Session: 01-05-03: Data-Driven Design and Optimization of Aerospace Structures Using AI/ML

Paper Number: 190186

190186 - Fast Wind-Farm Aep Evaluation and Layout Optimization via Low-Rank Wake Power Matrices 

Annual energy production (AEP) evaluation is a major computational bottleneck in wind-farm design and layout optimization because it requires repeated wake-model simulations over a wind-direction-wind-speed (WD-WS) grid. Coarse wind-rose discretization can reduce runtime but may introduce bias and numerical artifacts in wake-dominated regimes where power losses vary sharply with wind direction. We introduce a physics-informed low-rank reconstruction framework that accelerates AEP evaluation by querying only a small subset of inflow bins and reconstructing the full WD-WS wake-induced efficiency matrix.

We provide empirical evidence that wake-induced loss/efficiency matrices exhibit strong effective low-rank structure across diverse layout topologies. Singular value analysis shows that the leading mode captures the dominant variance of wake losses, while rank-1 reconstruction errors stratify by topology: isotropic (cluster-like) farms are nearly rank-1, whereas anisotropic string-like farms require additional modes due to sharper directional signatures and deep-array operating-point disparity. The dominant directional mode behaves as a geometric "topological fingerprint" that is weakly dependent on wind speed under sub-rated conditions, while the leading speed loading acts as an aerodynamic modulation consistent with thrust behavior, with higher-order activity concentrated near control transition regions (cut-in and rated).

Building on these observations, we propose an Energy-Weighted Adaptive Cross Approximation (EW-ACA) kernel for fast AEP computation. Pivot inflow states are selected using an energy proxy proportional to wind-bin probability weighted by the no-wake power curve. EW-ACA constructs a rank-1 efficiency skeleton from a single cross probe (one full-direction sweep at an energy-dominant wind speed and one full-speed sweep at an energy-dominant direction), reducing wake-model calls from O(N_WD N_WS) to approximately O(N_WD + N_WS). A lightweight adaptive refinement then injects a few additional informative direction rows to suppress non-separable residuals while preserving the low-rank backbone.

Across benchmark wind climates and a broad layout ensemble, the framework yields a robust accuracy-speed trade-off. Relative to full-grid references, the median AEP relative error decreases from ~0.178% (rank-1) to ~0.028% (adaptive), and the reconstructed efficiency-field error improves from median RMSE ~0.043 to ~0.005. Substantial acceleration is retained: relative to a sparse baseline that evaluates only WD-WS bins with nonzero wind probability, the methods achieve median speedups of ~9.8x (rank-1) and ~5.0x (adaptive). When integrated into a WFLO pipeline under a fixed wall-clock budget (64 CPU cores x 12 h, 10 runs per method) and assessed via full-fidelity re-evaluation of final designs, the accelerated evaluators enable broader search and yield layouts that remain consistent under re-evaluation. The optimization objective is normalized AEP, defined as farm AEP divided by the corresponding no-wake (wake-free) AEP under the same wind climate, and is reported here in percent. In one case, the rank-1 evaluator achieves 89.974% +/- 0.230% versus 89.207% +/- 0.298% for the full evaluator after re-evaluation, corresponding to an absolute difference of 0.767 percentage points. Rather than suggesting systematic gains over full evaluation, the results indicate that the low-rank accelerated evaluator can be used within WFLO without destabilizing the optimization, yielding full-fidelity re-evaluable outcomes while reducing evaluation cost.

Presenting Author: Zhikun Dong The Univiersity of Hong Kong

Presenting Author Biography: Zhikun Dong is currently a Ph.D. candidate in Civil Engineering at The University of Hong Kong, supported by the Hong Kong PhD Fellowship Scheme (HKPFS) and the HKU Presidential PhD Scholarship. He holds a Master’s and Bachelor’s degree in Civil Engineering from Shanghai Jiao Tong University, where he was recognized as a Shanghai Outstanding Graduate.His research focuses on wind resource assessment, wake dynamics, and wind farm layout optimization, with a particular emphasis on applying artificial intelligence and numerical simulation to offshore wind energy systems.