Main content area

Gaussian process regression modeling of wind turbines lightning incidence with LLS information

Sarajcev, P., Jakus, D., Mudnic, E.
Renewable energy 2020 v.146 pp. 1221-1231
Bayesian theory, artificial intelligence, autocorrelation, business enterprises, insurance, landscapes, lightning, models, normal distribution, probabilistic risk assessment, renewable energy sources, stochastic processes, storm damage, uncertainty, wind farms, wind turbines
This paper presents a machine learning (ML) approach to wind turbine (WT) lightning incidence analysis in complex terrain, based on the information obtained from a lightning location system (LLS). A particular ML model of the WTs lightning incidence is developed, using Bayesian statistical learning and Gaussian process regression, and trained on the actual LLS data. The model is developed around a known proposition that the lightning strike frequency data are emanating from a Poisson stochastic process. It further makes use of an attractive radius concept of lightning attachment, employs a sophisticated analysis of the WT effective height—which leverages terrain elevation data—and introduces spatial autocorrelation of lightning strikes. It provides a probabilistic risk assessment of WT lightning damage, along with a statistical measures of the associated monetized financial losses. Proposed ML model benefits from the Bayesian ability to quantify uncertainty of model parameters, and employ hierarchical model structure that informs model parameters through the mutual higher-level hyperpriors. Proposed model enables both investors and insurance companies to asses risks associated with lightning incidence to WTs, considering historical LLS data and future wind farm installation locations.