A forecast model of geomagnetic indices from the solar wind fluids observations based on long short-term memory neural network

Abstract

Geomagnetic indices including AE (Auroral Electrojet), AU (Upper envelopes of AE), AL (Lower envelopes of AE), and Dst (Disturbance Storm Time) are widely considered signatures of geomagnetic storms or substorms that are triggered by solar wind plasma fluids and magnetic fields impinging the Earth's magnetosphere. They are crucial for comprehending and predicting the particle dynamics within the near-Earth space. In the present study, a forecast model utilizing a Long Short-Term Memory (LSTM) Neural Network was constructed by training more than five decades of measurements of solar wind parameters near the first Lagrangian point (L1). Note that it is the first deep learning model to forecast the AE, AU, and AL indices, the overall correlation coefficient (R) between the 1-h-ahead forecast results and the observations reached > 0.85, and the root mean squared error (RMSE) reached 0.05. For the Dst prediction, the model achieves better performance with R and RMSE values reaching 0.981 and 0.011, respectively. It also exhibited an extremely high level in predicting 3-h-ahead Dst, as indicated by R and RMSE values of approximately 0.91 and 0.023, respectively. Further analysis presents that the predictions of the model can be used to precisely track the overall changes in AE and Dst during both quiet times and different stages of geomagnetic storms. It can be easily implemented to supply essential guidelines for examining and estimating the variations in inner magnetospheric particles and solar wind-magnetosphere couplings.