Robust wind farm layout optimization using pseudo-gradients
Erik Quaeghebeur
1 Wind energy systems
A wind energy system transforms wind into electrical power.
1.1 Wind turbines
© Hans Hillewaert / CC BY-SA 4.0
Wind turbines (picture left) are the elementary wind energy sys- tems. Important characteristics are its rated power, rotor diame- ter, and hub height.
0 v
cut-inv
cut-outv
p
rated power
c
t‘maximum’ thrust A high-level model
consists of the power curve and the thrust curve, which map wind speed at hub height
to power and force exerted on the wind (plot above right).
1.2 Wind farms
© Thomas Nugent / CC BY-SA 2.0
Wind farms are collections of wind turbines constrained to
a specific site (picture left).
The placement of turbines within a farm is its layout (drawing right).
The layout influences the farm cost via the cabling and substructure cost, due to cable layout and depth & soil variations.
1.3 Wake losses
Wakes are regions of complexly perturbed wind behind turbine rotors (picture right).
0°
45°
90°
135°
180°
225°
270°
315°
0.0 2.5
5.0 7.5
10.0 12.5
15.0 17.5
20.0
Computationally simple engi- neering wake models are used when calculating farm power
output (simulation below right).
In a farm, wakes may reduce the wind speed at downstream turbines, causing lower power production: wake losses. Wake wind speed deficits for a given layout depend on the wind direc- tion (plot above, corresponds to layout shown in Section 1.2).
© Vattenfall
© Harms et al. (2016 student project)
2 Wind resources
A wind resource is the wind available at a wind farm site.
2.1 Wind direction & speed distributions
0°
45°
90°
135°
180°
225°
270°
315°
5101520253035
0.0000 0.0025 0.0050 0.0075 0.0100 0.0125 0.0150 0.0175
The minimal wind re- source description re- quired is a joint wind direction & speed dis- tribution (plot left);
there is a dependency between both
components.
0°
45°
90°
135°
180°
225°
270°
315°
0.020.040.060.080.100.120.14
This joint is decomposed into the wind rose, the wind direction marginal (plot above right), and per- direction wind speed conditionals (plot right), for which Weibull distributions are often used.
0 v
cut-inv
cut-outv
225
°315
°2.2 Annual energy production of a wind farm
An essential quantity in the design of a wind farm is its annual energy production (AEP): the electrical energy produced by a farm for a given wind resource.
Equivalent is the capacity factor, the ratio between the ex- pected average power production and the farm’s rated power.
Also of interest is the power rose, the distribution over wind directions of relative wakeless power production (plot right).
0°
45°
90°
135°
180°
225°
270°
315°
0.020.040.060.080.100.120.14
2.3 Inter-year wind resource variation
0°
45°
90°
135°
180°
225°
270°
315°
0.020.040.060.080.100.120.140.16 We consider 35 yearly wind re- sources for a North Sea site from the Dutch meteorological institute’s
‘KNW atlas’ (plot left: orange lower, blue average, and red upper wind
roses for this set of distributions; plot right: corresponding power roses).
Note the substantial variation.
0°
45°
90°
135°
180°
225°
270°
315°
0.0250.0500.0750.1000.1250.1500.1750.200
4 Inter-year variation robustness
A wind farm’s layout is usually optimized for one wind resource, the estimated average one over the farm lifetime. However, inter-year production stability is important for the finan- cial attractiveness of a farm design. Making a farm robust against inter-year wind re- source variation is therefore of practical interest.
4.1 Goals
• Quantify inter-year wind resource variation (done).
• Quantify inter-year AEP variation (done).
• Determine existence of robust farm layouts (partly done).
• Develop robust layout optimiza- tion algorithm (not yet done).
4.2 Setup
• Realistic test site.
• Realistic & extensive set of yearly wind resources.
• Create optimized layout for
– a degenerate wind resource (‘
225
°’), – the uniform wind resource,– each wind resource in the set, – their average,
– their lower & upper envelopes.
4.3 Results
2.8 3.0 3.2 3.4 3.6 3.8 4.0
Wake loss percentage 19791980
19811982 19831984 19851986 19871988 19891990 19911992 19931994 19951996 19971998 19992000 20012002 20032004 20052006 20072008 20092010 20112012 all_years2013
Wind resource
year-optimized upper
all years lower uniform only 225°
4.4 Conclusions
• Inter-year variation is substantial.
• Observed inter-year variation is larger than inter-layout differences.
• The set of layouts with undominated production profiles is relatively small.
• No real trade-off achieved yet between robustness and optimality.
4.5 Recommendations
• Create a more diverse set of layouts:
– by varying the optimizer parameters,
– by using different optimization algorithms.
• Try out ideas for robust optimization:
– by each iteration using the maximin solution over wind resources,
– by following your suggestion.
3 Wind farm layout optimization
3.1 Objectives
• AEP: Maximize for expected power production only (used in our study).
• LCoE: Minimize levelized cost of energy,
the ratio between farm cost and power production (more realistic).
3.2 Constraints
Turbines in a farm must satisfy a distance constraint (drawing right, red circles) and site constraints (drawing right, red & blue lines).
3.3 Typical layout optimization algorithms
type gradient-based heuristic (usually random search-based) examples steepest ascent evolutionary, genetic, particle swarm
pros high-quality solutions flexible (generic)
cons computationally expensive, computationally expensive,
can get stuck in local optima, does not use domain knowledge, problem-specific preparation low-quality solutions
Computational cost is crucial in robustness studies, so we developed a fast heuristic approach that uses domain knowledge and produces medium-quality solutions.
3.4 Pseudo-gradient-based optimization
0°
45°
90°
135°
180°
225°
270°
315°
0.2 0.4
0.6 0.8
1.0 For one wind direction (plot left), the power deficit of a down-
stream turbine due to an upstream one determines a vector. Average
over all upstream turbines (plot right).
Variant: vectors pushing upstream turbines ‘back’ (plot far right).
0°
45°
90°
135°
180°
225°
270°
315°
0.020.040.060.080.100.120.14
Taking the expectation over all direc- tions (plot left) gives ‘pseu-
do-gradients’ usable in a local gradient ascent-type algorithm.
Applicable to all variants (plot right:
‘push down’; plot far right: ‘push back’).
We have created a layout optimization algorithm that each iteration:
0 10 20 30 40 50
0.0375 0.0400 0.0425 0.0450 0.0475 0.0500 0.0525
0.0550 • uses an adaptive step size,
• considers pseudo-gradients for each of the variants,
• greedily moves turbines
according to the best one, and
• corrects constraint violations be- tween steps by iteratively moving turbines to satisfying positions.
We obtain good convergence (plot above left, relative wake loss) and medium-quality layouts (drawing above right, turbine trajectories).