Source-linked AI summary
A Fast and Effective Local Search Algorithm for Optimizing the Placement of Wind Turbines
Markus Wagner, Jareth Day, Frank Neumann
TL;DR
Wind-farm layout optimization must account for wake effects, whose evaluation can take Ω(n^2) time per layout. The paper presents a problem-specific local search algorithm that achieves higher-quality results with effective speedups of up to 270.
Problem
Wind-farm layout optimization must account for turbine wakes because they greatly affect energy output, while Park wake-effect computation takes Ω(n^2) time per layout.
Source: §1 · p. 2Method
The paper presents a problem-specific local search algorithm that reduces layout-evaluation complexity for the Park wake model and adapts each turbine’s relocation parameters independently.
Source: §3.2 · p. 9, §6 · p. 15Results
Effective speed-up factors of up to 270 were observed, with higher-quality results achieved in minutes or hours instead of days or weeks.
Source: §6 · p. 15Takeaways & Limitations
The optimization algorithm can also be easily applied to other wake models, including the deep array wake model.
Source: §6 · p. 15Layouts are constrained to the farm area and to a minimum turbine separation set using rotor radius and a proximity factor; the study uses M = 8.
Source: §2.2 · p. 5
Abstract
from arXiv · showhide
The placement of wind turbines on a given area of land such that the wind farm produces a maximum amount of energy is a challenging optimization problem. In this article, we tackle this problem, taking into account wake effects that are produced by the different turbines on the wind farm. We significantly improve upon existing results for the minimization of wake effects by developing a new problem-specific local search algorithm. One key step in the speed-up of our algorithm is the reduction in computation time needed to assess a given wind farm layout compared to previous approaches. Our new method allows the optimization of large real-world scenarios within a single night on a standard computer, whereas weeks on specialized computing servers were required for previous approaches.
1 Introduction
Wind-farm layout optimization must account for wake effects, whose computation scales quadratically with turbine count for each evaluated layout. The paper introduces a problem-specific local search algorithm intended to speed evaluation, use layout characteristics, and handle infeasible regions.
1 Introduction
- Wake computation takes Ω(n^2) per layout, making iterative optimization costly as turbine counts increase.Wake effects arise when nonlinear turbulence from one turbine affects the inflow to turbines behind it.Source: §1 · pp. 1–2
1.1 Related Work
- Earlier approaches include coordinate-based evolution strategies and CMA-ES, while many methods simplify wake or wind scenarios and OpenWind lacks adaptive perturbations [20].CMA-ES can optimize layouts of up to 1,000 turbines but may require weeks on computing servers.Source: §1.1 · pp. 2–3
1.2 Our Contribution
- TDA evaluates a changed layout in Θ(kn), where k is the number of turbines moved, rather than the prior Ω(n^2) wake-computation cost.Its contributions also include faster layout evaluation, problem-aware layout generation, and handling infeasible regions.Source: §1.2 · p. 3
- The algorithm is designed to generate good layouts for hundreds of turbines quickly on standard computers, without specialized hardware.The paper cites the Horse Hollow Wind Energy Center, which has more than 300 turbines, as an example of large-farm scale.Source: §1.2 · p. 3
- The new algorithm outperforms CMA-ES and AWS OpenWind in solution quality and running time, according to the paper’s experiments.The introduction identifies the Horse Hollow Wind Energy Center as a large-farm example, with more than 300 turbines.Source: §1.2 · p. 4
2 Wind Farm Layout Optimization
The layout problem seeks turbine coordinates that maximize farm energy while satisfying spatial constraints. Energy estimates account for wake effects under a Park wake model.
2.1 Energy Output
- Farm energy depends on turbine coordinates because wake effects alter the wind resource available to each turbine.The Park model adjusts turbine-specific Weibull scale parameters using velocity deficits from other turbines.Source: §2 · p. 4, §2.1 · p. 5
- The expected-energy calculation combines wind speed, turbine power curves, wind-direction probabilities, and a direction-dependent Weibull distribution estimated from wind-resource data.The distribution accounts for wake effects through the turbine coordinates; its shape parameter is not changed by wakes.Source: §2.1 · p. 5
2.2 Constraints and Assumptions
- Turbines must remain within the farm bounds and at least M times the rotor radius apart; the paper uses M = 8.The setup assumes identical piecewise-linear power curves and a uniform wind resource across the farm, though these assumptions can be revised.Source: §2.2 · pp. 5–6
3 Turbine Distribution Algorithm
The Turbine Distribution Algorithm (TDA) combines a problem-specific local search with faster layout evaluation. It moves turbines to improve energy while respecting constraints, and adapts displacement behavior during the search.
