In this dissertation, we explore advanced deep learning methodologies for predicting rare events, with a specific focus on applications in space weather. We address key challenges in solar flare forecasting, including data scarcity, class imbalance, ordinality, and the need for explainability to advance the development of robust and reliable predictive models.

We make four key contributions. First, we introduce deep learning-based binary classification models for predicting solar flares using images of the solar full-disk and active region (AR) patches, demonstrating improved performance. Second, we propose a novel method for incorporating the inherent ordering of solar flare intensity classes into the loss function of a binary classifier, enabling the model to more effectively capture the ordinal nature of the prediction task. Third, we advance the application of explainable AI by developing an attention-based model and leveraging attention maps for qualitative analysis of the model’s spatial attention patterns. Additionally, we generate post hoc explanations for multiple models and systematically evaluate their quality using both human-centered and automated frameworks—thereby contributing to the methodological foundation for evaluating and inter- preting the model’s decision-making rationale. Fourth, we introduce techniques for analyzing explanation consistency, investigate multi-granular explanations as proxies for localized predictions, and conduct quantitative analysis of explanations through physically meaningful features, revealing systematic patterns that provide insight into model predictions.

Overall, we show how deep learning models, combined with rigorous explanation analysis, can improve the predictive performance and reliability of solar flare forecasting. Beyond this specific application, the methodologies introduced here offer a practical framework for rare event prediction in other domains with similar challenges.