High‐Precision Prediction of Auroral Substorm Expansion Phase Evolution Using a Spatio‐Temporal Attention Deep Learning Model

Abstract Auroral substorms are a key manifestation of solar‐terrestrial interactions and the release of energy within the Earth's magnetosphere. The expansion phase, which is characterized by rapid changes in auroral morphology and intensity, poses significant space weather risks to satellite operations and communication systems. However, accurately forecasting the complex spatio‐temporal evolution of auroras during this phase remains challenging with existing models. In this study, we introduce STA‐Aurora, a novel deep learning model utilizing a spatio‐temporal attention mechanism specifically designed to predict auroral evolution during the substorm expansion phase. Utilizing the initial four auroral images at the onset of expansion as input, STA‐Aurora employs a multi‐scale spatio‐temporal attention module to focus adaptively on regions with critical intensity enhancement and motion. To mitigate motion ambiguities and reduce prediction blurring, we incorporate a context gating unit. Furthermore, we utilize a combined loss function that integrates mean squared error, style loss, and micro‐dispersion regularization, ensuring that predictions maintain both morphological fidelity and motion coherence. Evaluated on Polar/UVI observations, STA‐Aurora demonstrates substantial improvements over the convolutional long short‐term memory baseline, including a 26.4% increase in structural similarity (SSIM), a 40.71% enhancement in auroral bright spot intensity prediction, a 5.58% improvement in bright spot coverage prediction, and an 18.5% improvement in auroral oval boundary location prediction. The average location errors were 2.31° MLAT for poleward boundaries and 0.68° MLAT for equatorward boundaries. These results highlight STA‐Aurora as a significant advancement in modeling auroral dynamics, offering a promising new approach to high‐resolution space weather forecasting.