The study aims to develop a system that can predict instances of space weather events, including solar flares, geomagnetic storms, and coronal mass ejections, using machine learning techniques. The researchers used data from the NASA Solar Dynamics Observatory (SDO) and the NASA Space Weather Database Of Notifications, Knowledge, Information (DONKI) to compile a dataset of solar magnetograms and the corresponding space weather event timings.
The researchers first matched the dates of solar flares and coronal mass ejections, then extracted magnetograms from the SDO for the matched dates, as well as dates with only one type of event. They also included dates with no observed events. The dataset was then resampled using the Synthetic Minority Over-sampling Technique (SMOTE) to address the imbalance between event and non-event cases.
The researchers designed a custom architecture convolutional neural network (CNN) model to process the magnetogram inputs and predict the occurrence of the three space weather events. The model was trained for 20 epochs with an early stopping mechanism to prevent overfitting.
The model achieved strong performance, with an accuracy of 90.27%, precision of 85.83%, recall of 91.78%, and an average F1 score of 92.14% across the three classes. The confusion matrices further revealed that the model performed well in predicting solar flares and geomagnetic storms, but had some challenges in accurately predicting coronal mass ejections, likely due to the imbalance in the training data.
The researchers conclude that using magnetogram data as an input for a CNN is a viable method for space weather prediction. Future work could involve predicting the magnitude of the solar events and incorporating time-series data to improve the reliability of the forecasts.
Vers une autre langue
à partir du contenu source
arxiv.org
Questions plus approfondies