CME Arrival Time Prediction Based on Coronagraph Observations and Machine-learning Techniques

The timely and precise prediction of the arrival time of coronal mass ejections (CMEs) is crucial in mitigating their potential adverse effects. In this study, we present a novel prediction method utilizing a deep-learning framework coupled with physical characteristics of CMEs and background solar wind. Time series images from synchronized solar white-light and EUV observations of 156 geoeffective CME events during 2000-2020 are collected for this study, according to the Richardson and Cane interplanetary CME directory and the SOHO/LASCO CME catalog of NASA/CDAW. The CME parameters are obtained from the CDAW website and the solar wind parameters are from OMNI2 website. The observational images are first fed into a convolutional neural network (CNN) to train a regression model as Model A. The results generated by the original CNN are then integrated with 11 selected physical parameters in additional neural network layers of Model B to improve the predictions. Under optimal configurations, Model A achieves a minimum mean absolute error (MAE) of 7.87 hr, whereas Model B yields a minimum MAE of 5.12 hr. During model training, we employed tenfold cross validation to reduce the occasionality of biased data. The average MAE of Model B on 10 folds is 33% lower than that of model A. The results demonstrate that combining the imaging observations with the physical properties of CMEs and background solar wind to train a machine-learning model can benefit the forecasting of CME arrival times.