One-Hour-Ahead Horizontal Geoelectric Fields Forecast Using Uncertainty-Quantification-based Machine Learning method

A fundamental but still unresolved problem in the Geospace environment is forecasting the occurrence of Geomagnetic Induced Current (GIC) calculated by horizontal geoelectric fields (including Exand Ey). Instead of ∆dB/dt , we develop a new model to directly forecast Ex and Ey 1-hour-ahead using advanced uncertainty-quantification(UQ) based Machine Learning (GRU) method, SuperMag data and Magnetotellurics (MT) survey results.

Ex and Ey are retrieved from SuperMag data and Magnetotellurics (MT) survey results over SuperMag stations inside Unite States with 1-min resolution. We train two separated models (Ex and Ey) for each station. A simple ML model and SWMF model are used for comparisons. Generally, the developed model significantly outperforms both simple ML model and SWMF model. Finally, the SHapley Additive exPlanations (SHAP) method is used to investigate the importance ofeach parameter to horizontal geoelectric fields.