Enhancing Multivariate Time Series-based Solar Flare Prediction with Multifaceted Preprocessing and Contrastive Learning

TL;DR

This research enhances solar flare prediction by utilizing advanced data preprocessing and classification methods on a multivariate time series-based dataset of photospheric magnetic field parameters and develops a novel classifier, termed ContReg, which employs dual learning methodologies, thereby further enhancing prediction performance.

Abstract

Accurate solar flare prediction is crucial due to the significant risks that intense solar flares pose to astronauts, space equipment, and satellite communication systems. Our research enhances solar flare prediction by utilizing advanced data preprocessing and classification methods on a multivariate time series-based dataset of photospheric magnetic field parameters. First, our study employs a novel preprocessing pipeline that includes missing value imputation, normalization, balanced sampling, near decision boundary sample removal, and feature selection to significantly boost prediction accuracy. Second, we integrate contrastive learning with a GRU regression model to develop a novel classifier, termed ContReg, which employs dual learning methodologies, thereby further enhancing prediction performance. To validate the effectiveness of our preprocessing pipeline, we compare and demonstrate the performance gain of each step, and to demonstrate the efficacy of the ContReg classifier, we compare its performance to that of sequence-based deep learning architectures, machine learning models, and findings from previous studies. Our results illustrate exceptional True Skill Statistic (TSS) scores, surpassing previous methods and highlighting the critical role of precise data preprocessing and classifier development in time series-based solar flare prediction.

Figures (8)

• Figure 1: The figure presents a stacked bar chart illustrating the distribution of different solar flare classes within each partition of the SWAN-SF dataset. This visualization is based on the current methodology of time series slicing used in SWAN-SF, which involves steps of 1 hour, an observation period of 12 hours, and a prediction span of 24 hours. Each slice of the MVTS is categorized according to the most intense flare reported within its prediction window.
• Figure 2: The figure demonstrates the different stages of our preprocessing pipeline for training and testing sets. Panel A showcases the training sets, while Panel B showcases the testing sets. No sampling methodologies are applied to the testing sets to avoid biased results.
• Figure 3: The figure illustrates the concept of Balanced Sampling, which includes both Balanced Over-Sampling and Balanced Random Under Sampling (RUS), along with the NDBSR strategy, a more sophisticated approach to addressing class imbalance. In the provided example, representing an approximation of the first partition from the SWAN-SF dataset, synthetic samples are generated for subclasses X and M in a controlled manner. The aim is to avoid generating an excessive number of synthetic samples, thereby preventing them from dominating the original samples. At the same time, a higher proportion of synthetic samples is generated for subclass X (500%) compared to subclass M (150%) to ensure a balanced representation between subclasses. Furthermore, in alignment with the NDBSR strategy, samples from subclasses B and C are completely removed from the minor-flaring class (comprising classes FQ, B, and C), while only a small portion of samples from class FQ is retained and utilized. This approach ensures that the distribution of the minor-flaring class (FQ) is aligned with that of the major-flaring samples (X and M classes), thereby promoting a balanced representation across the classes.
• Figure 4: The figure illustrates the architecture of ContReg, which employs three individual networks and utilizes a combined loss function to train the network and classify solar flare events. The three dots in the figure illustrate the concept of a fully connected layer. The technique combines contrastive learning with regression to create concise information that is fed into the final fully connected neural network, along with the actual input, to achieve higher classification performance.
• Figure 5: This figure showcases four distinct train-test sets employed in each classification experiment. This approach ensures a more comprehensive and accurate assessment of algorithms across all dataset partitions. In real-world time series forecasting, it is recommended that the training set precedes the test set chronologically, as the goal is always to predict the future. This approach leads to a more accurate and meaningful evaluation.