Source-linked AI summary
Solar Flare Prediction Model with Three Machine-Learning Algorithms Using Ultraviolet Brightening and Vector Magnetogram
N. Nishizuka, K. Sugiura, Y. Kubo, M. Den, S. Watari, M. Ishii
TL;DR
The paper asks how to predict the maximum solar-flare class over the next 24 h and which machine-learning algorithm is most suitable. It builds a model from multiwavelength solar observations and active-region features, then compares three algorithms. With fully shuffled data, TSS exceeded 0.9, k-NN performed best, and previous flare activity was the most important feature group.
Problem
The study addresses the need to predict solar flares from large observational datasets and determine which machine-learning algorithm performs best.
Method
The model detects active regions in 2010–2015 SDO and GOES observations, extracts magnetic, UV, X-ray, and temporal features, and compares SVM, k-NN, and ERT.
Results
TSS exceeded 0.9 on the fully shuffled dataset, k-NN had the highest average performance, and previous flare activity ranked as the most important feature.
Takeaways & Limitations
Feature importance emphasizes prior flare activity, magnetic neutral-line structure, unsigned flux, UV-brightening area, and 24 h feature changes.
Takeaways & Limitations
Year-separated training and testing markedly reduced prediction scores, while excluding near-limb regions improved feature accuracy but was less representative of operational forecasting.
Abstract
from arXiv · showhide
We developed a flare prediction model using machine learning, which is optimized to predict the maximum class of flares occurring in the following 24 h. Machine learning is used to devise algorithms that can learn from and make decisions on a huge amount of data. We used solar observation data during the period 2010-2015, such as vector magnetogram, ultraviolet (UV) emission, and soft X-ray emission taken by the Solar Dynamics Observatory and the Geostationary Operational Environmental Satellite. We detected active regions from the full-disk magnetogram, from which 60 features were extracted with their time differentials, including magnetic neutral lines, the current helicity, the UV brightening, and the flare history. After standardizing the feature database, we fully shuffled and randomly separated it into two for training and testing. To investigate which algorithm is best for flare prediction, we compared three machine learning algorithms: the support vector machine (SVM), k-nearest neighbors (k-NN), and extremely randomized trees (ERT). The prediction score, the true skill statistic (TSS), was higher than 0.9 with a fully shuffled dataset, which is higher than that for human forecasts. It was found that k-NN has the highest performance among the three algorithms. The ranking of the feature importance showed that the previous flare activity is most effective, followed by the length of magnetic neutral lines, the unsigned magnetic flux, the area of UV brightening, and the time differentials of features over 24 h, all of which are strongly correlated with the flux emergence dynamics in an active region.
1. Introduction
Solar-flare prediction addresses a persistent physical puzzle using increasingly large observational datasets and machine-learning methods. This study compares three algorithms, incorporates UV brightening and vector-magnetogram information, and ranks feature importance.
- Solar flares remain a long-standing puzzle, with energy storage and triggering linked to photospheric flux emergence.
- Magnetogram-derived indicators include magnetic flux, field gradients, magnetic neutral-line length, effective field, and unsigned flux near neutral lines.
- Vector magnetograms capture dynamic magnetic-configuration changes near neutral lines and provide features used in machine-learning flare prediction.
- Preflare UV emission gradually increases for 2–3 h, and 1600 Å observations capture chromospheric emission associated with flare-related activity.
- The study compares support vector machines, k-nearest neighbors, and another machine-learning approach while adding UV brightening, vector magnetograms, and feature-importance rankings.
2. Overview of our Prediction Model
The prediction model combines multi-instrument solar observations, active-region tracking, feature extraction, and flare labeling. It uses data from 2010–2015 and addresses the imbalance between common negative events and rarer flares.
- The pipeline downloads magnetogram, vector-magnetogram, UV-filtergram, and soft-X-ray observations from SDO and GOES.
- Active regions are detected in full-disk line-of-sight magnetograms, tracked through time, and used to calculate multiwavelength features and attach flare labels.
- Supervised learning predicts the maximum flare class during the following 24 h from observations sampled at 1 h cadence.
- The dataset spans June 2010–December 2015 and contains 29 X-class and 433 M-class disk flares.
- Negative events greatly outnumber positive events, making flare prediction an imbalanced classification problem.
3.1. Detection of ARs
Active regions are detected from hourly full-disk line-of-sight magnetograms using thresholded magnetic-field pixels, square grouping, and temporal tracking. Limb regions are excluded because projection and vector-magnetogram quality complicate their analysis.
- Approximately 10^5 full-disk magnetograms were processed at 1 h cadence to detect active regions, whose coordinates were transferred to other wavelengths.
- Pixels with absolute magnetic-field strength above Bth=140 G are grouped into 80”×80” squares, and overlapping squares are merged into whole-region boxes.
- Limb regions are neglected because projection effects obscure magnetic structure and vector-magnetogram data quality is poor near the limb.
- The detection procedure tracks active regions along solar rotation and assigns persistent IDs, yielding 11700 detected regions during 2010–2015.
- The threshold-based definition includes faint quasi-active regions, differing from NOAA’s definition and reducing missed small flares outside NOAA regions.
- Figures show the progression from full-disk magnetograms and thresholded points to grouped boxes, alongside multiwavelength active-region views.
3.2. Extraction of Solar Features
The model extracts magnetic, vector-field, UV-brightening, and X-ray-history features from each active region. These features include structural measures, time variability, chromospheric brightening, and prior flare activity.
