Wind farm layout optimization using a novel machine learning approach
Mehrshad Gholami Anjiraki, Christian Santoni, Samin Shapourmiandouab, Hossein Seyedzadeh, Jonathan Craig, Fotis Sotiropoulos, Ali Khosronejad · 2025
Read Wind farm layout optimization using a novel machine learning approach by Mehrshad Gholami Anjiraki, Christian Santoni, Samin Shapourmiandouab, Hossein Seyedzadeh, Jonathan Craig, Fotis Sotiropoulos, Ali Khosronejad (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-07868.txt.
Introduction
Where turbines should stand inside a wind farm is an optimisation puzzle taken up by Mehrshad Gholami Anjiraki and 6 co-authors in a 2025 engineering paper. Turbines steal wind from neighbours behind them, so layout decides total yearly energy. Accurate wake simulations are costly, while fast simplified models lose detail.
The team blends both worlds. High-fidelity large-eddy simulations and a cheaper Gaussian-curl hybrid model feed an autoencoder-based neural network that predicts the wake flow field reliably at low cost. With quick predictions available, the search over layouts can aim directly at maximum annual energy production. Better layouts squeeze more electricity from the same land and wind, a practical gain as wind farms spread.
Catalog record
| CREATOR | Mehrshad Gholami Anjiraki, Christian Santoni, Samin Shapourmiandouab, Hossein Seyedzadeh, Jonathan Craig, Fotis Sotiropoulos, Ali Khosronejad |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 69 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-07868
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
This article has been submitted to Applied Energy. Final version will appear at https://www.sciencedirect.com/journal/applied-energy Wind farm layout optimization using a novel machine learning approach Journal: Applied Energy Mehrshad Gholami Anjiraki, Christian Santoni, Samin Shapourmiandouab, Hossein Seyedzadeh, Jonathan Craig, Fotis Sotiropoulos, Ali Khosronejad Abstract We present a novel approach to optimize wind farm layouts for maximum annual energy production (AEP). The optimization effort requires efficient wake models to predict the wake flow and, subsequently, the power generation of wind farms with reasonable accuracy and low computational cost. Wake flow predictions using large-eddy simulation (LES) ensure high fidelity, while reduced-order models, e.g., the Gaussian-curl hybrid (GCH), provide computational efficiency. We integrate LES results and the GCH model to develop a machine learning (ML) framework based on an autoencoder-based convolutional neural network, allowing for a reliable and cost-effective prediction of the wake flow field. We trained the ML model using high-fidelity LES results as the target vector, while low-fidelity data from the GCH model serve as the input vector. The efficiency of the ML model to predict the AEP of the South Fork wind farm, offshore Rhode Island, was illustrated. Then, we integrated the ML model into a greedy optimization algorithm to determine the optimal wind farm layout in terms of turbine positioning. The optimized wind farm layout is shown to achieve a 2. 05% improvement in AEP over the existing wind farm. Highlights • A novel machine learning algorithm is proposed that allows efficient power prediction of utility-scale wind farms. • Validation of the machine learning model against large-eddy simulation results for an unseen utility-scale wind farm illustrates the reliability of the model. • The machine learning model is integrated into the Greedy algorithm to allow for precise and low-cost optimization of turbine positioning within a wind farm. • The integrated optimization approach obtains the optimal positioning of individual turbines within a farm to maximize its annual energy production. • Applying the optimization approach to an existing offshore wind farm, it was shown that the optimized wind farm layout achieves over 2% improvement in annual energy production. Corresponding author: Corresponding author: [email protected] 1 Introduction Over the millennium-long history of wind energy [ 1 , 2 ] , considering it as a widespread renewable resource is a relatively recent development. From 1990 to 2021, wind energy generation rose from 3.6 billion kWh in 16 countries to 1808 billion kWh in 128 countries [ 2 ] . Bojek 2025 projected that wind energy will continue to grow from 30% in 2023 to 46% of renewable energy production in 2033, surpassing hydropower to become second only to solar energy. To become a competitive global