基于数据融合的日冕物质抛射传播时间预报

Forecast of Coronal Mass Ejection Propagation Time Based on Data Fusion

摘要: 日冕物质抛射(CMEs)作为极具破坏性的空间天气现象，其传播时间精准预报对保障航天器安全、卫星导航及电力网络稳定至关重要。针对现有研究未充分融合LASCO观测图像与文本物理参数的空白，本文提出一套完整的数据融合方案与预报模型。首先通过时间戳匹配，实现255个日冕物质抛射事件的时序图像与19维文本物理特征(含运动学、源区磁场、背景太阳风参数)合并；随后基于迁移学习的ResNet50模型提取图像深度特征，经降维得到5维图像特征，与文本特征拼接形成24维融合数据集。在此基础上，本文构建了XGBoost、随机森林模型，并进一步沿用二者的堆叠集成学习模型。结果显示，堆叠集成学习模型表现最优，在测试集上的平均绝对误差低至7.8048小时，R²达0.7613，显著优于两种单一模型，充分验证了数据融合的有效性与集成学习的互补优势。此外，本文还将该学习模型与经典的DBM模型进行对比分析，发现模型学习趋势与实际物理规律契合，从而检验了模型的有效性。

Abstract: Coronal Mass Ejections (CMEs), as highly destructive space weather phenomena, require precise forecasting of their propagation time to ensure the safety of spacecraft, satellite navigation, and the stability of power networks. Addressing the gap in existing research where LASCO observational images and textual physical parameters have not been fully integrated, this paper proposes a comprehensive data fusion scheme and forecasting model. Initially, through timestamp matching, we merge the time-series images of 255 CME events with 19-dimensional textual physical features, encompassing kinematics, source region magnetic field, and background solar wind parameters. Subsequently, we employ a transfer-learning-based ResNet 50 model to extract image deep features, which are then reduced to 5-dimensional image features. These features are concatenated with the textual features to form a 24-dimensional fused dataset. Based on this, this paper constructs XGBoost and Random Forest models, and further employs their stacked ensemble learning model. The results show that the stacked ensemble learning model performs optimally, with an average absolute error as low as 7.8048 hours on the test set and an R² of 0.7613, significantly outperforming the two individual models. This fully validates the effectiveness of data fusion and the complementary advantages of ensemble learning. Furthermore, this paper compares and analyzes this learning model with the classic DBM model, finding that the model learning trend aligns with actual physical laws, thus verifying the model’s effectiveness.