Predicting Solar Energetic Particle Event Energy Spectra Using Machine Learning Methods

Abstract

The magnetic activity of the Sun can trigger various energy-releasing events, such as solar flares and coronal mass ejections (CMEs). These energy bursts can lead to acceleration and ejection of matter into interplanetary space, where they can be observed as solar energetic particle (SEP) events. On this thesis the focus is on predicting the proton peak energy fluxes observed during these events. It has been shown that the properties of the SEP event energy spectra are related to the properties of the associated phenomena, such as flares, CMEs and the solar wind conditions during the event. In the last years, machine learning (ML) models have also been applied to predicting SEP events. In their 2024 paper [1], Liu et al. research team presented an iterative decision tree model for predicting SEP event energy spectra. In this thesis our first goal was to recreate this ML model and then expand our approach to other classical machine learning methods. Using SEP event data with associated flare, CME and solar wind speed data we created machine learning models with ridge, K-nearest neighbor and decision tree regressors for predicting the SEP event energy spectra. In this thesis we also introduced feature cost analysis for the different input feature combinations. As the flare, CME and solar wind speed data are obtained from different instruments, we explored whether just some of these properties would be sufficient in predicting the SEP events. Our results indicate that the ridge and KNN regression models seem to be somewhat equal in prediction performance and overall better than the decision tree regression. The decision tree regressor performed very poorly in predicting the SEP energy spectra. However, the prediction performances across all model types were overall lacking. For the feature set cost analysis, the results show that models trained with flare strength parameters perform better than other models and models trained with only flare strength parameters seem to perform nearly as well as models with additional parameters. Based on these results, further research with more robust models and if possible, larger datasets, is encouraged.