Prediction of the solar wind propagation delay for L1 to Earth using machine learning

Abstract

Precise information on the impact of space weather phenomena are crucial for GNSS positioning in aviation, spacecraft operations in general, and many other fields of modern industries. In particular, it is of great importance for end-users that require urgent knowledge of the expected timing of these impacts. This work contributes to mitigating the timing problem by predicting the solar wind delay between Lagrangian point L1 and Earth using machine learning regression based on decision tree models.

The propagation delay is measured by identifying the instance when an interplanetary shocks passing the Advanced Composition explorer (ACE) satellite and the shock's subsequent arrival at Earth as measured by magnetometers. The database of this study covers the years from 1998 until 2018 and includes 380 interplanetary shocks with solar wind speeds from 300 to 1000 km/s. As input features we have chosen the solar wind speed in all three directions and the position of ACE around L1. In addition, we use the DST index for a simple representation of the state of Earth's magnetosphere. To assess the performance of the machine learning approach we apply a 10-fold cross validation and derive the RMSE of the prediction.

The results are compared to the flat propagation delay (in the x-direction) and the propagation delay based on the normal vector of the shock front (vector delay). The random forest algorithm can reduce the RMSE of the predicted delay to 4.8 minutes , achieving a 15% increase in performance compared to the vector delay method and an even greater improvement of 50% when compared to the flat delay method. To explore the near real-time capabilities of this method, we divide our data set into a training set ranging from 1998 to 2014 and test the algorithm on the remaining 2015-2018 cases. In this real time scenario, again the trained machine learning model output shows the smallest RMSE among the tested methods. Additionally, this study aims to provide greater interpretability of the machine learning approach to predicting the propagation delay. For that purpose also drop-column feature importance and Shapley values of the trained algorithm are presented.