Posters

Abstract

In order to make accurate predictions of off-shore wind turbine power production, numerical models are developed using Computational Fluid Dynamics (CFD). However, these are computationally expensive and must typically be repeated for each configuration and wind direction. Here, we propose a novel approach of a Machine Learning (ML) surrogate model that allows the prediction of wind speeds based on turbine position. This is done by encoding the turbine surroundings in descriptors in a way that is independent of the number of turbines in the environment and maintains positional and directional invariance. In this way we can assign a fingerprint to each turbine, that is a set of three computationally inexpensive descriptors depending on relative distance and angle between turbines and the wind direction. We then trained a Gaussian Process (GP) surrogate model with CFD simulation data of two- and three-turbine configurations to predict the wind speed at each turbine in dependence of its fingerprint. In unseen test data, the maximum absolute error rate of the prediction was 0.35 m/s, within the precision bounds of the CFD data. This approach could be used to provide fast high-fidelity predictions of CFD wind speeds in off-shore wind farms.