Deep Learning‐Based Prediction of High‐Speed Solar Wind Streams: Spatio‐Temporal Dependencies in Coronal Hole Dynamics

Abstract High‐speed solar wind streams (HSS), originating from coronal holes (CH), are key drivers of space weather disturbances and heliospheric dynamics. However, forecasting HSS remains challenging due to the evolving morphology of CH. In this study, we present a deep learning‐based framework that models the spatiotemporal relationship between CH and HSS.We applied preprocessing techniques that included the Stonyhurst projection, removal of off‐limb structures, transient events, and background noise, thus isolating persistent CH features. We developed two convolutional neural network|convolutional neural networks (CNN) models: one using full‐disk extreme ultraviolet images of the sun at 193 Å, 171 Å, and 304 Å wavelengths; the other using binary CH maps derived from 193 Å wavelength. Both models are trained and evaluated across different solar cycle phases using a meta‐learning strategy to retain optimal checkpoints based on validation loss. We find that, over the entire solar cycle (SC) period, our model outperforms the benchmark models, achieving a best correlation of , a root mean square error of Km/s, and a threat score of 0.71 with the observed solar wind (SW) at a 3‐day lead time. The explainability attribution method utilized confirms the model's ability to focus on CH regions and track their evolution over time. Both models learn physically consistent features, highlighting CH structures near the central meridian as key HSS sources. SpeedNet‐BM further demonstrates stable and coherent activation responses, reinforcing its predictive capability and the importance of domain‐specific preprocessing, phase‐aware training, and interpretable deep learning to improve HSS forecasting.