- Features are extracted from line-of-sight and vector magnetograms, GOES X-ray data, and 1600 Å UV observations, with UV chromospheric brightening introduced as a new feature source.
- Line-of-sight features include active-region area, magnetic-field statistics, unsigned flux, field gradients, and counts and lengths of magnetic neutral lines.
- Vector-magnetogram features include vertical current, current helicity, Lorentz force, and mean total-field gradient, together with time derivatives.
- Time derivatives are calculated over 24 h, 12 h, and 2 h to track active-region variability across multiple timescales.
- UV features measure maximum intensity, brightening area, and total brightening intensity from AIA 1600 Å filtergrams.
- The brightening area is defined using a 700 photons cm^-5 s^-1 pix^-1 intensity threshold, while total intensity integrates values above that threshold.
- GOES X-ray measurements provide prior-flare indicators including one-day maximum intensity and counts of earlier flares in an active region.
3.3. Classification by Machine Learning
The paper compares SVM, k-NN, and ERT classifiers for flare prediction, with each algorithm using a distinct supervised-learning strategy and ERT additionally supporting feature-importance analysis.
- The model compares support vector machine (SVM), k-nearest neighbors (k-NN), and extremely randomized trees (ERT) as flare-class classifiers.
- SVM: SVM determines a maximum-margin hyperplane from labeled training samples by solving an optimization problem.
- k-NN: k-NN assigns an object to the most popular class among its nearest k training instances, with this study setting k=1.
- ERT: ERT randomly selects candidate features and thresholds, then chooses the best randomly generated threshold for each split.
- ERT: ERT feature importance averages weighted impurity decreases across tree nodes and trees, using the Gini index in this study.
3.4. Standardization, Evaluation, and Cross-Validation
The study standardizes heterogeneous solar features, evaluates forecasts with TSS, and uses repeated shuffle-and-split cross-validation to assess validity while accounting for imbalanced flare samples.
- Standardization: Z-value standardization is required because extracted solar features have different units and scales, and it strongly affects prediction accuracy.
- Standardization: Each standardized feature is expressed in standard deviations from its mean, with large-scale variables logarithmically transformed first.
- Data splitting: The 1 h-cadence database was fully shuffled and split 7:3 into training and testing datasets, with labels marking samples within 24 h of X/M-class flares.
- Evaluation: TSS measures hit rate relative to false alarm rate and ranges from -1 to +1, with 0 indicating no skill and 1 perfect prediction.
- Evaluation: TSS was selected because flare prediction is imbalanced, with negative events much more frequent than positive events.
- Cross-validation: Shuffle-and-split cross-validation used a 7:3 training-validation ratio and was repeated 10 times, with average results reported.
4. Prediction Results
The prediction results show that k-NN achieved the strongest TSS among the three algorithms, while feature rankings emphasize flare history, magnetic structure, UV brightening, and 24 h feature changes.
- Algorithm comparison: For X-class flares, TSS was 0.91±0.03 for k-NN, 0.88±0.03 for SVM, and 0.82±0.04 for ERT.
- Algorithm comparison: For ≥M-class flares, TSS was 0.912±0.005 for k-NN, 0.870±0.007 for SVM, and 0.71±0.02 for ERT.
- Algorithm comparison: k-NN had the highest TSS for both flare classes, followed by SVM and then ERT, although ERT had the smallest false-positive count.
- Feature contribution: Without previous flare activity and selected X-ray-history features, X-class TSS was 0.91±0.02 for k-NN, 0.86±0.02 for SVM, and 0.62±0.03 for ERT.
- Feature importance: The most effective features included total X/M-class flare history, previous-day maximum X-ray intensity, magnetic neutral-line length, unsigned magnetic flux, and UV brightening area.
- Feature importance: UV brightening area and total intensity were relatively important, while 24 h feature derivatives were effective and 12 h and 2 h derivatives were ineffective.
- Detection area: Excluding near-limb active regions improved TSS for all algorithms; for disk-center X-class prediction, TSS was 0.94±0.02 for k-NN, 0.92±0.03 for SVM, and 0.88±0.06 for ERT.
5. Summary and Discussion
The model achieved TSS above 0.9 on fully shuffled data, with k-NN outperforming SVM and ERT. Feature rankings emphasized prior flare activity, magnetic neutral-line structure, unsigned flux, UV brightening, and 24-hour feature changes, while performance depended on data splitting and detection area.
- TSS exceeded 0.9 on the fully shuffled dataset, and k-NN achieved the highest average performance among the three algorithms.The model used standardized data and optimization, including Manhattan distance, to improve prediction accuracy.
- The model appeared to outperform human operations, whose TSS values were 0.21 for X-class and 0.50 for ≥M-class forecasts at NICT.SIDC achieved a TSS of 0.34 for ≥M-class flares during 2004–2012.
- Previous flare activity ranked as the most important feature, followed by magnetic neutral-line configurations, unsigned flux, and UV-brightening area.The importance ranking directly highlighted flare history and maximum prior-day X-ray intensity, while the study introduced chromospheric brightening as a useful indicator.
- Twenty-four-hour feature derivatives were useful for prediction, whereas 12-hour and 2-hour derivatives were not.The authors relate the 24-hour timescale to magnetic relaxation and reconnection processes discussed in prior work.
- Prediction scores depended strongly on dataset separation and detection area, with yearly separation producing a more difficult evaluation than fully shuffled splitting.Excluding near-limb regions increased accuracy, although including them is more realistic for operational forecasting.