Simultaneously forecasting global geomagnetic activity using Recurrent Networks

Many systems used by society are extremely vulnerable to space weather events such as solar flares and geomagnetic storms which could potentially cause catastrophic damage. In recent years, many works have emerged to provide early warning to such systems by forecasting these events through some proxy, but these approaches have largely focused on a specific phenomenon. We present a sequence-to-sequence learning approach to the problem of forecasting global space weather conditions at an hourly resolution. This approach improves upon other work in this field by simultaneously forecasting several key proxies for geomagnetic activity up to 6 hours in advance. We demonstrate an improvement over the best currently known predictor of geomagnetic storms, and an improvement over a persistence baseline several hours in advance.