A Deep Learning Approach to forecasting solar wind properties

Abstract:

Deep Learning has proven extremely successful both in supervised learning problems, such as image and speech recognition, and in unsupervised tasks, such as games and videogames playing. Whether this technology is applicable to the space weather problem is still an open question. However, given the high-risk/high-gain nature of this approach, we believe that it is something worth investigating.

Hence, we have systematically explored a Deep Learning approach to forecast several solar wind properties at L1, such as wind speed, magnetic field, density, and temperature, mostly based on solar images.

The task falls in the semi-supervised category, because even though in-situ solar wind measurements are available, a proper ‘ground truth’ does not exist due to uncertainties in the propagation time.

We present a number of approaches that include both forecasts with a fixed lead time and the prediction of the propagation time itself. Finally, pure black-box technologies are compared to gray-box approaches where physical models are included in the Deep Learning pipeline.