Characterizing Geomagnetic Storm Data for Machine Learning Models of Geomagnetically Induced Currents

Geomagnetically Induced Currents, GICs, can negatively affect the power grid as well as other electronic components and conductors on Earth. The MAGICIAN project is trying to use machine learning-based models to predict ground magnetic fluctuations, to be used as proxies for GIC occurrences. In order to train these neural networks, sufficient data is needed from the solar wind, particularly during geomagnetic storms where it is more likely for GICs to occur. Unfortunately, although geomagnetic storms occur frequently, they still represent less than 10% of the time, which results in very biased datasets. The problem lies in isolating the correct data to use to train these neural networks. In this work, we evaluate different ways to separate the data and identify patterns that would be useful forecasting the magnetic field perturbations (dBH/dt spikes). We considered the effects of evaluating different lengths of storms, MLT dependence, latitudinal dependence, geomagnetic index dependence among others to understand how to improve existing machine learning-based models.