Modelling of turbine power and local wind conditions in wind farm using an autoencoder neural network

Abstract

Neural network models have been widely applied in wind energy to enable data-driven modelling of turbine’s power, loads, damage, and other quantities in wind farm. They have shown advantages in the flexibility, robustness, and efficiency, as well as the ease of application. An important task in data-driven modelling of wind farm is the characterization of wake-induced effects within wind farm. Because of the large number of wind turbines in wind farm and the broad range of conditions required for power output assessment, it has become relevant to use computationally efficient surrogate models that are based on parameterizations of the wind farm layout. Previous studies on surrogate models of wake-induced effects have focused on parameterizations based on manually extracted features from the wind farm geometry. The present study applies an autoencoder neural network to automatically learn the features representing the wind farm layout. Specifically, two autoencoders are studied and compared. The autoencoder provides dimension reduction and feature learning that allow representing the entire wind farm layout in just a few latent variables, thus allowing efficient and accurate surrogate models of wake-induced effects. The outputs of the autoencoder, together with any ambient wind condition variables, serve as inputs to a standard feedforward neural network that predicts the wake-induced effects for any turbine in the wind farm. Our results show that the autoencoder provides an automatic and effective way to parameterize the wind farm geometry, and facilitates the building of efficient surrogate models that are capable of accurate modeling of wake-induced effects in a wind farm with arbitrary layout.