⚡ electrical engineering

Intelligent Optimization of Wind Farm Layout Under Wind Turbine Wake Effect Based on Deep Learning and Fluid Dynamics

This paper proposes an end-to-end collaborative framework integrating physics-informed neural networks, graph attention networks, and deep reinforcement learning to simultaneously enhance wind farm wake prediction accuracy and optimize turbine layouts, achieving significant reductions in wake losses and inference time while maintaining physical consistency and scalability.

Original authors: Shujie Liu, Shengqiao Chen, Bo Xiang

Published 2026-09-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shujie Liu, Shengqiao Chen, Bo Xiang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Wind energy has become a cornerstone of the global power system, with vast fields of turbines stretching across coastlines and plains to capture the wind. However, a hidden problem limits how much electricity these farms can generate. When a wind turbine spins, it pulls energy from the air, leaving a turbulent, slower-moving trail of air behind it, known as a wake. If a second turbine sits in this wake, it receives less wind energy and produces less power, while also suffering from increased stress that can shorten its lifespan. In large wind farms, these wakes overlap and interact, causing the entire farm to lose between ten and twenty percent of its potential annual energy output. For decades, engineers have tried to solve this by rearranging the turbines, but the physics of how these air trails interact is incredibly complex. Traditional methods either rely on simplified guesses that miss the details or use supercomputer simulations so slow and expensive that they cannot be used to test thousands of different layouts.

A team of researchers has developed a new way to solve this puzzle by teaching a computer to understand the physics of the wind while learning from vast amounts of data. Instead of treating the wind farm as a collection of separate machines, they built a system that sees the farm as a connected network, where every turbine influences its neighbors. The researchers created a three-part intelligent system. First, they trained a neural network to predict how the wind moves through the farm, but they forced this network to obey the fundamental laws of fluid motion, ensuring its predictions remained physically realistic even in areas where data was scarce. Second, they used a graph-based approach to map the specific relationships between turbines, allowing the system to learn exactly how much a turbine upstream affects one downstream. Finally, they employed a learning agent that acts like a strategic planner, testing millions of layout variations to find the arrangement that generates the most power while minimizing the interference between turbines.

When the researchers tested this system on a benchmark wind farm equivalent to a real-world site with eighty turbines, the results were significant. The new method predicted the wind flow with far greater accuracy than previous models, reducing the error in its predictions by nearly two-thirds compared to standard engineering formulas. More importantly, when the system designed a new layout for the farm, it increased the total annual energy production by 5.7 percent. This improvement came from shifting the turbines slightly to break up the overlapping wakes, which reduced the energy loss caused by the wind shadows from 11.4 percent down to 6.5 percent. While the system requires a training period of roughly 28 to 96 hours depending on the farm size, once trained, it performs calculations in a fraction of a second, a speed that makes it possible to test and refine designs in real-time, something that was previously impossible with older, slower simulation tools.

The study also showed that this approach is robust, meaning it works well even when the wind conditions are difficult or changeable. Whether the wind blew from a single direction, from two opposing directions, or from all around, the system consistently improved energy production by between 4.4 and 5.9 percent. It even handled scenarios with high turbulence and strong vertical wind shear, where other methods often fail. The researchers found that the system could scale up to handle wind farms with two hundred turbines without losing its speed or accuracy, maintaining a response time of just over one hundred milliseconds. This suggests that the method could be applied to massive offshore or onshore projects, offering a practical tool to squeeze more clean energy out of existing wind resources.

By combining deep learning with the strict rules of fluid dynamics, the researchers have created a tool that bridges the gap between theoretical physics and practical engineering. The system does not just guess at the best layout; it learns the underlying patterns of how wind interacts with a group of machines and uses that knowledge to make decisions. This approach moves beyond the limitations of older techniques, which often had to choose between being fast but inaccurate or being accurate but too slow to be useful. The new framework offers a path forward for designing wind farms that are not only more efficient but also more resilient to the complex and changing nature of the wind, potentially saving millions of dollars in lost energy over the life of a project.

Technical Summary: Intelligent Optimization of Wind Farm Layout Under Wind Turbine Wake Effect Based on Deep Learning and Fluid Dynamics

Problem Statement
Large-scale wind farm expansion is significantly hindered by wake effects, where upstream turbines reduce inflow velocity and increase turbulence for downstream units, leading to 10%–20% losses in annual energy production (AEP). Existing optimization approaches face a critical trade-off between accuracy, efficiency, and robustness. Traditional analytical models (e.g., Jensen, Bastankhah-Gauss) rely on simplified assumptions that fail to capture complex turbulence superposition and wake deflection in dense layouts. Conversely, high-fidelity Computational Fluid Dynamics (CFD) simulations, such as Large Eddy Simulation (LES), offer accuracy but are computationally prohibitive (hours per condition) for iterative optimization. Furthermore, recent data-driven deep learning methods often lack physical consistency in data-sparse regions and treat flow field prediction and layout optimization as decoupled tasks, limiting their ability to generalize across high-dimensional, variable wind conditions.

Methodology
The authors propose an end-to-end collaborative framework integrating deep learning with fluid dynamics, structured around three tightly coupled modules:

Physics-Informed Neural Network (PINN) Wake Field Proxy:
- This module serves as a surrogate for high-fidelity flow field simulations.
- It embeds the residuals of the Reynolds-Averaged Navier-Stokes (RANS) momentum and continuity equations as soft constraints within the loss function.
- This ensures the network adheres to conservation laws (mass and momentum) even in data-sparse regions, preventing non-physical predictions during extrapolation.
- The architecture utilizes an 8-layer fully connected backbone with SiLU activation and multi-scale residual blocks to capture wake gradients.
Graph Attention Network (GAT) for Multi-Turbine Coupling:
- Turbines are modeled as nodes in a directed graph, with upstream-downstream relationships defined by wind direction projection as directed edges.
- Edge features encode geometric distance, azimuth angles, and thrust coefficients.
- The GAT employs multi-head attention to learn adaptive coupling weights between turbines, explicitly capturing the spatial structure of wake interactions rather than relying on fixed adjacency matrices or grid-based convolutions.
- The output is a wind farm-level embedding representing the global state of turbine interactions.
Deep Reinforcement Learning (DRL) for Layout Optimization:
- The layout selection problem is formalized as a Markov Decision Process (MDP).
- The state space consists of the global graph embedding from the GAT and current turbine positions.
- Actions involve discrete positional adjustments (North, South, East, West) for individual turbines.
- The Proximal Policy Optimization (PPO) algorithm is trained with a multi-objective reward function balancing AEP improvement, wake intensity uniformity, and engineering constraints (spacing, boundaries).
- The system undergoes a two-stage training process: pre-training the PINN-GAT agent on LES data, followed by fixing these parameters to train the PPO policy.

Key Contributions

Physics-Constrained Surrogate Modeling: Development of a lightweight wake field proxy that reduces inference time to the sub-second level while maintaining LES-level accuracy by enforcing physical laws via soft constraints.
Graph-Based Coupling Representation: Introduction of a graph attention mechanism that treats turbine spatial coupling as a learnable component of the state space, overcoming the limitations of regular grid representations in capturing complex multi-machine interactions.
End-to-End Strategy Optimization: A unified framework where layout generation is directly driven by a policy network rather than external heuristic searches, enabling the optimization algorithm to learn global strategies guided by a physically consistent environment simulator.

Experimental Results
The framework was evaluated on an 80-turbine benchmark (equivalent to Horns Rev I) using a dataset of 5,000 LES-generated samples across 15 wind speeds and 12 wind directions.

Prediction Accuracy: The PINN-GAT model achieved an $R^2$ of 0.97 against LES ground truth. It reduced the Root Mean Square Error (RMSE) by 67% compared to the Bastankhah-Gauss model and by 49% compared to a CNN-only model.
Efficiency: Single-condition inference time was compressed to 41 ms, representing an acceleration of approximately $3.3 \times 10^5$ times compared to LES (3.8 hours).
Optimization Performance: On the 80-turbine benchmark, the method increased AEP by 5.7% and reduced wake loss from 11.4% to 6.5%. This outperformed baseline methods (GA+Jensen, PSO+Gauss, BO+CNN) by margins ranging from 1.6% to 3.1% in AEP improvement.
Robustness and Scalability: The method demonstrated consistent AEP improvements (4.4%–5.9%) across five distinct wind condition categories, including high turbulence and vertical shear. Scalability tests showed that training duration and inference time scale approximately linearly with turbine count, maintaining 108 ms inference time for 200 turbines.

Significance and Claims
The paper claims to deliver a reusable methodology chain that successfully balances accuracy, efficiency, and robustness for intelligent wind farm layout optimization. By integrating physical constraints into deep learning and utilizing graph structures for spatial coupling, the approach addresses the "accuracy-speed" trade-off that has long hindered the industry. The authors assert that this framework enables a shift from "post-examination" to "pre-design" optimization, allowing for real-time evaluation of candidate layouts and the generation of Pareto-optimal solutions that balance power generation with wake uniformity. The study suggests that a 5.7% AEP gain on a 400 MW farm could yield nearly 1 billion yuan in cumulative economic value over a 25-year lifecycle.

The authors modestly acknowledge limitations, noting that the training data relies on LES rather than real-world SCADA or LiDAR data, and that the DRL training cost, while acceptable, remains higher than simple heuristics. Future work is proposed to include real-site data fine-tuning and meta-learning strategies to reduce training costs and enhance "plug-and-play" capabilities.

Technical Summary: Intelligent Optimization of Wind Farm Layout Under Wind Turbine Wake Effect Based on Deep Learning and Fluid Dynamics

More like this