About the project or challenge area

Given the complexity and stochastic nature of the relevant physics, it is possible that the use of machine learning algorithms could provide a practical approach to the optimization of performance of wind farms on a project scale. We want to adapt largely autonomous data-driven new control strategies for wind turbines based on existing sensor information to optimize wind farm scale performance. We will investigate the use of historical wind farm measurements to develop a non-physical model of the system, using a statistical approach to the prediction of future performance from information contained in the wind farm data. A supervised learning algorithm will be utilized to analyse the wealth of operational data recognizing conditional state patterns. This will be coupled with additional expert knowledge of wind farm operation to form the basis of a model to optimize power output and minimize fatigue loading for control.

We propose a model based on training data from the DNV GL dynamic wind farm simulator and real wind farm data from public sources. Wind farm data is used in cross-validation and parameter tuning, thus eliminating any bias from simulated environments. The use of high-frequency SCADA data holds potential for real accuracy gains within this framework. A data-driven machine learning model benefits from real-time updated learning with incoming operational data, reduced computational time for implementation, and a lack of dependence on physical model assumptions. The research problem is nontrivial as it involves complex data, which are characterized by nonlinearity, nonstationary and high dimensionality. The state space is very high-dimensional, which leads to problems in many machine learning methods. Solution methods must deal with incomplete system observations. SCADA has missing or erroneous data. Controllers work on live feeds. Another challenge is given by the existence of different time scales in typical real systems. In particular, both the quick and the slower dynamic components must be captured without devaluing each other as noise.

Why choose this project?

The project will allow you to conduct state-of-the art machine learning research in partnership with DNV GL, a world leading renewable energy consultancy. 

About you

This project would suit a candidate with strong interests in wind energy, machine learning and control methods.

How to apply

All students can apply using the button below, following the Admissions Statement (PDF, 188kB). Please note that this is an advertised project, which means you only have to complete Section A of the Research Statement.

This project is not funded, for further details please use this link.

 

Before applying, we recommend getting in touch with the project's supervisors. If you are interested in this project and would like to learn more about the research you will be undertaking, please use the contact details on this page.