Global Cross-Time Attention Fusion for Enhanced Solar Flare Prediction from Multivariate Time Series

Anonymous, Onur Vural, Shah Muhammad Hamdi, Soukaina Filali Boubrahimi, Onur Vural, Shah Muhammad Hamdi, Soukaina Filali Boubrahimi · 2025

Read Global Cross-Time Attention Fusion for Enhanced Solar Flare Prediction from Multivariate Time Series by Anonymous, Onur Vural, Shah Muhammad Hamdi, Soukaina Filali Boubrahimi, Onur Vural, Shah Muhammad Hamdi, Soukaina Filali Boubrahimi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-12955.txt.

Introduction

Onur Vural, Shah Muhammad Hamdi, and Soukaina Filali Boubrahimi predict strong solar flares. Intense flares can disrupt power grids and satellites, so warnings matter. Magnetic field measurements of solar active regions are converted into multivariate time series. The learning problem is lopsided: major flares are rare next to minor ones, and models can look accurate while missing every event that counts. Their model adds learnable global tokens that gather patterns across the whole sequence.

Through cross-attention, those tokens act as a summary that is fused back into the time series representation, letting the model use distant, non-adjacent moments that signal a coming flare. On the standard solar flare benchmark, the model detected intense flares effectively and improved predictive performance. The paper was accepted at an IEEE big data conference. It suggests flare forecasting benefits from attention that is explicitly global, not only local in time.

Catalog record

CREATORAnonymous, Onur Vural, Shah Muhammad Hamdi, Soukaina Filali Boubrahimi, Onur Vural, Shah Muhammad Hamdi, Soukaina Filali Boubrahimi
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-12955

