A neural network‐based local model for prediction of geomagnetic disturbances

Abstract

This study shows how locally observed geomagnetic disturbances can be predicted from solar wind data with artificial neural network (ANN) techniques. After subtraction of a secularly varying base level, the horizontal components X Sq and Y Sq of the quiet time daily variations are modeled with radial basis function networks taking into account seasonal and solar activity modulations. The remaining horizontal disturbance components Δ X and Δ Y are modeled with gated time delay networks taking local time and solar wind data as input. The observed geomagnetic field is not used as input to the networks, which thus constitute explicit nonlinear mappings from the solar wind to the locally observed geomagnetic disturbances. The ANNs are applied to data from Sodankylä Geomagnetic Observatory located near the peak of the auroral zone. It is shown that 73% of the Δ X variance, but only 34% of the Δ Y variance, is predicted from a sequence of solar wind data. The corresponding results for prediction of all transient variations X Sq + Δ X and Y Sq + Δ Y are 74% and 51%, respectively. The local time modulations of the prediction accuracies are shown, and the qualitative agreement between observed and predicted values are discussed. If driven by real‐time data measured upstream in the solar wind, the ANNs here developed can be used for short‐term forecasting of the locally observed geomagnetic activity.