Sunspot Classifications & Solar Flare Prediction: Does machine learning improve upon Poisson-based prediction models?

Historically, McIntosh classifications of sunspots have been utilised for the prediction of solar flares, with modern day operational flare forecast services still reliant upon these classifications for their predictions. Here, building upon previous Poisson-based flare forecasting models that make use of Mcintosh classifications, a set of various machine learning (ML) techniques are applied to construct a set of new models to predict flares within a 24-hr period.

These ML algorithms are trained and tested using data from a range of independent solar cycle periods, cross-validation techniques are applied and the relative performance of each algorithm is compared. In order to make a direct comparison to Poisson-based forecasts, skill scores are calculated and the performance of each model is presented, results showing that the ML models perform well across multiple metrics. The implications these results have when compared with the previous Poisson-based approach are discussed as well as the problem of solar cycle dependence. Additionally, an exploration of the importance of the individual features (i.e., McIntosh components) on the performance of each prediction model and their physical implications are presented.