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Global Cross-Time Attention Fusion for Enhanced Solar Flare Prediction from Multivariate Time Series Anonymous, Onur Vural, Shah Muhammad Hamdi, Soukaina Filali Boubrahimi, Onur Vural, Shah Muhammad Hamdi, Soukaina Filali Boubrahimi Abstract Multivariate time series classification is increasingly investigated in space weather research as a means to predict intense solar flare events, which can cause widespread disruptions across modern technological systems. Magnetic field measurements of solar active regions are converted into structured multivariate time series, enabling predictive modeling across segmented observation windows. However, the inherently imbalanced nature of solar flare occurrences, where intense flares are rare compared to minor flare events, presents a significant barrier to effective learning. To address this challenge, we propose a novel Global Cross-Time Attention Fusion (GCTAF) architecture, a transformer-based model to enhance long-range temporal modeling. Unlike traditional self-attention mechanisms that rely solely on local interactions within time series, GCTAF injects a set of learnable cross-attentive global tokens that summarize salient temporal patterns across the entire sequence. These tokens are refined through cross-attention with the input sequence and fused back into the temporal representation, enabling the model to identify globally significant, non-contiguous time points that are critical for flare prediction. This mechanism functions as a dynamic, attention-driven temporal summarizer that augments the model’s capacity to capture discriminative flare-related dynamics. We evaluate our approach on the benchmark solar flare dataset and show that GCTAF effectively detects intense flares and improves predictive performance, demonstrating that refining transformer-based architectures presents a high-potential alternative for performing solar flare prediction tasks. I Introduction An intense solar flare erupted from the Sun’s surface in 1859, unleashing a powerful geomagnetic storm that disrupted telegraph systems across Europe and North America, sparked widespread auroral displays visible near the equator, and induced electrical currents strong enough to ignite fires in communication lines [ 8 ] . This occurrence, known as the Carrington Event, is a striking and extreme historical example of the effects of an intense solar flare [ 3 ] . Solar flares are explosive manifestations of solar magnetic energy, erupting from active regions on the Sun’s surface and extending into the corona and interplanetary space [ 13 ] . These bursts are classified on a logarithmic scale as A, B, C, M, and X, based on peak soft X-ray intensity in the 1 to 8 Ångström band, with M and X flares indicating intense flare events [ 1 ] . Beyond their astrophysical intrigue, such eruptions bear serious consequences: they unleash radiation across the electromagnetic spectrum, including gamma rays and extreme ultraviolet, that can threaten astronaut safety, compromise satellites and power grids, and disrupt navigation and communications infrastructures. Furthermore, an occurrence of similar scale to the Carrington Event could lead to devastating societal and economic consequences in our technology-driven world [ 6 ] . In light of these risks, the scientific community continues to advance solar flare prediction by analyzing solar magnetic field data and developing predictive models. Magnetic field measurements from solar active region vector magnetogram data have been increasingly organized into structured multivariate time series (MVTS) over segmented time windows, enabling MVTS-driven predictive modeling that outperforms single-timestamp approaches [ 3 ] . However, severe class imbalance due to rarity of intense flares is a central challenge, biasing models to favor the majority class [ 9 ] . Traditional classifiers and sequence models typically exploit only temporal dependencies in MVTS data, limiting capture of complex inter-variable patterns and resulting in suboptimal performance in identifying intense flaring events. To address these limitations, transformer-based architectures emerge as promising candidates, utilizing self-attention mechanisms to dynamically weigh the relevance of each time step and feature in a sequence, capturing both local and global dependencies without the sequential limitations of traditional models. In solar flare prediction, where critical magnetic patterns may be subtle and dispersed across time, such expressiveness is particularly advantageous [ 2 ] . Additionally, cross-attention mechanisms, which relate different modalities or segments of data, can further help to identify rare but significant flares. In this work, we introduce a transformer-based classification framework, Global Cross-Time Attention Fusion (GCTAF), tailored to improve the prediction of solar flare events from MVTS data. Unlike generic transformer models that rely solely on self-attention over local temporal interactions, GCTAF introduces learnable global tokens that attend to the entire input sequence via a cross-attention mechanism. This enables the model to capture globally salient temporal dependencies that conventional self-attention may overlook. These global representations are fused with local temporal features extracted through stacked transformer encoder modules, allowing the model to integrate both fine-grained and holistic information. GCTAF is assessed using the MVTS solar flare benchmark dataset, where enhanced prediction performance indicates higher effectiveness in identifying intense flares. These findings underscore the efficacy of refined transformer models in addressing the inherent challenges of solar flare prediction. II Related Work II-A A Brief History of Solar Flare Prediction Early solar flare prediction relied on expert systems grounded in sunspot classification [ 13 ] . With advances in observational technology, magnetic field data became increasingly accessible, shifting prediction toward data-driven approaches [ 17 ] . NASA’s Solar Dynamics Observatory (SDO) mission, particularly its Helioseismic and Magnetic Imager (HMI) instrument, enabled large-scale, high-cadence magnetogram data analysis [ 1 ] . Consequently, flare prediction evolved into a machine learning task employing traditional classifiers such as SVMs [ 4 ] , and logistic regression [ 15 ] . Building on these efforts, the Space Weather Analytics for Solar Flares (SWAN-SF) dataset was introduced [ 3 ] , capturing magnetic field evolution as MVTS data. Since then, MVTS-based approaches have become central to advancing solar flare prediction. These include preprocessing-enhanced decision trees [ 12 ] , contrastive representation learning to emphasize class distinctions [ 18 , 17 , 20 ] , deep LSTM architectures for temporal sequence modeling [ 14 ] , and graph-based methods [ 19 ] . II-B Transformer-Based Time Series Classification The emergence of transformers represents a breakthrough in deep learning approaches by employing self-attention mechanisms to capture long-range dependencies in sequential data without relying on recurrent or convolutional operations [ 16 ] . This innovation enabled parallel processing and greater flexibility in modeling sequence data, which has since propelled transformers beyond natural language processing into diverse domains, including time series classification. In time series tasks, transformers have shown strong potential by learning contextual temporal features across multiple channels while accommodating variable-length sequences [ 21 ] . Selected examples of recent advancements include transformer-based unsupervised MVTS representation learning for improved downstream classification [ 23 ] , gated transformer architectures for enhanced feature extraction [ 11 ] , shapelet-based modules to capture both class-specific and general patterns [ 10 ] , and the use of hierarchical pooling, adaptive frequency learning strategies, and feature-map-wise attention mechanisms to model multi-scale temporal dependencies [ 22 ] . Collectively, these developments emphasize the transformer’s versatility and growing impact in time series classification research. III Dataset Although machine learning has been applied to solar flare prediction, earlier efforts primarily relied on static, point-in-time observations [ 4 ] . To address this limitation, the SWAN-SF dataset was introduced [ 3 ] , providing a temporally resolved framework that captures the dynamic evolution of solar magnetic fields through MVTS data. Each SWAN-SF instance is an

Read the complete text