Extended Lead-Time Geomagnetic Storm Forecasting with Solar Wind Ensembles and Machine Learning

Geomagnetic storms are large disruptions of the magnetosphere, which can impact satellites, communications systems, and power grids, causing significant technological and economic impacts. Current forecasting models utilize L1 satellite data, constraining lead time to a few hours, often insufficient for effective mitigation. We investigate how to extend the lead times of these forecasts with solar data. Associated spatial and propagation uncertainties of solar data are captured with a solar-wind ensemble, of the computationally efficient one-dimensional HUXt numerical model. The solar-wind ensemble once propagated to Earth is processed through logistic regressions, weighting ensemble members by comparison with historical observed velocities, effectively filtering out high error ensemble members. Performance was evaluated across different storm intensities and lead times, demonstrating the models predictive capabilities in a variety of circumstances. Although not including transient phenomena such as Coronal Mass Ejections, our approach demonstrates strong predictive capability, achieving a Brier Skill Score relative to climatology BSSclim of 0.595 and a Receiver Operating Characteristic Area Under the Curve (ROC AUC) of 0.751 at 6-hr lead time for storms defined as Hp30MAX>5 within a 24-hr forecast window. Overall, these results highlight the strong potential of the coupled numerical model and machine learning framework to extend the forecast lead time for geomagnetic storms.