Forecasting and Modelling Space Weather with Deep Learning Methods

Abstract

With the ever increasing number of spacecraft providing essential services for society, the accurate modeling and forecasting of conditions for these spacecraft becomes increasingly important. The conditions that these spacecraft operate in is often referred to as space weather. The advent of deep learning has unlocked the ability to use large datasets to model and forecast these conditions. This thesis principally describes a set of methodological improvements, considerations and proof of concept systems that use extreme ultra-violet (EUV) solar images and deep learning techniques to forecast and model space weather conditions. Firstly, vision transformers are used to forecast solar wind speed from solar EUV images, with improvements over previous work. Secondly, solar irradiance is forecast using pre-trained vision transformers that consume nine solar EUV/UV image channels, with its performance explored. Thirdly, autoencoders are trained to create new solar indices that can be used to forecast various space weather phenomena to significant effect, motivating the use of such indices in production systems. Lastly, thermospheric density models are trained that can significantly outperform existing physics-based models.