Timing of the solar wind propagation delay between L1 and Earth based on machine learning

Abstract

Erroneous GNSS positioning, failures in spacecraft operations and power outages due to geomagnetically induced currents are severe threats originating from space weather. Knowing the potential impacts on modern society in advance is key for many end-user applications. This covers not only the timing of severe geomagnetic storms but also predictions of substorm onsets at polar latitudes. In this study, we aim at contributing to the timing problem of space weather impacts and propose a new method to predict the solar wind propagation delay between Lagrangian point L1 and the Earth based on machine learning, specifically decision tree models. The propagation delay is measured from the identification of interplanetary discontinuities detected by the advanced composition explorer (ACE) and their subsequent sudden commencements in the magnetosphere recorded by ground-based magnetometers. A database of the propagation delay has been constructed on this principle including 380 interplanetary shocks with data ranging from 1998 to 2018. The feature set of the machine learning approach consists of six features, namely the three components of each the solar wind speed and position of ACE around L1. The performance assessment of the machine learning model is examined based on of 10-fold cross-validation. The machine learning results are compared to physics-based models, i.e., the flat propagation delay and the more sophisticated method based on the normal vector of solar wind discontinuities (vector delay). After hyperparameter optimization, the trained gradient boosting (GB) model is the best machine learning model among the tested ones. The GB model achieves an RMSE of 4.5 min concerning the measured solar wind propagation delay and also outperforms the physical flat and vector delay models by 50% and 15% respectively. To increase the confidence in the predictions of the trained GB model, we perform a performance validation, provide drop-column feature importance and analyze the feature impact on the model output with Shapley values. The major advantage of the machine learning approach is its simplicity when it comes to its application. After training, values for the solar wind speed and spacecraft position from only one datapoint have to be fed into the algorithm for a good prediction.