3.1 Computational Speedup
- After a one-time influence-matrix setup, subsequent evaluation takes Θ(kn) when k of n turbine locations change, rather than quadratic computation for the influence matrix.The experiments use k = 1; the matrix stores mutual turbine influences across wind directions.Source: §3.1 · p. 6
3.2 The Algorithm
- TDA initializes turbines in a maximally spaced grid, then reduces an illegal displacement until the turbine remains within bounds and outside prohibited proximity zones.For nn = 2, the illustrated displacement is influenced by the two nearest neighbors.Source: §3.2 · p. 7, Figure 1 · p. 9
- TDA repeatedly modifies one randomly selected turbine in a copy of the best-so-far layout and compares the candidate layout's energy.The algorithm is a purpose-built local search that displaces a single turbine while enforcing constraints.Source: §3.2 · p. 7
- TDA derives a displacement direction from the turbine's nearest neighbors, then samples direction and distance; the vector may be reversed to allow closer placements.When a move improves quality, that turbine's distance standard deviation increases for exploration; after a lower-quality move, it decreases for exploitation.Source: §3.2 · pp. 7, 9
4 Experimental Investigations
The experiments evaluate TDA on realistic wind scenarios, examine its settings, and compare its solution quality and runtime with CMA-ES* and AWS OpenWind. TDA often matches or exceeds CMA-ES* with fewer evaluations, while OpenWind’s optimizer improves layouts less.
4.1 Algorithm Settings and Scenario Information
- The test scenario used wind from a broad 105° sector with non-zero probability in every direction, and the algorithm adjusted parameters according to recent displacement success.The experiments used the wind resources of Scenario 2 in.Source: §4.1 · pp. 9–10
4.2 Impact of the Nearest Turbines
- For 100-turbine scenarios, nn = 2 yielded 1.1% more energy on average than nn = 8; the reported average differences were 0.6% for 40 turbines and 0.1% for 400.The authors suggest opposing neighbors can cancel displacement vectors when four or eight neighbors are considered, producing less exploratory directions.Source: §4.2 · pp. 10–11, Figure 2 · p. 10
4.3 Experimental Results
- The comparison covered varying turbine counts and farm sizes, using 10,000 and 200,000 evaluations for TDA and up to 200,000 for CMA-ES* and OpenWind.Each scenario began from the same maximally spaced grid and was repeated 30 times; CMA-ES* was tuned by increasing its initial standard deviation after poor results on 70- and 80-turbine scenarios.Source: §4.3 · p. 11, Figure 3 · p. 13
- Up to 1.4% more energy than CMA-ES* was achieved by TDA with 20 times the evaluations, and TDA was significantly faster.For 40 turbines, TDA reached a better gain in three minutes versus CMA-ES*’s two hours; for 400 turbines, TDA took less than an hour versus CMA-ES*’s more than 200 hours.Source: §4.3 · pp. 11, 13, Table 1 · p. 12, Figure 3 · p. 13
- OpenWind’s optimizer averaged only 2% improvement over initial layouts, while TDA adapts relocation parameters per turbine and shifts between exploration and exploitation.The authors attribute OpenWind’s results most likely to a lack of self-adaptation that leads it to local optima, typically within the first 1,000 assessments.Source: §4.3 · p. 13
- As turbine density increased in the 10–100 turbine scenarios, wake-loss energy increased, while the benefit of adding turbines relative to the next smaller scenario decreased.For 200–1,000 turbines, farms were sized to contain roughly 2n turbines, and the reported wake-loss percentage was relatively constant.Source: §4.3 · p. 13
5 A Real-World Problem: Dealing With Infeasible Areas
TDA handles infeasible areas by representing them as shapes and correcting displacements that would place a turbine inside them. A Woolnorth-based example demonstrates the modeling capability, but is explicitly a proof of concept.
5 A Real-World Problem: Dealing With Infeasible Areas
- When a displacement would move a turbine into an infeasible region, TDA reduces its distance until the move is legal.The method models infeasible regions as sets of shapes and handles them like proximity and farm-area constraints.Source: §5 · p. 14
- The Woolnorth example modeled 37 turbines at the north site, but did not use the site’s actual wind characteristics or internal map and terrain information.Only coastal details were represented, and the experiment used the wind resource from Scenario 2 in.Source: §5 · pp. 14–15
- In a modeled scenario using coastal details but not Woolnorth wind characteristics or actual-site map or terrain, turbines quickly formed staggered north/south columns by evaluation 20,000 under predominantly western winds (120°–225°).Source: §5 · p. 15
6 Conclusion
The conclusion attributes TDA’s performance to problem-specific layout optimization and reduced Park-model evaluation complexity. The authors also state that the algorithm can be applied to other wake models.
6 Conclusion
- For the Park wake model, layout evaluation drops from Ω(n²) to Θ(kn), where k is the number of turbines moved from the previous layout.The authors report higher-quality results than existing approaches and assessment speedups of up to 270, allowing optimization in minutes or hours rather than days or weeks.Source: §6 · p. 15
- Although the experiments used the Park wake model, the authors say TDA can also be applied to models such as the deep array wake model.Source: §6 · p. 15