Solar Flare Prediction and Feature Selection Using a Light-Gradient-Boosting Machine Algorithm

Solar flares are among the most severe space-weather phenomena, and they have the capacity to generate radiation storms and radio disruptions on Earth. The accurate prediction of solar-flare events remains a significant challenge, requiring continuous monitoring and identification of specific featur...

Full description

Saved in:
Bibliographic Details
Published in:Solar physics Vol. 298; no. 11; p. 137
Main Authors: Vysakh, P. A., Mayank, Prateek
Format: Journal Article
Language:English
Published: Dordrecht Springer Netherlands 01.11.2023
Springer Nature B.V
Subjects:
Algorithms
Astrophysics and Astroparticles
Atmospheric Sciences
Labels
Machine learning
Physics
Physics and Astronomy
Sampling techniques
Solar activity
Solar flares
Solar physics
Space Exploration and Astronautics
Space Sciences (including Extraterrestrial Physics
Temporal resolution
Weather
Feature selection
Forecasting
Machine learning
Solar flare
ISSN:0038-0938, 1573-093X
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Solar flares are among the most severe space-weather phenomena, and they have the capacity to generate radiation storms and radio disruptions on Earth. The accurate prediction of solar-flare events remains a significant challenge, requiring continuous monitoring and identification of specific features that can aid in forecasting this phenomenon, particularly for different classes of solar flares. In this study, we aim to forecast C- and M-Class solar flares utilising a machine-learning algorithm, namely the Light Gradient Boosting Machine. We have utilised a dataset spanning nine years, obtained from the Space-weather Helioseismic and Magnetic Imager Active Region Patches (SHARP), with a temporal resolution of 1 h. A total of 37 flare features were considered in our analysis, comprising of 25 active-region parameters and 12 flare-history features. To address the issue of class imbalance in solar-flare data, we employed the Synthetic Minority Over-sampling Technique (SMOTE). We used two labelling approaches in our study: a fixed 24-h window label and a varying window that considers the changing nature of solar activity. Then, the developed machine-learning algorithm was trained and tested using forecast-verification metrics, with an emphasis on evaluating the true skill statistic (TSS). Furthermore, we implemented a feature-selection algorithm to determine the most significant features from the pool of 37 features that could distinguish between flaring and non-flaring active regions. We found that utilising a limited set of useful features resulted in improved prediction performance. For the 24-h prediction window, we achieved a TSS of 0.63 (0.69) and an accuracy of 0.90 (0.97) for ≥C- (≥M)-Class solar flares.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0038-0938
1573-093X
DOI:10.1007/s11207-023-02223-5