Renewable energy sources are vital to meeting the world’s energy needs whilst mitigating climate change. Although wind power is already playing a prominent role in the UK’s energy supply mix, its use is expected to rapidly increase over the coming years. In particular, the UK government has the ambition to increase offshore wind power capacity fivefold by 2030, which will require building more, and ever larger, wind farms.
Building wind farms is an expensive undertaking, though: a typical offshore facility requires over £1 billion of capital investment, with the cost significantly increasing in recent months in line with supply chain inflation and rising interest rates. Policy makers, energy companies and funders therefore need to be confident that proposed future wind farms are designed in a way that maximises their energy production and minimises construction costs. This is the job of engineers, and their efforts are critical to ensuring wind farms are economically viable.
A key component of this endeavour are mathematical models of offshore wind farms – idealised representations of the physics of the wind and its interactions with the turbines. When implemented on a computer, these models can be used to predict a wind farm’s energy production. However, designing the models is a significant technical challenge. Behind every turbine is a wake region where the wind is slower than around other parts of the turbine. When these wakes impact the airflow around nearby turbines, they can reduce the overall farm power by up to 30%. Accurate modelling of the physics creating these wakes is therefore essential for designing wind farms.
The industry standard for approaching this challenge is a class of models called ‘engineering wake models’, where wakes from individual turbines are simply added together to estimate their overall impact. Unfortunately, though, the larger the wind farm, the larger the error from these models is likely to be: a significant problem given our needs for larger wind farms. An alternative modelling approach is ‘large eddy simulations’, which are more accurate but more computationally expensive. For context, on a laptop, one of these simulations would take over a year to run, and engineers typically need to consider thousands of these simulations to design a wind farm.
Our project – a collaboration between the Turing, the University of Oxford, University College London and the Met Office – is using machine learning (ML) to tackle this problem. Our ML model combines the results from a small number of expensive but accurate large eddy simulations with the results from many cheap but less accurate engineering wake model simulations. This is known as a ‘multi-fidelity’ approach because it uses simulations at different accuracy levels.
This multi-fidelity ML model is around twice as accurate at predicting the effects of different turbine layouts as using the small number of large eddy simulations alone. It is also more than five times more accurate than using only the engineering wake models. Our approach therefore offers a novel way to accurately predict the energy output of an offshore wind farm. It could be used to design more efficient wind farms, substantially reducing the cost of wind energy in the UK. Our next step is to further improve the model’s accuracy by incorporating weather data, which will allow us to take into account the impact of atmospheric conditions such as wind turbulence on wind farm performance.
Read the paper:
Data-driven modelling of turbine wake interactions and flow resistance in large wind farms
About the project:
This research is part of the Fundamentals of statistical machine learning research project in The Alan Turing Institute’s data-centric engineering programme. It is also supported by Met Office Academic Partnerships between the Met Office and University College London and the University of Oxford. Andrew Kirby is funded by a studentship from the NERC-Oxford Doctoral Training Partnership in Environmental Research.
Top image: Sue Burton / Shutterstock
