Investigating whether machine learning alone can predict solar wind parameters at L1 from solar images
Abstract
The solar wind is a stream of energized charged particles, primarily electrons and protons, flowing outward from the Sun. Predicting solar wind properties such as speed, density and magnetic field near Earth is important for space weather applications. However, making accurate predictions has proven to be challenging, and many forecasting tools do not significantly outperform a 27-day persistence model. Most forecasts rely on physical approximations, for example to extrapolate magnetic fields or to compute so-called flux-tube expansion factors. Here we investigate whether such physical approximations can be replaced by a machine learning model, by studying to what extent such a 'naive' model can predict solar wind properties at L1 directly from solar images. For this purpose, we use a data set containing over 15000 solar images at different wavelengths, and corresponding solar wind data from the OMNI database. The two main challenges are discussed. First, the solar image data is high-dimensional compared to the observed solar wind parameters. We explore the performance of several dimensionality reduction strategies for the solar image data. Second, there is a variable time lag between the emission of the solar wind and its arrival at L1. This time lag depends on the location of emission and on the solar wind speed. We investigate how to best account for this time lag, and whether it is possible to do so dynamically.
This project has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No 776262 (AIDA, www.aida-space.eu).- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMNG0040027T
- Keywords:
-
- 1914 Data mining;
- INFORMATICS;
- 7833 Mathematical and numerical techniques;
- SPACE PLASMA PHYSICS;
- 7924 Forecasting;
- SPACE WEATHER;
- 7959 Models;
- SPACE WEATHER