Abstract
One of the limiting factors in training data-driven, rare-event prediction algorithms is the scarcity of the events of interest resulting in an extreme imbalance in the data. There have been many methods introduced in the literature for overcoming this issue; simple data manipulation through undersampling and oversampling, utilizing cost-sensitive learning algorithms, or by generating synthetic data points following the distribution of the existing data. While synthetic data generation has recently received a great deal of attention, there are real challenges involved in doing so for high-dimensional data such as multivariate time series. In this study, we explore the usefulness of the conditional generative adversarial network (CGAN) as a means to perform data-informed oversampling in order to balance a large dataset of multivariate time series. We utilize a flare forecasting benchmark dataset, named SWAN-SF, and design two verification methods to both quantitatively and qualitatively evaluate the similarity between the generated minority and the ground-truth samples. We further assess the quality of the generated samples by training a classical, supervised machine learning algorithm on synthetic data, and testing the trained model on the unseen, real data. The results show that the classifier trained on the data augmented with the synthetic multivariate time series achieves a significant improvement compared with the case where no augmentation is used. The popular flare forecasting evaluation metrics, TSS and HSS, report 20-fold and 5-fold improvements, respectively, indicating the remarkable statistical similarities, and the usefulness of CGAN-based data generation for complicated tasks such as flare forecasting.
Introduction
Predicting solar flares is a high-stakes challenge for space weather forecasting, as these events can disrupt power grids, GPS systems, and satellite communications. However, flares are statistically rare, leading to extreme class imbalance in datasets like SWAN-SF, where "no-flare" instances outnumber actual flares by approximately 50 to 1. This imbalance often causes machine learning models to become biased toward the majority class, failing to predict the critical rare events. This study addresses this gap by utilizing Conditional Generative Adversarial Networks (CGANs) to generate synthetic, high-quality multivariate time series (MVTS) data to balance the training set.
Methodology: CGAN Architecture
The researchers designed a CGAN architecture tailored for temporal data. Unlike standard GANs, the conditional component allows the model to specifically generate samples belonging to the rare "flare" class.
Key technical features include the integration of Long Short-Term Memory (LSTM) layers as the backbone for both the Generator and Discriminator. This choice is crucial for learning the complex temporal "flow" of magnetic field parameters leading up to a flare. The model was trained on four critical magnetic parameters (TOTUSJH, ABSNJZH, SAVNCPP, and TOTBSQ) known to be highly relevant to solar activity.
Evaluation and Statistical Mimicry
To verify the quality of the synthetic data, the authors compared the statistical distributions of the generated samples against real ground-truth data. They focused on features such as the mean, median, and standard deviation of the magnetic flux parameters.
As training progressed, the synthetic data's statistical "fingerprint" began to overlap almost perfectly with the real data. The researchers identified that the most realistic samples were generated between 201 and 250 training epochs, where the adversarial accuracy reached an equilibrium, suggesting the Generator had become highly effective at mimicking real solar flare signatures.
Key Results and Impact
The impact of augmenting the training data with synthetic samples was transformative for forecasting performance. When an SVM classifier was trained on the balanced dataset and tested on unseen real data, the results showed:
- True Skill Statistic (TSS): Improved from a baseline of 0.04 to 0.81, a 20-fold increase.
- Heidke Skill Score (HSS): Improved from 0.08 to 0.45, a 5-fold increase.
These results demonstrate that the synthetic data was not just statistically similar but also carried the necessary predictive information to train a robust classifier.
Conclusion
This study proves that CGAN-based data augmentation is a viable and powerful method for addressing class imbalance in multivariate time series forecasting. By successfully mimicking the physics-based distributions of solar flares, the model provides a path forward for more accurate rare-event prediction in space weather and beyond. Future research will explore scaling this approach to the full 51-parameter SWAN-SF dataset and utilizing more advanced GAN variants like WGAN for improved stability.
Summary updated