Speaker
Description
Relativistic electrons (>2 MeV) in geostationary orbit (GEO) are of great scientific interest because, during disturbed geomagnetic conditions, their fluxes can increase by several orders of magnitude on timescales of just hours. This phenomenon causes internal charging in satellites, damaging their electrical circuits and affecting their operation. Various studies have shown that the flux of relativistic electrons in GEO presents a strong correlation with different geomagnetic indices, as well as with plasma parameters and the magnetic field of the solar wind.
In this work, a predictive model based on Deep Learning techniques is developed, using Convolutional Neural Networks (CNN) to forecast relativistic electron fluxes of >2 MeV in GEO, comparing single-output and multi-output configurations. The model was trained with historical electron flux data measured by GOES satellites through the EPEAD instrument, along with solar wind parameters from the OMNIWeb database. The model performance was evaluated for prediction horizons of 1, 12, and 24 hours. Likewise, models trained exclusively with data corresponding to periods of geomagnetic storms were implemented, analyzing predictive capability under disturbed conditions.
The results show that CNN-based models, when compared to each other, present consistent performance across different prediction horizons, maintaining similar RMSE values. Likewise, no significant difference is observed between single-output and multi-output configurations. On the other hand, when training exclusively with geomagnetic storm data, a deterioration in the overall performance of the models is observed, reflecting the difficulty of generalization under disturbed conditions. In this scenario, for a 24-hour horizon, the multi-output approach shows a better ability to capture the system dynamics compared to the single-output approach.