← Back to Use Cases

Machine Learning for Wind Energy Modelling

Authors:
Name Affiliation Phone Number Email Address
David Standingford Zenotech david.standingford@zenotech.com
Industrial Sectors:

Energy

1. DESCRIPTION OF USE CASE

The United Kingdom is one of the best locations for wind power in the world, and is considered to be the best in Europe. Wind power contributed

11% of UK electricity generation in 2015, and 17% in December 2015.

Allowing for the costs of pollution, onshore wind power is the cheapest form of energy in the United Kingdom. The use of computational fluid dynamics (CFD) to assess the wind energy re- source for a prospective new site is an established method that has been used for many years, however the inclusion of wake interaction effects - particularly for larger arrays of turbines - is less mature.

The SWEPT2 consortium (led by DNV-GL, with SSE, ORE Catapult, STFC, CFMS, Zenotech and the universities of Surrey, Strathclyde, Bristol and Imperial College) has been developing new tools to improve the utility and accuracy of CFD-based wake interaction modelling. Large amounts of data can now be generated and consolidated to inform engineering and investment decisions.

Objectives: The consortium is interested to know how well machine learning methods might be able to infer the interaction patterns and consequent power production from sites, given suffi- cient training data.

2. KEY UQ&M CONSIDERATIONS
2.1 Process Inputs

The datasets include a range of wind directions, terrain maps (including roughness parameters and tree canopies) and full CFD fields (millions of data- points, including velocity, pressure, temperature and turbulence quantities of interest). We will provide CFD data for a wind farm site in a Python notebook, with links to supporting datasets as needed.

2.2 Propagation

If the study group wants to run additional simulations then the software can be made available - either on a local HPC cluster or on ARCHER. We will challenge the study group to reconstruct the wind data for a wind direction that is not part of the original dataset. This will requires the machine learning system to develop a an internal representation of the principal wake interaction modes as a function of wind direction, including nonlinear fluid dynamics effects and terrain!

2.3 Interpretation and Communication of Results

3. CURRENT STATE OF MATURITY

References: