Major research outcome
- Dr. Myeonghwan Ahn, Dr. Jun Park and Professor Sang Lee developed a three-dimensional AI-based surrogate model for long-term prediction of turbulent wake dynamics in wind farms
- 관리자 |
- 2026-05-16 20:22:33|
- 87
- 2026-05-16 20:22:33|
Wake interactions in large wind farms reduce power output and increase structural fatigue loads, making accurate wake prediction essential for wind farm design and control. However, large-eddy simulation (LES) is computationally expensive for repeated evaluations, while existing low-fidelity models have limited capability to capture three-dimensional unsteady wake behavior and complex multi-turbine interactions. Under the supervision of Professor Sang Lee in the Department of Aerospace Engineering at KAIST, Dr. Myeonghwan Ahn and Dr. Jun Park proposed a deep learning framework for long-term prediction of three-dimensional turbulent wake flows in wind farms by combining a Swin Transformer backbone with a U-Net neural network architecture. The proposed model was trained using LES data that include various turbine tilt configurations, enabling the framework to effectively capture three-dimensional wake dynamics.
The model accurately predicts three-dimensional velocity fields over long time horizons, successfully reproducing large-scale wake structures and turbine-induced flow variations. Furthermore, detailed physics-based analyses demonstrated that the framework accurately captures velocity deficit profiles, wake centerline trajectories, and rotor inflow velocities, while also showing strong generalization capability for unseen tilt configurations. The proposed framework serves as an efficient surrogate model for LES, reproducing key wake dynamics and multi-turbine interactions at substantially lower computational cost.
This work was published online in May 2026 in Physics of Fluids, a Q1 international journal ranked within the top 6% in the field of fluid mechanics.
The model accurately predicts three-dimensional velocity fields over long time horizons, successfully reproducing large-scale wake structures and turbine-induced flow variations. Furthermore, detailed physics-based analyses demonstrated that the framework accurately captures velocity deficit profiles, wake centerline trajectories, and rotor inflow velocities, while also showing strong generalization capability for unseen tilt configurations. The proposed framework serves as an efficient surrogate model for LES, reproducing key wake dynamics and multi-turbine interactions at substantially lower computational cost.
This work was published online in May 2026 in Physics of Fluids, a Q1 international journal ranked within the top 6% in the field of fluid mechanics.
