.png){kind=icon}
.png){kind=icon}
Return
Addressing scale discrepancy in wind turbine wake turbulent kinetic energy predictions via quantile transformation and bayesian deep learning
.png){kind=icon}
PRE
AI
Z
M
Y
D
DOI：10.1007/s42241-026-0041-x.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Wind energy is a crucial component of the global transition to green energy. The development of wind resources relies on wind turbines, which are typically clustered in wind farms. The wake generated by upstream turbines significantly impacts downstream turbines, leading to reduced power generation and increased fatigue loads. turbulent kinetic energy (TKE) is a key physical quantity in wind turbine wake prediction, governing wake mixing and recovery processes, thus holding substantial research importance. However, a major challenge in data-driven wake prediction is the significant scale discrepancy in TKE distribution. TKE values are predominantly low across most of the wake field, with only a few critical regions (e.g., behind the blade tips and hub) exhibiting high values. This highly skewed distribution causes conventional data-driven models to often treat these high-value points as outliers, leading to poor learning of critical features and increased model uncertainty. This study proposes a novel data-driven framework combining quantile transformation and Bayesian deep learning to accurately predict the TKE distribution in wind turbine wakes. The nonlinear quantile transformation effectively mitigates the scale discrepancy issue, enabling the neural network to learn features from the highly imbalanced data more effectively. Meanwhile, the Bayesian deep learning approach quantifies the predictive uncertainty, enhancing the model’s reliability. The proposed method is validated against high-fidelity computational fluid dynamics (CFD) data, demonstrating superior performance compared with traditional scaling methods like Min-Max scaler.
Journal
IF:
3.5
Papers:
2.4K
Citations:
4.0K
Organization
Cited Papers