum, the mean, the root mean square, the standard deviation, the shape factor, the crest factor, and the pulse factor—are extracted to reduce the dimensions of the raw sequences. Two Bayesian network models are built using raw sequential data (BN_R) and feature extracted data (BN_F), respectively. The explanations of these models are consistent with physical analyses of experts. The performances of the BN_R and the BN_F appear comparable with other methods. More importantly, the comprehensibility of the Bayesian network models is better than other methods.