Speaker
Description
Coronal Mass Ejections (CMEs) are major drivers of Space Weather (SWx) effects on Earth, and predicting their arrival is a major aspect of SWx forecast. Several CME propagation models have been developed for this purpose, but the overall arrival time error still exceeds 12 hours. In this study, we aim to improve these predictions by employing machine learning (ML) techniques that utilize the differences between observed and simulated solar wind (SW) profiles ahead of CMEs. Our analysis includes 160 CME events from March 2012 to March 2023. We found that the mean average error (MAE) for the WSA-ENLIL+CONE (WEC) modeling of these events is ~ 12.54 hours. We use OMNI data to get the near-Earth observed SW profile. The simulated background SW is borrowed from publicly available simulation results obtained by the ENLIL model. These datasets include key SW parameters such as speed, density, temperature, magnetic field strength, and total pressure. The mean difference between these observed and simulated parameters, from the time of insertion to the time of arrival, is calculated for each CME. We employ three ML models—1) k-nearest neighbors (KNN), 2) support vector machine (SVM), and 3) linear regression (LR) in our analysis. These models use errors in the upstream SW parameters as features. Our ML models are set up to estimate the error in the WEC prediction based on the errors in the upstream SW parameters. Our results identified the mean SW pressure difference as the most effective parameter for estimating WEC error. We were able to reduce the MAE given by the WEC model by 1.15 hours using the SVM model. This shows that the incorrect SW background created using WSA-ENLIL contributes to MAE, and our proposed ML-based methods can reduce the MAE by quantifying this contribution.
