Forecasting the geomagnetic activity with deep learning techniques ; Apprentissage profond pour la prévision de l'activité géomagnétique

The study of Sun–Earth interactions, in particular through the solar wind–magnetosphere coupling, is at the core of space weather related issues. In this work we focus on the forecasting of geomagnetic indices within a few days, which can be used to drive the Earth's radiation belts models. During the last decade, many studies have shown that artificial neural networks can predict these indices in a very efficient way, based on measurements of the solar wind near the Earth.In our work, we first propose a new model for the prediction of the geomagnetic index Dst, consisting of a neural network with recurrent layers. This new model produces better probabilistic forecasts than the current state of the art for prediction horizons shorter than 6 hours. In order to make our model more operationally useful, we adapt it for the prediction of the new geomagnetic index Ca, designed to better account for the geoeffectiveness of geomagnetic events from the perspective of the electron radiation belts. By conducting a comprehensive evaluation of our model, we show that it loses its usefulness in an operational context for prediction horizons longer than a few hours.Based on this observation, and facing the limits shown by the current physical models of solar wind propagation, we study the use of solar imaging to directly forecast the geomagnetic index Kp from 2 to 7 days ahead. To do so, we build SERENADE, the first geomagnetic index forecasting model driven only by images of the Sun. This model is a neural network with a complex architecture combining layers of different nature. We show that our model performs at least as well as some simple empirical models (and yet currently the most efficient) for forecasting the daily maximum of Kp. We show that our model, although still immature for an operational context, is able to account for the geoeffectiveness of some solar events directly from solar imagery alone. By identifying the limitations of our model and their causes, our results open the way to a data-driven modeling of ...