Efficient Assessment of Lightning Outages in Transmission Lines Using Surrogate Models

Estimating lightning disruptive effect (DE) in overhead transmission lines is crucial for insulation coordination and reliability assessment. However, conventional approaches, which usually rely on the Monte Carlo method, are computationally expensive, limiting their practical use. To overcome this limitation, this article proposes an approach that replaces part of the transient simulations with a machine learning (ML) algorithm, aimed at estimating the DE of direct lightning strikes. For this purpose, the ML algorithm is trained using samples randomly extracted from a dataset generated via an electromagnetic transients program-type, the alternative transients program (ATP), considering key input parameters, including current peak, phase angle at the moment of the strike, grounding impedance, and point of incidence along the line span. A Gaussian Process Regression model is adopted, as it naturally provides both the estimated value and its associated confidence interval. In the proposed approach, this aspect is exploited so that, if a prediction falls within a region of high uncertainty, an additional ATP simulation is executed to refine the result. The performance of the algorithm is computed using traditional metrics in the regression field. Its robustness is assessed by repeating multiple times the extraction of the samples to be used to train the ML algorithm and analyzing the related outcomes. The case study considered refers to four scenarios, differing in the soil resistivity and in the grounding impulsive impedance. According to results, the combination of the ML algorithm with the eventual additional ATP simulation allows obtaining accurate DE estimates and achieves average speedup of around 25 times.