Scholarlink

<p>This study aims to enhance the accuracy of arrival-time forecasts for coronal mass ejections (CMEs) by employing a hybrid model that integrates deep learning (DL) and machine learning (ML) techniques. The primary goal is to minimize errors and improve the reliability of space weather forecasting, particularly in scenarios characterized by data limitations and noise. The model was composed of two stages: a DL stage at the front end, and a Random Forest (RF) algorithm at the back end. Dynamic input features were extracted from time-series CME physical and solar wind datasets. The stochastic gradient descent with momentum (SGDM) algorithm was chosen to optimize the model. Tuning hyperparameters such as long short-term memory (LSTM) parameters, batch normalization layers, and dropout layer configurations were applied. After the deep learning stage, the RF stage uses time-related features to predict the CME arrival time. In addition, the LSTM component recognizes time-series data, whereas the RF model enhances its ability and robustness. Compared with previously proposed models that use physical and solar wind parameters, our model can predict the CME arrival time more accurately. Our model achieves a mean absolute error (MAE) of the arrival time of only 8.1&#xa0;h. The results indicate that mixing time-based patterns with group learning greatly improves the forecasting of CME arrivals, which will aid future real-time, data-driven space and weather efforts.</p> Graphical Abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CME arrival time forecasting based on dynamic physical and solar wind parameters using a LSTM-RF model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月19日

查看全文
  • Pronthep Pipitsunthonsan,
  • Wanida Panup,
  • Rattachai Wongtanawijit,
  • Sittiporn Channumsin,
  • Damrongrit Niammuad,
  • Mitchai Chongcheawchamnan

摘要

This study aims to enhance the accuracy of arrival-time forecasts for coronal mass ejections (CMEs) by employing a hybrid model that integrates deep learning (DL) and machine learning (ML) techniques. The primary goal is to minimize errors and improve the reliability of space weather forecasting, particularly in scenarios characterized by data limitations and noise. The model was composed of two stages: a DL stage at the front end, and a Random Forest (RF) algorithm at the back end. Dynamic input features were extracted from time-series CME physical and solar wind datasets. The stochastic gradient descent with momentum (SGDM) algorithm was chosen to optimize the model. Tuning hyperparameters such as long short-term memory (LSTM) parameters, batch normalization layers, and dropout layer configurations were applied. After the deep learning stage, the RF stage uses time-related features to predict the CME arrival time. In addition, the LSTM component recognizes time-series data, whereas the RF model enhances its ability and robustness. Compared with previously proposed models that use physical and solar wind parameters, our model can predict the CME arrival time more accurately. Our model achieves a mean absolute error (MAE) of the arrival time of only 8.1 h. The results indicate that mixing time-based patterns with group learning greatly improves the forecasting of CME arrivals, which will aid future real-time, data-driven space and weather efforts.

Graphical Abstract