Short-term solar flare forecasting with a hybrid ConvNeXtV2 and transformer–temporal convolutional network
- 1omputer Key Laboratory, Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming, China
Abstract
Solar flares are violent energy-release events in the solar atmosphere, accompanied by intense electromagnetic radiation and energetic particles that threaten satellites, radio communications, and power infrastructure, making accurate flare prediction essential for space weather early warning. This paper proposes a deep learning model integrating ConvNeXtV2 with a Transformer–Temporal Convolutional Network (TCN) to forecast whether a solar active region will produce a ≥C-class or ≥M-class flare within the next 24 h. Using SHARP magnetograms and magnetic parameters from SDO/HMI (2010–2019), ConvNeXtV2 first extracts spatial structural and deep semantic features from the magnetograms; the resulting feature sequences are then fed into the Transformer–TCN module, where the Transformer captures global temporal dependencies and the TCN models local dynamic features of active-region evolution. The two components jointly predict flare occurrence probability, with the True Skill Statistic (TSS) as the primary metric. Under active-region-isolated five-fold cross-validation, the model achieves TSS values of 0.861±0.021 and 0.844±0.041 for the ≥C-class and ≥M-class tasks, respectively, and performs at least as well as the strongest baselines retrained on the same data and the same folds. These results indicate that combining ConvNeXtV2 with the Transformer–TCN characterizes both the magnetic field structure of active regions and its dynamic evolution, offering a further option for short-term solar flare forecasting.
Related
- Solar Flare Prediction Using a Hybrid Convolutional Neural Network and Transformer Model
- A Swin Transformer for Multiparameter Solar Wind Forecasting from MHD Simulations
- Identification of Magnetic Reconnection in Plasma Turbulence Using a Dual-branch Lightweight Hybrid Convolutional Neural Network
- Fast Radio Burst Classification on Dispersed Dynamic Spectra Using a Wavelet-based Convolutional Network
- Stacked Hybrid Recurrent Convolutional Neural Network Reconstruction of X-Ray Influence on 21 cm Brightness Temperature
BibTeX
@article{zheng2026shortterm,
title = {{Short-term solar flare forecasting with a hybrid ConvNeXtV2 and transformer–temporal convolutional network}},
author = {Mingjian Zheng and Xu Xiong and Hui Li},
journal = {Frontiers in Astronomy and Space Sciences},
year = {2026},
month = oct,
doi = {10.3389/fspas.2026.1999089},
url = {https://doi.org/10.3389/fspas.2026.1999089},
note = {CC BY},
}
RIS
TY - JOUR
TI - Short-term solar flare forecasting with a hybrid ConvNeXtV2 and transformer–temporal convolutional network
AU - Mingjian Zheng
AU - Xu Xiong
AU - Hui Li
PY - 2026
DA - 2026/10/26
JO - Frontiers in Astronomy and Space Sciences
DO - 10.3389/fspas.2026.1999089
UR - https://doi.org/10.3389/fspas.2026.1999089
AB - Solar flares are violent energy-release events in the solar atmosphere, accompanied by intense electromagnetic radiation and energetic particles that threaten satellites, radio communications, and power infrastructure, making accurate flare prediction essential for space weather early warning. This paper proposes a deep learning model integrating ConvNeXtV2 with a Transformer–Temporal Convolutional Network (TCN) to forecast whether a solar active region will produce a ≥C-class or ≥M-class flare within the next 24 h. Using SHARP magnetograms and magnetic parameters from SDO/HMI (2010–2019), ConvNeXtV2 first extracts spatial structural and deep semantic features from the magnetograms; the resulting feature sequences are then fed into the Transformer–TCN module, where the Transformer captures global temporal dependencies and the TCN models local dynamic features of active-region evolution. The two components jointly predict flare occurrence probability, with the True Skill Statistic (TSS) as the primary metric. Under active-region-isolated five-fold cross-validation, the model achieves TSS values of 0.861±0.021 and 0.844±0.041 for the ≥C-class and ≥M-class tasks, respectively, and performs at least as well as the strongest baselines retrained on the same data and the same folds. These results indicate that combining ConvNeXtV2 with the Transformer–TCN characterizes both the magnetic field structure of active regions and its dynamic evolution, offering a further option for short-term solar flare forecasting.
SN - 2296-987X
ER -