Forecasting Geomagnetically Induced Currents From Solar Wind and Geoelectric Field Data Using Deep Learning: Application in Fujian, China

Abstract

We developed a multi‐module convolutional neural network combined with a multi‐head attention mechanism (MM‐CNN‐MHA) to forecast geomagnetically induced currents (GICs) at the Huangmeishan (HMS) 220‐kV substation in Fujian Province, China. The model was trained using solar wind parameters, geomagnetic disturbances, and geoelectric field observations from 2010 to 2021. Five GIC estimation or prediction methods were evaluated: (a) GICs calculated from geoelectric fields predicted by MM‐CNN‐MHA; (b) direct GICs prediction using MM‐CNN‐MHA; (c) GICs calculated from measured geoelectric fields; (d) GICs calculated using a one‐dimensional (1D) conductivity model; and (e) GICs calculated using magnetotelluric (MT) impedance data. The results show that the GICs calculated from measured geoelectric fields achieved the highest accuracy, with a root‐mean‐square error (RMSE) of 0.24 A, and better reproduced the observed GICs peaks. The GICs calculated from MM‐CNN‐MHA‐predicted geoelectric fields outperformed the directly predicted GICs, with RMSE values of 0.26 and 0.28 A, respectively. In contrast, the GICs calculated using the 1D conductivity model and MT impedance method generally overestimate GICs peaks, with RMSE values of 0.52 and 0.39 A, respectively. These findings further highlight the importance of local geoelectric field observations for GIC monitoring and forecasting and demonstrate that preserving the physical chain from solar wind, geomagnetic disturbances, and geoelectric fields to GICs improves the stability and reliability of GIC predictions.