Presentations

Abstract

Introduction: Farm-to-farm wake impact modelling using advanced weather models like the Weather Research and Forecasting (WRF) Wind Farm Parameterization (WFP) is gaining momentum, offering impressive accuracy in capturing long wake phenomena, particularly in stable atmospheric conditions. These methods outperform traditional engineering models and have become more accessible with the declining costs of simulations. However, while single-site studies are now feasible, running dozens or even hundreds of scenarios—such as evaluating national or regional wind farm deployment strategies—is still cost-prohibitive. This creates a compelling case for leveraging deep learning to scale scenario analyses. In this study, we trained and deployed a convolutional neural network (CNN) capable of predicting wind and power fields based on WRF WFP simulations in New Mexico, USA. The goal was to create a tool that drastically reduces computational cost and time while maintaining high fidelity to WRF WFP results, enabling large-scale scenario analyses to inform strategic wind farm deployment. Methodology:  The project was executed in three key phases: 1. Data Preparation and Analysis: * Conducted 20 high-resolution WRF WFP simulations for wind farm layouts with diverse parameters, including variations in turbine capacity density, spacing, and deployment scale, covering a 1-km resolution domain in New Mexico. * Simulations were based on a representative Typical Meteorological Year (TMY), capturing long-term atmospheric conditions such as wind speed, wind direction, turbulence, and atmospheric stability. * Prepared a training dataset of hourly snapshots (175,000 samples), structured into arrays compatible with deep learning workflows.  2. Model Development: * Designed a CNN architecture optimized for spatial data analysis, featuring: * Convolutional layers for extracting localized wind and wake features. * Pooling layers to reduce dimensions while retaining critical spatial patterns. * Upsampling layers to maintain high spatial resolution for predictions across the domain. * Trained the model to learn the relationships between turbine layouts, meteorological inputs, and wake effects, iteratively optimizing hyperparameters to ensure accuracy and efficiency. * Validated the model using wind farm performance data at sites across New Mexico 3. Deployment and Application: * Use the trained CNN model to evaluate and optimize wind farm location and sizing in New Mexico. Results: The CNN model achieved high fidelity in replicating WRF WFP results, significantly reducing the computational cost and time required for scenario analyses. This approach facilitated rapid evaluation of interconnected impacts from multiple wind farm layouts, revealing valuable insights into how turbine density, spacing, and atmospheric conditions influence wake patterns and AEP losses on a regional scale. Conclusion: This study demonstrates the transformative potential of combining WRF-based wake modelling with deep learning for scalable wind farm scenario analysis. While full-year simulations are now affordable for single-site studies, the ability to train a CNN for large-scale deployment scenarios is a game-changer for renewable energy planning. This innovative approach bridges the gap between computational efficiency and accuracy, providing a valuable tool to optimize wind farm layouts and support informed decision-making at regional and national scales.