Creation of input data sets when constructing neural network models

The paper solves the problem of reducing the dimension of input features when constructing neural models for the SuperMAG electrojet index forecast. In order to decrease the input data dimension, estimates of correlation indexes values and computations of neural network errors are used. Feature sets providing the least error for the SuperMAG electrojet index modeling are obtained. Modelling of the geomagnetic data was performed based on the obtained feature sets. Recurrent neural networks were used for modelling. The dependence of the modelling quality on the input data is shown.