Abstract
This study explores the integration of solar north and south polar field strengths into machine learning models to enhance solar flare prediction. By supplementing local active region data with global polar field information, we provide predictors with a broader context of the solar cycle. We evaluate four machine learning architectures, including a novel probabilistic Mixture of Experts (MOE) model designed to effectively leverage polar field data. Our results demonstrate that incorporating these global features significantly improves prediction performance, with the Heidke Skill Score (HSS2) increasing by up to 10.1%.
Introduction
Solar flares represent sudden, intense bursts of radiation originating from magnetic energy release in the Sun's atmosphere. These events can have profound impacts on Earth's technological infrastructure, including satellite operations and power grids. Accurate prediction is critical but remains challenging due to the inherent rarity of high-magnitude (M and X class) flares.
Traditional prediction models primarily rely on "local" magnetic field characteristics within specific Active Regions (ARs). While effective, these models often neglect the "global" magnetic state of the Sun, which is governed by the 11-year solar cycle. This paper bridges the gap between long-term solar cycle dynamics and short-term flare forecasting by incorporating polar field data as a predictive feature.
Methods
We investigated four machine learning algorithms to assess the impact of global polar field data:
- Logistic Regression (LR): Serves as a baseline for linear classification.
- Multi-layer Perceptron (MLP): A standard feed-forward neural network to capture non-linear relationships.
- Recurrent Neural Network (RNN): Specifically an LSTM model, which is the current state-of-the-art for processing temporal sequences in flare prediction.
- Mixture of Experts (MOE): Our primary architectural contribution. This model employs a gating function that utilizes polar field data to weigh the outputs of specialized sub-models (experts). This approach allows the system to adapt its predictive logic based on the current phase of the solar cycle.
To mitigate noise in the polar field measurements, a Kalman filter was applied to provide a smoothed, more reliable estimate of the magnetic state over time.
Datasets
The experimental framework utilized data spanning from 2010 to 2020, encompassing the majority of Solar Cycle 24. Three primary sources were integrated:
- SHARP (Space-weather HMI Active Region Patches): Provides local magnetic parameters such as flux, energy, and complexity from the SDO/HMI instrument.
- GOES Flare List: Used for ground-truth labeling, focusing on a binary classification between "Strong" (M/X class) and "Weak" (B class) flares.
- SPF (Solar Polar Fields): Mean radial magnetic field strengths from the Sun’s poles (latitudes ±60° to ±90°), sourced from the Wilcox Solar Observatory and SDO/HMI.
Key Insights: The Global-Local Link
The central finding of this research is that the Sun's global magnetic state provides a vital context for local flare activity.
- Anti-Correlation as a Feature: Polar fields are approximately 180 degrees out of phase with sunspot activity. As active regions decay and their magnetic flux migrates toward the poles, they reinforce the polar fields. This physical relationship allows polar data to serve as a "signature" of the solar cycle phase.
- Solar Cycle Gating: The success of the MOE model suggests that different physical regimes may govern flare production during solar minimum versus maximum. By using polar fields to "gate" the model, we can effectively train specialized experts for these different phases.
Results
The inclusion of polar field data yielded consistent improvements across all tested models. The most notable result was a 10.1% increase in the Heidke Skill Score (HSS2) for the RNN model when using active-region-based data splitting. The novel MOE model also performed exceptionally well, showing a 9.9% improvement and achieving parity with more complex RNN architectures. These results indicate that explicitly modeling the solar cycle's influence is a powerful alternative to relying solely on deep temporal models.
Conclusion and Future Work
This study demonstrates that global solar features, specifically polar field strengths, are highly beneficial for short-term solar flare prediction. Future research will focus on extending this analysis across multiple solar cycles to ensure long-term robustness. Additionally, we aim to move beyond binary classification to predict specific flare classes (C, M, X) and explore the use of high-resolution magnetogram images within Convolutional Neural Networks (CNNs) to further enhance predictive accuracy.
Summary updated