Machine Learning as a Flaring Storm Warning Machine: Was a Warning Machine for the 2017 September Solar Flaring Storm Possible?

Abstract

Machine learning is now one of the methodologies of choice for flare forecasting, and supervised techniques, in both their traditional and deep versions, are becoming more frequently used for prediction in this area of space weather. Most studies assess the prediction effectiveness of machine-learning methods by computing confusion matrices, which are typically highly non-diagonal, particularly in applications concerning the forecasting of X-class flares. The present study suggests that the reliability of the outcomes of a supervised machine-learning method could be better assessed by using it as a warning machine, sounding binary alerts unrolled over time, and by comparing the number of alerts sounded by the machine in specific time windows with the number of events actually observed in those time windows. Indeed, when applied to the prediction of the events associated with the 2017 September solar storm, a hybrid LASSO algorithm was able to sound alerts every day a flare actually occurred; it also identified the corresponding flare class. In addition, the machine was able to predict with some accuracy a reliable proxy of the energy budget daily released by magnetic reconnection during the entire course of the storm. Finally, the analysis shows that the combination of sparsity-enhancing machine learning and feature ranking could allow the identification of the prominent role that the Ising energy played as an active region property in the forecasting process.