Intelligent prediction model for geoeffectiveness of coronal mass ejections

Predicting the geoeffectiveness of coronal mass ejections (CMEs) is a core issue in operational space weather forecasting. In recent years, deep learning methods have been increasingly applied to CME geoeffectiveness prediction. However, existing studies mostly rely on single data sources, with image-based and parameter-based approaches each having their own limitations and failing to achieve effective complementarity, while also facing challenges of high false alarm rates and extreme class imbalance. In this study, we compiled a dataset of 15,983 CME events observed by the SOHO/LASCO coronagraph from 2001 to 2023, comprising 116 geoeffective events and 15,867 non-geoeffective events, with a positive-to-negative sample ratio of approximately 1:136. Based on this dataset, we adopted a data-driven approach to first extract the morphological and spatiotemporal evolution features of CMEs from coronagraph running-difference image sequences, establishing a TimeSformer-based image prediction model. We then incorporated CME characteristic parameters including speed, mass, and angular width through a cross-modal attention mechanism to achieve complementary integration of image morphology and CME characteristic parameters, and designed a dedicated training strategy to address the extreme class imbalance problem, ultimately constructing a multimodal CME geoeffectiveness prediction model. The dataset is partitioned chronologically at the event level into a training set (2002–2014, 10,663 events with 74 geoeffective), a validation set (2015–2018, 2,810 events with 19 geoeffective), and a held-out test set (2019–2023, 2,510 events with 23 geoeffective), so that no event from a later time window can leak information into an earlier-trained model. Experimental results on the held-out test set demonstrate that after incorporating CME characteristic parameters the model’s receiver operating characteristic area under the curve (ROC-AUC) improves from 0.782 (95% bootstrap CI [0.653,0.898]) to 0.852 ([0.754,0.939]), and false alarms decrease from 317 to 222 (a reduction of 30.0 % ) at the validation-selected operating threshold, with the CME characteristic parameters effectively helping to exclude false positive events that are morphologically similar but lack sufficient speed or deviate in propagation direction. Experiments demonstrate that multimodal fusion effectively combines image morphological information with parametric physical constraints, providing a feasible technical solution for operational space weather forecasting.