Session: 01-05-01: Foundations of AI and Machine Learning for Aerospace Structural Analysis

Paper Number: 190367

190367 - Phoenix: Physics-Informed Hierarchical Generative Deep Learning for Long-Horizon Dynamic Wake Modeling 

Accurate yet efficient prediction of unsteady wind-turbine wakes is critical for wind-farm power assessment, layout optimization, and load analysis. Physics-based dynamic wake meandering (DWM) models can represent key mechanisms (wake deficit, meandering, and rotor-induced turbulence), but remain computationally expensive for large ensembles and optimization loops. Meanwhile, purely data-driven wake surrogates often suffer from limited physical consistency and error accumulation in long sequence forecasting, leading to instability and physically implausible wake evolution.

This work presents PHOENIX (PHysics-infOrmed gEnerative deep learniNg for hIerarchical dynamic wake modeling eXploration), a hierarchical dynamic wake surrogate that frames wake evolution as a conditional image-to-image translation problem. A conditional generative adversarial network (cGAN) with a U-Net generator and PatchGAN discriminator is used as a time-stepping solver, predicting the next-step wake snapshot from (i) the instantaneous ambient wind field and (ii) the previous wake snapshot. To enforce physical plausibility under convection-dominated wake transport, PHOENIX incorporates a physics-informed loss that penalizes violations of (a) upstream-region invariance and (b) side-boundary/interface continuity, reflecting the limited crosswind extent of wake influence and the continuity of the ambient field.

Training and evaluation data are generated using the NREL FAST.Farm DWM implementation for the NREL 5-MW reference turbine. Three ambient wind conditions (hub-height mean wind speeds 8/9/10 m·s⁻¹, turbulence intensity 10%) are simulated over a 5 km × 6 km × 340 m domain. Both freestream and waked inflow cases are included (300 and 1800 simulations, respectively), producing ~1.68 million 2D velocity-field snapshots (streamwise and spanwise components) after preprocessing and resampling to 128 × 128. The trained base model achieves excellent single-step accuracy on the test set (MAPE ≈ 0.46%, R² ≈ 0.999 for both velocity components).

To address long-horizon instability, PHOENIX introduces hierarchical temporal aggregation, training multiple sub-models at lead times of 10 min, 5 min, 1 min, 10 s, and 2 s, and invoking them via a greedy largest-factor representation to reduce sequential rollouts. Over an 800 s forecasting horizon, the hierarchical strategy maintains stable performance (maximum mean MAPE ≈ 0.9%, minimum R² ≈ 0.995) while a non-hierarchical rollout exhibits rapid error growth and potential pattern “crash.” Computational tests show PHOENIX delivers substantial acceleration relative to the physics-based DWM solver (≈ 274× on CPU, up to 4890× using a single RTX 3090 GPU). A 3×3 wind-farm case study further demonstrates engineering relevance, yielding a mean absolute percentage error of ~5.03% in turbine power prediction under waked conditions.

PHOENIX provides a fast, stable, and physically consistent surrogate for dynamic wake fields, enabling scalable wake evaluation for wind-farm design, control-oriented studies, and optimization under realistic unsteady inflow and wake-interaction conditions.

Presenting Author: Xiaowei Deng The University of Hong Kong

Presenting Author Biography: Dr. Xiaowei Deng is an Associate Professor in the Department of Civil Engineering at The University of Hong Kong (HKU). He received his Ph.D. in Aerospace Engineering from the California Institute of Technology (Caltech) in 2012, where he was awarded the William F. Ballhaus Prize for outstanding dissertation. Prior to joining HKU, he served as a Postdoctoral Fellow at the University of California, San Diego.
Dr. Deng’s research interests lie at the intersection of computational fluid dynamics (CFD), fluid-structure interaction, and machine learning, with a specific focus on wind resource assessment, wind turbine wake modeling, and wind farm layout optimization. He has led and contributed to numerous major research projects funded by the RGC and NSFC, aiming to enhance the performance and resilience of offshore wind energy systems in complex environments. His work has resulted in over 90 journal publications and an H-index of 37.