Posters

Abstract

In this work, we propose two methods for building neural-network based surrogate models for wake-induced effect estimation in large wind farms. The key innovation is the approach used for wake effect parameterization - i.e., how the wake conditions experienced by a given turbine are encoded as variables suitable to serve as inputs for a surrogate model. Two approaches are used to obtain these variables describing the wake sources. In the first approach, two parameters per turbine (e.g. crosswind and upwind distances) are used to describe the position of up to N closest upwind turbines within a given relative wake angle, for example, 16 degrees. In this study, up to 20 closest upwind turbines are chosen. If there are fewer than 20 closest turbines, the unnecessary variables are set to zeros. As a result, a vector of 40 variables is used to describe the wake sources. This approach can be applied to both regular and irregular wind farm layouts. The second approach is based on only four latent variables obtained by encoding the above vector of variables obtained in the first approach for describing the wake sources. These four latent variables are obtained with a novel deep learning neural network model. It is similar to an autoencoder in the sense that it also has a bottleneck layer for the purpose of dimensionality reduction and feature learning. However, the proposed deep learning neural network model does not map the input to an approximation of the input itself as it would be the case with a classical autoencoder. In contrast, our model maps the inputs directly to the wake-induced effect on turbine power while passing the information through the bottleneck layer. This enables the latent variables to represent the wake sources not purely based on their geometric configuration but also based on their physical effect, i.e. the wake-induced effect on the power. In the training of this deep learning neural network model, 20 randomly generated wind farm layouts are used with a mock turbine model. After training, the trained encoder part, i.e. from the input to the bottleneck layer, can be applied to a new wind farm without retraining. This approach can also be applied to both regular and irregular wind farm layouts. These two approaches for parameterization of the wake sources are applied to build surrogate models for an operational offshore wind farm. In addition to the variables describing wake parameterization, the ambient wind speed is also included as an additional input for model training. The data for wake-induced effects on turbine power is generated using the wind farm simulation tool PyWake which supports engineering wake models and turbulence models. The results show that both approaches used to describe the wake sources result in high-performance neural network models in terms of correlation and error between simulated and estimated data. However, the encoding of the wake sources into latent variables enables a smaller number of input variables and a simpler neural network model, and requires less data for training.