Deep Learning Kp Computational Prediction Combined Multimodal Data of Solar Wind and EUV Images

This study presents a deep learning framework for multi-scale prediction of the geomagnetic Kp index by integrating multi-modal solar data. Unlike models that rely solely on upstream solar wind measurements from the Advanced Composition Explorer (ACE), which offer a short-term forecast horizon of 0.5~2 h, our approach synergistically incorporates Extreme Ultraviolet (EUV) images of coronal holes from the Solar Dynamics Observatory (SDO). These images provide precursor signals of solar activity, enabling medium-term forecasts. This study constructed a multi-year dataset (2011－2019) with carefully designed features to ensure physical meaning and sample balance. And several models were developed and evaluated: a direct solar wind model, an indirect multi-scale model using EUV images, and a model incorporating coronal hole location information. While all models showed similar and superior performance in 3-hour-ahead predictions, the model incorporating coronal hole location information demonstrated a significant advantage at extended forecast horizons, maintaining a Correlation Coefficient (CC) of approximately 0.23 even at a 72 h lead time. This work underscores the critical value of fusing solar source images with in-situ measurements to achieve reliable, long-lead space weather forecasts.