Abstract
Geomagnetic storms pose a significant risk to our increasingly technology-dependent society, with the potential to disrupt power grids, satellite operations, and global communication systems. These disturbances are primarily driven by solar wind, a stream of charged particles emitted by the Sun, which interacts with the Earth's magnetosphere. The dynamic nature of solar activity, notably characterized by the solar cycle's periodicity of approximately 11 years, further complicates the prediction and understanding of such events. A critical challenge in space weather forecasting is the development of reliable models that can accurately predict geomagnetic disturbances in a timely manner. Traditional modeling approaches often struggle to capture the complex temporal dependencies and the multivariate nature of solar wind data. To address this challenge, we developed a Long Short-Term Memory (LSTM) neural network model designed to harness the sequential nature of solar wind measurements. This study leverages LSTM's capability to retain long-term temporal relationships, utilizing a dataset composed of various solar wind parameters collected from NASA's ACE and NOAA's DSCOVR satellites. The model was trained to predict the disturbance storm time (Dst) index, a measure of geomagnetic activity, using a feature set that included interplanetary magnetic field components, solar wind proton density, speed, and ion temperature. The LSTM model demonstrated a substantial learning capacity, evidenced by a Root Mean Squared Error (RMSE) of 14.25, indicating strong predictive performance. In a binary classification setup, the model achieved an accuracy of approximately 80.35% with an Area Under the Curve (AUC) score of 0.818, signifying its effective discrimination between high and low geomagnetic disturbance events. These results underscore the potential of utilizing LSTM models for space weather forecasting, which could significantly enhance our ability to mitigate the risks associated with geomagnetic storms. The incorporation of sunspot data could allow for calibration to the solar cycle, further refining model predictions. This study lays the groundwork for future research aimed at integrating more diverse data sources and applying advanced machine learning techniques to improve the accuracy and reliability of geomagnetic disturbance forecasts.