Modelling global vertical total electron content with machine learning

This study introduces a new two-step neural-network based approach for modelling vertical total electron content (VTEC) on a global level. The inputs to the neural network are chosen and evaluated through different feature selection techniques, namely time-lagged Pearson cross-correlation, mutual information, random forests and permutation feature importance. The feature sets consist of geomagnetic and solar wind indices, their time histories and geomagnetic and geographic coordinates. The parameters of the neural networks are tuned with cross validation and the final model is tested in extended time intervals covering a wide range of solar activity conditions. The proposed approach increases computational efficiency and provides with a high generalization skill.