Forecasting solar energetic particles using multi-source data from solar flares, CMEs, and radio bursts with machine learning approaches.

This study presents a consistent method to the inherently imbalanced problem of predicting solar energetic particle (SEP) events, using a variety of datasets that include solar flares, coronal mass ejections (CMEs), and radio bursts. We applied several machine learning (ML) methods, including Random Forests (RF), Decision Trees (dtree), and Support Vector Machines (SVM) with both linear (linSVM) and nonlinear (svm) kernels. To assess model performance, we used standard metrics such as Probability of Detection (POD), False Alarm Rate (FAR), True Skill Statistic (TSS), and Heidke Skill Score (HSS). Our results show that the RF model consistently outperforms the other algorithms across datasets containing flares, CMEs, and radio bursts. For the sweep frequency dataset, RF achieved a POD of [Formula: see text], a FAR of [Formula: see text], a TSS of [Formula: see text],and a HSS of [Formula: see text]). For the fixed-frequency dataset, RF produced a POD of [Formula: see text], a FAR of [Formula: see text], a TSS of [Formula: see text] ,and a HSS of [Formula: see text]). Key features for SEP prediction include CME linear speed and angular width across both datasets. For sweep frequency, flare intensity and integral soft X-ray (SXR) flux are crucial, while for fixed frequency, the rise time and duration of radio bursts at 1415 MHz are significant.