Speaker
Description
Providing actionable forecasts of geomagnetic storm occurrence on timescales of several hours is essential for space weather services aimed at protecting critical infrastructure and supporting operational decision-making. We present GeoStormAlert, a machine-learning-based forecasting system that leverages real-time in-situ solar-wind measurements at L1 to predict geomagnetic storm conditions with lead times up to four hours.
GeoStormAlert ingests high-cadence solar-wind plasma and interplanetary magnetic-field observations, together with derived physically meaningful parameters such as magnetic energy and helicity. These upstream quantities describe the drivers of magnetospheric disturbances. The inputs are processed by a Long Short-Term Memory (LSTM) neural network designed to capture temporal dependencies and the evolving solar-wind–magnetosphere coupling. To enhance robustness and interpretability, feature-ranking techniques are applied to assess the relative importance of input variables and to identify the most predictive subset across multiple forecast horizons.
The system generates warning alerts when geomagnetic indices are forecast to cross operational thresholds associated with moderate to intense storms. By delivering multi-horizon alerts, GeoStormAlert enables timely, risk-informed decisions for satellite operators and infrastructure managers.
| Numerical model | GeoStormAlert |
|---|
