Forecasting GICs and geoelectric fields from solar wind data using
LSTMs: application in Austria
Abstract
The forecasting of local GIC effects has largely relied on the
forecasting of dB/dt as a proxy and, to date, little attention has been
paid to directly forecasting the geoelectric field or GICs themselves.
We approach this problem with machine learning tools, specifically
recurrent neural networks or LSTMs by taking solar wind observations as
input and training the models to predict two different kinds of output:
first, the geoelectric field components Ex and Ey; and second, the GICs
in specific substations in Austria. The training is carried out on the
geoelectric field and GICs modelled from 26 years of one-minute
geomagnetic field measurements, and results are compared to GIC
measurements from recent years. The GICs are generally predicted better
by an LSTM trained on values from a specific substation, but only a
fraction of the largest GICs are correctly predicted. This model had a
correlation with measurements of around 0.6, and a root-mean-square
error of 0.7 A. The probability of detecting mild activity in GICs is
around 50%, and 15% for larger GICs.