Prediction of Space Weather Events through Analysis of Active Region Magnetograms using Convolutional Neural Network

TL;DR

This work tackles the problem of forecasting space weather events—solar flares, geomagnetic storms, and CMEs—by exploiting active-region magnetograms from SDO/HMI and event timings from DONKI. A custom convolutional neural network trained on magnetograms captured 24 hours before events demonstrates that multi-class predictions are viable, achieving 90.27% accuracy and a high F1 score across classes, with strongest performance for solar flares and geomagnetic storms. The authors address data imbalance with SMOTE and provide thorough evaluation via precision, recall, and normalized confusion matrices, noting CME predictions remain challenging due to limited and imbalanced data. The study contributes a practical pipeline for magnetogram-based space weather forecasting and suggests avenues to improve CME prediction through more balanced datasets and time-series analyses, potentially enabling proactive mitigations for technology infrastructure.

Abstract

Although space weather events may not directly affect human life, they have the potential to inflict significant harm upon our communities. Harmful space weather events can trigger atmospheric changes that result in physical and economic damages on a global scale. In 1989, Earth experienced the effects of a powerful geomagnetic storm that caused satellites to malfunction, while triggering power blackouts in Canada, along with electricity disturbances in the United States and Europe. With the solar cycle peak rapidly approaching, there is an ever-increasing need to prepare and prevent the damages that can occur, especially to modern-day technology, calling for the need of a comprehensive prediction system. This study aims to leverage machine learning techniques to predict instances of space weather (solar flares, coronal mass ejections, geomagnetic storms), based on active region magnetograms of the Sun. This was done through the use of the NASA DONKI service to determine when these solar events occur, then using data from the NASA Solar Dynamics Observatory to compile a dataset that includes magnetograms of active regions of the Sun 24 hours before the events. By inputting the magnetograms into a convolutional neural network (CNN) trained from this dataset, it can serve to predict whether a space weather event will occur, and what type of event it will be. The model was designed using a custom architecture CNN, and returned an accuracy of 90.27%, a precision of 85.83%, a recall of 91.78%, and an average F1 score of 92.14% across each class (Solar flare [Flare], geomagnetic storm [GMS], coronal mass ejection [CME]). Our results show that using magnetogram data as an input for a CNN is a viable method to space weather prediction. Future work can involve prediction of the magnitude of solar events.

Figures (12)

• Figure 1: Magnetogram of the Sun. Courtesy of NASA/SDO and the AIA, EVE and HMI science teams.
• Figure 2: Visualization of overall methodology. Dates between solar flare and CME events were first matched using DONKI database. Then, labels and images were automatically uploaded into dataframe using Google AppsScript which obtained dates from SDO.
• Figure 3: Visualization of matching process for CME and Solar Flare classes. Google Sheet processes would determine whether a date from CME column would match with Solar Flare date.
• Figure 4: Dataframe example. Contains image file along with three classes, Flare, CME, and GMS, with binary labels where 1 represents event and 0 represents no event.
• Figure 5: Visualization of the SMOTE algorithm. The SMOTE algorithm Smote synthesizes data using real data from the minority class of the dataset. Courtesy of Analytics Vidhya.