Abstract
In this study we apply Support Vector Machine (SVM) to the prediction of geo-effective halo coronal mass ejections (CMEs). The SVM, which is one of machine learning algorithms, is used for the purpose of classification and regression analysis. We use halo and partial halo CMEs from January 1996 to April 2010 in the SOHO/LASCO CME Catalog for training and prediction. And we also use their associated X-ray flare classes to identify front-side halo CMEs (stronger than B1 class), and the Dst index to determine geo-effective halo CMEs (stronger than -50 nT). The combinations of the speed and the angular width of CMEs, and their associated X-ray classes are used for input features of the SVM. We make an attempt to find the best model by using cross-validation which is processed by changing kernel functions of the SVM and their parameters. As a result we obtain statistical parameters for the best model by using the speed of CME and its associated X-ray flare class as input features of the SVM: Accuracy=0.66, PODy=0.76, PODn=0.49, FAR=0.72, Bias=1.06, CSI=0.59, TSS=0.25. The performance of the statistical parameters by applying the SVM is much better than those from the simple classifications based on constant classifiers.
Introduction
Coronal Mass Ejections (CMEs) are massive eruptions of solar plasma and magnetic fields. When these eruptions are directed toward Earth—appearing as "halo" CMEs in coronagraphs—they can trigger geomagnetic storms, disrupting critical infrastructure like satellites, GPS, and power grids. Predicting which CMEs will be "geo-effective" (causing a storm) remains a significant challenge. Traditional methods often rely on simple threshold-based "constant classifiers" (e.g., "if speed > 500 km/s, then storm"), which frequently lead to high false-alarm rates or missed events. This research applies Support Vector Machines (SVM) to identify non-linear patterns in solar data, aiming to improve the reliability of space weather forecasts.
Methodology
The researchers implemented a Support Vector Machine (SVM) framework to classify CMEs as geo-effective or non-geo-effective. The process involved scaling input features, training the model on historical data, and using the resulting model to predict outcomes for new events.
Key technical steps included:
- Feature Selection: Combinations of CME linear speed (S), angular width (W), and associated X-ray flare class (X) were evaluated.
- Preprocessing: Features were scaled to a [-1, 1] range to prevent magnitude bias. Flare classes (B, C, M, X) were converted to numerical values (1–4).
- Optimization: The Radial Basis Function (RBF) kernel was found to be the most effective. Hyperparameters $C$ and $\gamma$ were tuned via grid-search and cross-validation.
Data Selection and Criteria
The study integrated data from the SOHO/LASCO CME Catalog, GOES X-ray flare lists, and the Dst index from the World Data Center in Kyoto. The dataset was filtered to focus on "front-side" halo CMEs likely to impact Earth.
A CME was labeled geo-effective if the Disturbance Storm Time (Dst) index dropped below -50 nT within a 24-hour window of the predicted arrival time.
Results and Discussion
The most successful SVM model utilized a combination of Speed and X-ray Flare Class (SX) with an RBF kernel. This configuration highlights that the initial eruption's energy (flare) is a vital complement to the CME's kinetic energy (speed) for determining Earth-side impact.
The optimized model achieved:
- Accuracy: 0.66
- Probability of Detection (PODy): 0.76 (76% of storms correctly identified)
- Critical Success Index (CSI): 0.59
This represents a substantial improvement over simple constant classifiers. For instance, a baseline speed threshold of ≥ 500 km/s yields a CSI of only 0.39; the SVM approach provides a roughly 50% improvement in this metric. The superior performance of the RBF kernel suggests that the relationship between solar parameters and geomagnetic impact is inherently non-linear and localized in the feature space.
Conclusion and Future Directions
While the SVM significantly improves detection rates, the False Alarm Ratio (0.72) remains high, reflecting the inherent complexity of space weather. Future improvements could include incorporating the solar source location (latitude/longitude) of the eruption, which was omitted in this study due to data sparsity. Integrating these machine learning models into real-time automated pipelines could provide essential early warnings for satellite and grid operators.
Summary updated