Forecasting the transit times of earth-directed halo CMEs using artificial neural network: A case study application with GCS forward-modeling technique

Abstract

Mitigating the lethal threats caused by coronal mass ejections (CMEs) on human and space operations can be accomplished with a fast and accurate forecast of Earth-directed CME transit times. The current paper presents a robust Cascade Forward Neural Network (CFNN) framework to predict the transit times of Earth-directed halo CMEs using a total of 290 CME/Interplanetary coronal mass ejections (ICME) pairs of datasets for the past two and half decades (solar cycle; SC 23, 24, 25). It is the first time incorporating deprojected speeds into a machine learning framework to mitigate uncertainties due to projection effects during CME transit time prediction. The CFNN model forecasted the transit times of 87 Earth-directed CMEs and recorded a mean absolute error (MAE) of 7.3 h. In addition, 5 selected fast-moving (energetic) halo CME episodes during the deep phases (solar minima) of SC 25 were reconstructed using the Graduated Cylindrical Shell (GCS) forward-modeling technique. The events were predicted using the CFNN model (MAE = 4.5 h) in comparison with the Drag-Based model (DBM; MAE = 6.2 h) and Empirical Shock Arrival model (ESA; MAE = 13.5 h) to evaluate the robustness and the flexibility of the CFNN framework as well as the reliability of the CME 3D speed as a proxy for the space speed. The CFNN framework demonstrated satisfactory performance in contrast to previously used models by minimizing CME arrival time prediction errors due to projection effects. Hence, the study validated the efficiency of the GCS model for studying the 3D kinematics of CMEs and emphasized the essence of utilizing deprojected speeds in machine learning frameworks as better alternatives for fast, reliable, and accurate CME arrival (transit) time predictions.