Combining Solar Imagery and In Situ Data for Prediction of Geomagnetic Storms

Abstract In this work, we propose the use of the Most Extreme Space Weather Events (MESWE) dataset, which combines 12 solar corona and solar wind parameters to predict the disturbance storm time (Dst) index 4 or 8 hr ahead by deep learning models. The dataset focuses on periods of extreme solar activity in the past 30 yr that triggered great and severe geomagnetic storms, as measured by geomagnetic indices. Using the MESWE dataset and leveraging the gated recurrent units type of recurrent neural network, we obtain a significantly reduced persistence effect compared to models trained with conventional methodologies in the Dst index forecasting community. The performance of the models is compared quantitatively by using the dynamic time warping method. In addition, we introduce a methodology to reduce the risk of the split of the validation test set that can lead to misleading performance assessments. The methodology of combining imagery and in situ data and the results presented in this work are directly aligned with the objectives of the European Space Agency Vigil mission.