APPLICATION OF MACHINE LEARNING MODELS IN PREDICTING THE SOLAR WIND PROPAGATION FROM L1 MONITORS TO THE EARTH’S BOW SHOCK

Abstract

Understanding the near-Earth solar wind (SW) and interplanetary magnetic field (IMF) is crucial for space weather operations and modeling the magnetosphere and ionosphere. Studies on SW and its impact on Earth primarily rely on spacecraft data from the Lagrangian point L1. Accurate forecasting of ionospheric and magnetospheric conditions depends on precisely predicting the arrival time of SW disturbances from L1 to Earth’s upstream region and magnetosphere. A widely used data source, OMNIWeb, estimates SW propagation time from L1 to Earth’s bow shock. However, differences between the timeshifted IMF provided by OMNIWeb and the best-matched IMF often exceed OMNIWeb’s uncertainties. This study addresses SW propagation delay issues by applying artificial intelligence, e.g., machine learning (ML) and deep learning models (DL), to enhance SW delay predictions. ML models [e.g., gradient boosting (GB) and random forest (RF)] and DL models [e.g., Long Short-Term Memory (LSTM) Network and Multilayer perceptron (MLP)] are trained, tested, and validated using solar wind features (e.g., solar wind speed, temperature, magnetic field, positions of L1, and near-Earth monitors) and target delays to predict the propagation time from L1 monitors to a given location upstream or at the bow shock. The target delays are estimated solar wind propagation delays using a statistical approach by comparing SW features at L1 and upstream of the bow shock. The performances of these ML and DL models are evaluated on a test dataset. The overall performance of the ensemble-based ML models is better than that of the neural network model in solving the SW propagation delay issues. The root mean square error (RMSE) values for ML models (RMSE of RF is 1.5% and of GB is 2.5%) are lower than those of DL models (RMSE of LSTM is 3% and MLP is 4%). Additionally, the machine learning models’ predicted delays are compared with the predictions using physics-based models, i.e., flat delays and OMNIWeb-provided delays. In the selected cases, the delay predicted by the machine learning model results in a better match between the IMF features at the L1 point and those near Earth, compared to the delay provided by OMNIWeb.