renewable resource, there has been an exponential optimization effort in turbine modeling since the latter half of the 20th century. As asserted by Chen et al. 2016 , wind farm layout optimization is pivotal to the efficient use of wind power as a global renewable resource. The simple analytical wake model established by Jensen 1983 served as the foundation for subsequent models aimed at optimizing wind farm power generation and advancing optimization technologies [ 6 ] . By the end of the 1990s, Crespo et al. 1999 compiled a survey of wake modeling methods for wind turbine design. According to Berg et al. 2014 , individual wind turbines in the 2000s nearly approached the theoretical Betz limit [ 9 ] , but there had not been enough research on wind farm optimization to reduce interference between turbines. Various studies reported that unoptimized turbine layout suffered 10-20% of power losses [ 10 , 11 ] , 10-40% [ 8 ] , or as much as 60-80% in extreme cases [ 12 ] . Turbine spacing is not only a fundamental problem but also a multifaceted challenge in offshore wind farm layout optimization. There are natural and practical constraints on wind farm layout, including seabed morphology, national boundaries, and commercial fishing zones, which entail turbine wake interactions with lower velocity and higher turbulence from upstream to downstream [ 13 , 14 , 15 ] . Excessive lateral wake displacements could cause severe power penalties on top of reduced available energy for downstream turbines [ 16 ] . However, optimal turbine spacing in an array might significantly raise power production [ 11 , 17 , 18 , 19 ] . Until the end of the 2000s , there was a discrepancy between contemporary wind farm design and the development of wind farm optimization theory and measurements, whereby wind farms in practice often had smaller spacing between turbines than what Meyers and Meneveau 2012 found to be optimal. Albeit limited at the time, the growing availability and integration of computational modeling catalyzed layout optimization at the level of multiple turbines [ 21 , 22 , 23 , 4 , 24 , 20 ] . Other studies on layout optimization by computer modeling investigated the intersection between layout optimization and near-wake aerodynamic behavior [ 25 ] , modification of wind energy software tools to test standards [ 21 , 23 ] , and even machine learning for more efficient optimizations [ 4 , 24 ] . Stevens 2016 extended the analysis of Meyers and Meneveau 2012 to underscore the significance of wind farm length in optimal turbine spacing. For the iterative design process of wind farm layouts in computational models, computational efficiency and accurate rendering of design parameters and control features are required [ 27 ] . Hence, the determination of the ideal balance between wind farm layout and turbine wake interaction has remained the primary optimization problem. Analytical turbine wake models are widely used to estimate wake velocity deficits, with upsides such as low computational cost and emphasis on physics through conservation of fundamental flow properties [ 19 , 28 , 13 , 29 , 30 , 31 , 15 ] . Until the mid-2000s, computational fluid dynamics (CFD) was not economically viable due to the high computational cost and limited development [ 32 ] . By as late as 2010, Jimenez et al. 2009 suggested that CFD techniques such as large-eddy simulation (LES) could supplement simple analytical predictions with higher resolution of turbine wakes, which is essential for wake control and wind farm optimization [ 34 , 13 , 35 ] . Such shortcomings provided an opportunity for the development of CFD simulation approaches to achieve high-fidelity wake modeling that is not attainable by analytical methods alone [ 36 , 37 , 38 , 13 ] . CFD models elucidate turbine wake dynamics including wake meandering, atmospheric stability, influence from topology, and effects from the tower and nacelle [ 39 , 30 , 22 ] . Moreover, the actuator methods in CFD lower computational demand by simplifying the turbine geometry and the required resolution of the flow field [ 40 ] . These methods include the actuator disk (AD), actuator line (AL), and actuator surface (AS). All these methods respectively offer increasing fidelity for more computational demand and detail [ 41 , 42 , 43 , 44 , 40 ] . Nonetheless, there has been an inherent obstacle due to the greater computational cost of high-fidelity approaches such as LES. Recently, the integration of analytical and CFD models has emerged as a central approach