Algorithm for Generating Synthetic Datasets of Lightning Flashovers on Distribution Lines
Abstract
This paper describes an algorithm for generating synthetic datasets of lightning flashovers on medium voltage overhead distribution lines (OHL). Datasets are heterogeneous, hierarchical and class-imbalanced. They are designed specifically for training machine learning (ML) models (i.e. binary classifiers) for predicting lightning flashovers on OHLs, and they are built from both direct and indirect lightning strikes. Algorithm for creating datasets employs the Monte Carlo method and makes use of the several underlying mathematical models of lightning interaction with the OHL, which introduce noise and make training of the ML models more challenging. Algorithm is versatile enough to account for the different statistical properties of lightning interactions with distribution lines, starting from the bivariate probability distribution of lightning-current parameters. Example of the synthetic dataset will be presented, using a typical medium voltage OHL geometry with a horizontal conductors arrangement.
