Forecasting High‐Speed Solar Wind Streams From Solar Images

Abstract

Abstract The solar wind, a stream of charged particles originating from the Sun and transcending interplanetary space, poses risks to technology and astronauts. In this work, we develop a prediction model to forecast the solar wind speed (SWS) at the Earth. We focus on high‐speed streams (HSSs) and their solar source regions, coronal holes. As input, we use the coronal hole area, extracted from solar extreme ultraviolet (EUV) images and mapped on a fixed grid, as well as the SWS 27 days before. We use a polynomial regression model and a distribution transformation to predict the SWS with a lead time of 4 days. Our forecast achieves a root mean square error (RMSE) of 68.1 km/s for the SWS prediction and an RMSE of 76.8 km/s for the HSS peak velocity prediction for 2010 to 2019. We also demonstrate the applicability of our model to the current solar cycle 25 in an operational setting, resulting in an RMSE of 80.3 km/s and an HSS peak velocity RMSE of 92.2 km/s. The study shows that a small number of physical features explains most of the solar wind variation, and that focusing on these features with simple machine learning algorithms even outperforms current approaches based on deep neural networks and MHD simulations. In addition, we explain why the typically used loss function, the mean squared error, systematically underestimates the HSS peak velocities, aggravates operational space weather forecasts, and how a distribution transformation can resolve this issue.

Keywords

• solar wind speed
• machine learning
• coronal holes
• solar images
• high‐speed streams
• prediction