Probabilistic Geomagnetic Storm Forecasting via Deep Learning

By causing time variation in Earth's external magnetic field, geomagnetic storms can induce damaging currents in ground-based conducting infrastructure, such as power and communication lines.  The physical link between solar activity and Earth's magnetosphere, while complicated, provides the basis for attempts to forecast geomagnetic storms. Fortunately, we have abundant observational data of both the solar disk and solar wind, which are ameable to the application of data-hungry neural networks to the forecasting problem. To date, almost all neural networks trained for geomagnetic storm forecasting have utilized solar wind observations from the Earth-Sun first Lagrangian point (L1) or closer and have generated deterministic output without uncertainty estimates. Furthermore, existing models generate forecasts for indices that are also sensitive to induced internal magnetic fields, complicating the forecasting problem with another layer of non-linearity. In this work, we present neural networks trained on observations from both the solar disk and the L1 point. Our architecture generates reliable probabilistic forecasts over Est, the external component of the disturbance storm time index, showing that neural networks can learn measures of confidence in their output.