back to list

Project: Efficient surrogates for the Ainslie wind turbine wake model


In wind farms, one source of reduction in power generation by the turbines is the reduction of wind speed in the wake downstream of each turbine's rotor. Namely, a turbine downstream in the wind direction of another will effectively experience wind with a reduced speed. This reduction in wind speed results in a reduction in power production relative to the situation where the wind turbine would not be in that wake of another one.

Accurately and efficiently quantifying the effect of such wakes is important both for estimation of a wind farm's power production and optimizing the design of wind farms. Namely, by judiciously choosing the location of the farm's turbines (its layout), the effect of the wake on power production can be reduced relative to arbitrary regular or random placements. For estimating the farm's power production after it has been designed, the focus is on estimation accuracy. For estimating the wake effect during the farm's design, the computational efficiency of the wake model becomes more important, and a trade-off with accuracy must be made, because there is insufficient time during the design stage to optimize the layout using wake models with state-of-the-art accuracy.

The Ainslie wake model is a popular wake model in industry. It is used both for the estimation of power production and for wind farm layout optimization. It is a differential equation model, a simplification of the Navier–Stokes equations, and therefore requires a sequential solver to determine the entire wake based on specific initial conditions. There exist other wake models that are defined by direct analytical expressions for the wake, which means that the quantities defining the wake—wind speed, primarily—can be directly calculated at any downstream point. Such engineering models therefore have a computational efficiency advantage over the Ainslie model, which, however, is considered to be more accurate.

In this project, the goal is to derive surrogates for the Ainslie model and investigate their performance relative to the Ainslie model itself and engineering wake models, both in terms of accuracy and computational efficiency. A surrogate model is a model with the same inputs and outputs of the original model, but which is much simpler, computationally speaking. Given data from the original model, it approximates it efficiently and thereby makes it possible to perform analyses or activities that would otherwise be computationally infeasible.

The relevance for industry of such an investigation is the following:

  1. If surrogates are obtained that make a good trade-off between these two objectives, it may replace the typical engineering models in wind farm design tools, contributing to more efficient wind farms.
  2. If no surrogates are found that make a good trade-off between these two objectives, the analysis will have generated insight into what other investigations in this research area are most promising.

We have a contact in industry who is specifically interested in results of and insights gained from this project.
They will be involved as outside advisor. So while the project is conducted at the TU/e, there is a direct link to industry to strengthen its societal relevance.

The approaches to the definition of the surrogate models are not fixed in advance, but their determination forms part of the project. However, there needs to be some diversity in their nature, e.g., simple versus advanced, so that the insights needed for point 2 above can be generated. Example approaches are linear regression models, Gaussian process models, and deep learning models.

Requirements & Activities
This project will require you to immerse yourself in both the field of wake models and of surrogate models. It will require the mathematical maturity to understand and work with both classes of models, including learning the latter from data. The data used to create the surrogate models will also need to be generated during the project. Both the  data generation, model learning, and analysis will require a familiarity with scientific programming, e.g., in Python.

Erik Quaeghebeur
Secondary supervisor
Laurens Bliek (Information Systems, IE&IS)