Diagnostics of Geoinduced Currents in High Latitude Power Systems Using Machine Learning Methods

It is known, that geoinduced currents (GICs) take place in spatially distributed current-carrying technical systems (main pipelines, power transmission lines and telegraph lines, railway infrastructure facilities, etc.) due to geomagnetic variations (GMVs), which rate of change in high-latitude regions is often about several hundred nT/min. Flowing through the grounded windings of power transformers of system-forming electrical circuits, extreme GICs are capable of transferring their magnetic systems into saturation mode, which, in turn, can cause a failure of the corresponding electrical systems. However, due to little knowledge of the mechanisms of the emergence and development of GIC, as well as the fragmentation and heterogeneity of the available empirical data, the problem of their predicting and diagnostics today is associated with many uncertainties and remains practically unsolved. The research based on machine learning methods examines approaches to diagnostics of gas and electric power level in the electric network ‘‘Severnyi Transit.’’ In this case, both geomagnetic data recorded by magnetic stations in the subregion (Kola Peninsula, Russia) and natural (visible) indicators of extreme geomagnetic activity are used as input parameters. Using an annual sample of more than 35 000 records as an example, it was shown that the approach to GICs diagnostics, based on multiple linear regression, provides a root mean square error (RMSE) of \(\sim\) 0.122 A \({}^{2}\) . The use of an artificial neural network with the ReLU activation function can slightly improve the diagnostic accuracy (RMS \(\sim\) 0.119 A \({}^{2}\) ). However, the interpretability and theoretical significance of the model is significantly reduced. The application, in turn, of the Bayesian classifier to the data of optical observations of auroras showed that the posterior probability of the fact that in the north the GIC level at the Vykhodnoy station during auroras will exceed 2 A is 5.78 \(\%\) , while the probability of exceeding this value during auroras in the zenith and south are 10.04 and 14.93 \(\%\) , respectively. In the absence of auroras, the model indicates that the probability of achieving a GIC of a similar level does not exceed 0.26 \(\%\) , and the probability of exceeding 3 A is practically zero.