Forecasting global auroral particle precipitation and boundaries with novel multi-task deep learning techniques

Abstract:

We have developed multiple Deep Learning models to predict the global auroral particle precipitation, including a novel multi-task model that uses solar wind, geomagnetic indices, and their time histories as input to predict the total electron energy flux, total number flux, and auroral boundaries globally. We show results for nowcasts (prediction at the time of observation) and forecasts from one up to several hours. More than 40 satellite years of Defense Meteorological Satellite Program (DMSP) precipitation data provide the observations, which are split into training and validation sets.. DMSP data provide local information along the orbital paths of the spacecraft, and require intensive data treatment. We detail the data preparation process as well as the model development that will be illustrative for many similar time series global regression problems in space weather. We highlight a central result of importance is our design of a new multi-task deep learning network that simultaneously predicts the auroral region and the precipitation. Details of the influence of model structure are also highlighted, convolutional vs dense layers for example, as well dedicated designs for the multi-modal nature of the rare and extreme space storms.