Uncertainty Bounds for Anomalous Geomagnetic Storm Forecasting - A Deep Learning Approach

We present a novel approach to forecast geomagnetic storms by employing a multitask multivariate transformer-based methodology. Our model demonstrates the ability to predict the Sym-H, Ap, and Kp indices simultaneously, which serve as widely recognized proxies for assessing geomagnetic activities. Diverging from prior models that heavily rely solely on solar-wind indices, our novel methodology aims to overcome their limitations and yield improved forecasting capabilities using a multitask prediction approach. Through rigorous training on a meticulously curated high-resolution dataset spanning from the 1990s, our model attains remarkable accuracy and reliability in predicting severe geomagnetic storms. Considering the profound implications of these storms on electric grids and communication systems, our overarching objective is to provide comprehensive forecasts with well-defined uncertainty bounds.