Presentations

Abstract

Farm-to-farm wake impact modelling using advanced weather models such as the Weather Research and Forecasting (WRF) Wind Farm Parameterization (WFP) is gaining momentum due to its ability to accurately capture long wake phenomena, particularly under stable atmospheric conditions. These physics-based methods increasingly outperform traditional engineering models and are becoming more accessible as simulation costs decline. However, while single-site studies are now feasible, conducting dozens or hundreds of simulations—such as those required to evaluate regional or national wind deployment strategies—remains cost-prohibitive. This motivates the use of deep learning to scale such scenario analyses. In this study, we trained and deployed a convolutional neural network (CNN) to replicate WRF WFP-predicted wind and power fields in an active onshore wind development region. Our goal was to drastically reduce computation time and cost while maintaining high fidelity to WRF WFP results, enabling large-scale scenario analyses to inform wind farm siting and optimization. The project followed three key phases. First, we prepared a dataset of 20 high-resolution WRF WFP simulations that varied turbine capacity density, spacing, and deployment scale over a 1-km resolution domain, based on a representative Typical Meteorological Year. These simulations yielded 175,000 hourly samples formatted for deep learning workflows. Second, we developed a UNet architecture tailored for spatial data, incorporating convolutional, pooling, and upsampling layers to extract wake features and retain domain-wide resolution. The model was trained to map turbine layout and meteorological conditions to wake effects, with hyperparameters tuned for accuracy and efficiency, and was validated against actual wind farm performance. Finally, the trained CNN was used to evaluate and optimize wind farm location and sizing. The results show that the CNN closely replicates WRF WFP outputs, enabling rapid assessment of inter-farm interactions and revealing how layout and atmospheric stability affect wake losses and annual energy production at scale. This work highlights the transformative potential of combining physics-based wake modeling with deep learning to support cost-effective, high-resolution planning for regional and national wind energy development.