Solar flares are sudden, violent events that occur in the sun’s atmosphere and release large amounts of radiation. These phenomena can impact technological systems on Earth and in orbit, causing financial losses and endangering human life. Thus, forecasting these flares is crucial to mitigating their effects. Solar activity is monitored using specialized instruments that provide data for machine learning analysis. Researchers use various algorithms for this purpose. However, most works focus on the 24h forecasting window, while short-term approaches, like nowcasting (forecasting within 6h), are less explored. In this work, we developed solar flare forecasting models with a 6h window, using numerical data from the SHARP dataset. We implemented different machine learning algorithms, including Transformers, MLP, SVM, and LSTM, to compare their performance in a short-term scenario. For Transformers, we explored four approaches: (i) a conventional classification model; (ii) the FT-Transformer, a model optimized for tabular data; (iii) an adaptation for time series; and (iv) TabPFN, a pre-trained transformer model. TabPFN was pre-trained on over a million diverse datasets. Besides comparing Transformers and classical models, we examined whether using these four transformer approaches improves results within our 6h window using SHARP parameters. For evaluation, we did not select a single decision threshold; instead, we assessed how different thresholds (0.3, 0.5, and 0.7) affected outcomes. Lower thresholds favored TSS and TPR, while higher ones favored HSS and TNR. TabPFN achieved TSS = 0.79 and required much less training time than other deep learning-based models in this study.
Ferreira et al. (Sun,) studied this question.