Long-Distance Wind Farm Wake Modelling Using Convolutional Neural Networks

Abstract Wake effects between neighboring wind farms can extend beyond 50 km downstream, yet traditional engineering models often underestimate these long-distance interactions, leading to significant power production losses. This study presents a novel approach using a Convolutional Neural Network (CNN), specifically a modular U-Net architecture, to predict wake deficits at spatial scales of up to 768 rotor diameters. The model introduces two key innovations: a pixel density method for converting continuous turbine coordinates into a CNN-compatible format, and a root wake deficit transformation that improves training stability. Trained on 1,000 diverse wind farm layouts generated using an engineering wake model (TurbOPark), our model achieves a mean absolute error of 0.35 percentage points compared to the test dataset while maintaining millisecond-scale execution times independent of farm size. The approach enables rapid evaluation of large-scale wind farm cluster interactions, providing a practical tool for optimizing layout designs and improving energy yield predictions. This work demonstrates how data-driven techniques can bridge the gap between computationally intensive high-fidelity simulations and simplified engineering models in wind energy applications.