Relating transmission line overvoltages and lightning location: a machine learning–based procedure

Purpose This study aims to investigate the use of machine learning–based algorithms in the field of lightning stroke localization. This work is an important step ahead with respect to the research recently started by the authors, i.e. the possibility of locating a lightning discharge from the voltage induced on overhead transmission lines; more in detail it seeks new insights into the inclusion of both first and subsequent return strokes of negative cloud-to-ground flashes. Design/methodology/approach This study uses a quantitative approach, using supervised learning techniques for a regression problem (data preprocessing, model selection, training, testing, validation or algorithm optimization). Data are collected from a dedicated lightning-induced overvoltage simulator and analyzed using a specific machine learning–based procedure developed and programmed by the authors. Findings The results reveal significant improvements in localization accuracy for both first and subsequent strokes, with respect previous works, indicating that the novel approach is promising for future investigation with more complex power system configurations and the use of experimental data. These findings provide evidence that dedicated models for each type of stroke yield better performance, offering significant implications for the integration of machine learning–based lightning location systems into the existing power infrastructure. Originality/value The proposed method is, to the best of the authors’ knowledge, entirely new and constitute an innovation with respect to the present literature, both of the same authors and of other research groups. In particular, a new preprocessing procedure of the voltage data is proposed, and the performances of different neural networks are evaluated, both for the first and the subsequent stroke. The application to both first and subsequent stroke is an innovation itself because it has not been proposed before.