Transcript Towards the Evolution of Novel Vertical Axis Turbines

Richard Preen & Larry Bull
UWE, Bristol
Introduction
 Evolutionary computing has been applied widely.
 Over 70 examples of “human competitive”
performance have been noted [Koza, 2010].
 Most of this work has included simulation or models.
 When simulations are costly, surrogate models can be
used.
 Data mining techniques are used to create
approximations of the function space from sample
points gathered from the simulator/model.
Embodied Evolution 1
 Some hard problems are difficult to model or simulate
in a useful way.
 Small amount of work using simulated evolutionary
design directly on a task:
 Jet nozzle [Rechenberg, 1971]
 Mobile robotics [Nolfi, 1992]
 Electronic circuits [Thompson, 1998]
 Unconventional computing [Harding & Miller, 2004]
 Chemical systems [Theis et al., 2007]
Embodied Evolution 2
 Explore use of surrogate models in conjunction with




direct solution evaluation only for complex tasks.
No best-guess simulator or model used.
Combine with emerging rapid-fabrication (3D
printing) technology.
Potential for truly unexpected results in a wide range
of domains.
Many issues, of course: time, noise in evaluations,
representations, kinds of surrogates, etc.
An Example: Wind Turbines
Wind Energy
 “In theory, small-scale wind energy has the potential to
generate 41.3 TWh of electricity and save 17.8 MtCO2 in the
UK annually” [Carbon Trust, 2008].
 Wind flow is rarely constant and consistent, rather it is
usually veering and turbulent, and the influences of nearby
obstacles can significantly alter wind flow patterns.
 Vertical axis wind turbines (VAWT) represent a very
effective approach to harnessing wind power in many
situations – especially urban areas.
 In comparison to the more common horizontal axis wind
turbine (HAWT), VAWT can also be easier to manufacture,
may scale more easily, and are typically inherently lightweight with little or no noise pollution.
A Simple Representation
 Confined design space to a four-blade Savonius VAWT with
alterations in blade shape (profile and twist) possible.
 Genome [5,8,2,4] defines offsets from central spindle.
Embodied Evolution of VAWT
 Turbines of 30mm3 volume
 Population size 20
 Tournament selection
 Tip speed used as fitness measure
 Direct evaluations for first three generations (60 fabs)
 MLP surrogate model of fitness
 MLP trained for 1000 epochs per generation
 Best and random individual fabricated per generation
Single VAWT
Single VAWT: Z-axis design
Beyond Modelling
 Evolutionary computing has previously been used to
design both HAWT and VAWT via CFD models.
 Essentially impossible for turbine arrays where
interactions are considered.
 Dabiri et al. have recently highlighted how the spacing
constraints of HAWT arrays often do not apply for
VAWT, and even that performance can be increased by
exploitation of inter-turbine flow effects.
 View wind farm as an energy-capture ecosystem and
coevolve heterogeneous, interacting VAWT.
Embodied Coevolution
 Two “species” of turbine.
 Two populations and surrogate models (L, R).
 Each evaluated with best individual from other




population.
Evolve alternately.
Seed with 10 best from single VAWT experiments.
Allow counter-rotation.
All other features the same as before.
Paired VAWT
Example
Results
 Speciation seen – L and R physically different.
 Homogeneous pairing of either species not as effective
as the heterogeneous case.
 Counter-rotation not seen in fittest solutions here.
 Approach potentially unaffected by array size increase.
 CFD modelling impossible.