The geomagnetic index K is a fundamental metric for quantifying geomagnetic activity and understanding solar-terrestrial physics. Accurately modeling and predicting its non-linear fluctuations remains a significant challenge in the study of space climate. In this research, we propose a robust statistical approach using a Long Short-Term Memory (LSTM) neural network for the time-series forecasting of the K index. Our models are trained and evaluated on a comprehensive dataset spanning a 25-year period from 1999 to 2024, utilizing high-resolution solar wind data sourced from the NASA OMNIWeb database. We incorporate eight critical physical parameters as input features, including three-dimensional magnetic field components and plasma properties. A critical aspect of sequential modeling is determining the optimal historical context. Therefore, we systematically evaluate the impact of various input time windows (6, 12, 24, 48, 72, and 120 hours) on the predictive performance of the network. Our empirical analysis reveals that the 6-hour input window yields the highest precision, achieving an R² score of 0. 7233 and minimizing the root mean square error (RMSE). These results indicate that short-term solar memory contains the most relevant dynamical features for forecasting geomagnetic variations. This study highlights the effectiveness of LSTM architectures in capturing the short-term dynamical evolution of the Earth's magnetic environment, offering valuable insights for future data-driven research in space physics.
Ranjbar et al. (Thu,) studied this question.