Beruflich Dokumente
Kultur Dokumente
Abstract
The Monte Carlo simulation method for turbomachinery uncertainty analysis often requires performing a huge number
of simulations, the computational cost of which can be greatly alleviated with the help of metamodeling techniques. An
intensive comparative study was performed on the approximation performance of three prospective artificial intelligence
metamodels, that is, artificial neural network, radial basis function, and support vector regression. The genetic algorithm
was used to optimize the predetermined parameters of each metamodel for the sake of a fair comparison. Through test-
ing on 10 nonlinear functions with different problem scales and sample sizes, the genetic algorithm–support vector
regression metamodel was found more accurate and robust than the other two counterparts. Accordingly, the genetic
algorithm–support vector regression metamodel was selected and combined with the Monte Carlo simulation method
for the uncertainty analysis of a wind turbine airfoil under two types of surface roughness uncertainties. The results
show that the genetic algorithm–support vector regression metamodel can capture well the uncertainty propagation
from the surface roughness to the airfoil aerodynamic performance. This work is useful to the application of metamodel-
ing techniques in the robust design optimization of turbomachinery.
Keywords
Support vector regression, artificial neural network, radial basis function, uncertainty analysis, wind turbine airfoil
Creative Commons CC-BY: This article is distributed under the terms of the Creative Commons Attribution 3.0 License
(http://www.creativecommons.org/licenses/by/3.0/) which permits any use, reproduction and distribution of the work without
further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/
open-access-at-sage).
2 Advances in Mechanical Engineering
chaos (PC) techniques, can directly yield the uncertain turbomachinery uncertainty analysis. Therefore, in this
outputs through a one-time solution of certain refor- study, an intensive comparison among these three AI
mulated equations; however, they are severely limited metamodels was particularly conducted, where the per-
to low-dimensional, low-order, and short-time systems formance evaluation strategies were carefully designed
due to the mathematical and/or numerical difficulties and tested through a total of 10 nonlinear functions
arising from the reformulated equations. The latter, with different problem scales and sample sizes.
such as Monte Carlo simulation (MCS) method or its Additional efforts were then devoted to applying the
variants, are often easy to implement in combination AI metamodeling technique to the uncertainty analysis
with the available deterministic solvers, yet requiring a of a wind turbine airfoil. As exposed to harsh environ-
huge number of simulations for a statistical mental conditions, the in-service wind turbine blades
convergence. unavoidably suffer from insect and/or dust contamina-
Modern engineering analysis and design are fre- tions, blade erosion and/or corrosion, and massive sand
quently based on high-fidelity simulation tools such as build-up, which cause increased blade surface rough-
the computational fluid dynamics (CFD) for fluid flow ness to varying degrees. So far, extensive numerical and
and the finite element analysis (FEA) for structural experimental studies have concluded that the increased
strength and vibration. These high-fidelity simulation surface roughness, in particular around the blade lead-
tools are often difficult, if not impossible, to be directly ing edge (LE), can induce boundary layer disturbances,
used for the uncertainty analysis and design due to the earlier turbulence transition and eventually wind tur-
heavy computational load in spite of the continuing bine performance degradation.26–31 However, most of
advances of computing power. The MCS-based uncer- these studies only obtained deterministic results and
tainty analysis would be prohibitively expensive because ignored the wind turbine performance variations due to
of the requirements of even more CFD and/or FEA the stochastic nature of the surface roughness in real
simulations of sampling points. An efficient measure to conditions. In view of this, the motivation of the pres-
tackle this issue is to build variable-fidelity metamodels, ent uncertainty analysis was to investigate the impact
which are regarded as a ‘‘model of the model,’’3that is, a of surface roughness uncertainties on the aerodynamic
fast, inexpensive, effective but approximate model of performance of a wind turbine airfoil, although the
the high-fidelity model. So far, various metamodels used methodology can be extended without modifica-
have been developed, such as polynomial regression tions to other types of uncertainties.
(PR),4 Kriging (KG),5 multivariate adaptive regression This article is organized as follows. Section ‘‘AI
splines (MARS),6 artificial neural network (ANN),7,8 metamodels’’ mainly presents the basic formulations of
radial basis function (RBF),9,10 and support vector the three AI metamodels: ANN, RBF, and SVR. In
regression (SVR).11,12 Due to the greatly reduced com- section ‘‘Numerical assessment of metamodeling per-
putational cost, the metamodeling techniques have been formance,’’ a total of 10 analytical functions are tested
widely used for the analysis and design of engineering to compare these three AI models in terms of accuracy
devices.13,14 Typically, the applications of metamodel- and robustness, where the impacts of problem scale
ing techniques in turbomachinery design and analysis and sample size are carefully examined. The focus of
can be found in Chahine et al.,15 Mackman et al.,16 section ‘‘Uncertainty analysis of wind turbine airfoil’’ is
Sugimura et al.,17 Ma et al.,18 and Ju and Zhang.19 on the application of the metamodeling techniques to
One of the most important issues in metamodeling the uncertainty analysis of a wind turbine airfoil under
engineering problems is to evaluate the performances surface roughness uncertainties. Some conclusions are
of metamodels and then select the most appropriate drawn in section ‘‘Conclusion.’’
one for their design and analysis. Jin et al.20 performed
a systematic comparison among PR, KG, MARS, and
RBF. They aimed at developing standard procedures AI metamodels
and multiple criteria for metamodel evaluation. In the ANN
following studies, some other metamodels were also
compared in some particular applications.16,21–25 All The mathematical form of an ANN model with one
these comparative studies, to some extent, revealed the hidden layer and one output neuron can be written as7
great potential and promise of ANN, RBF, and SVR ! !
for approximation. These three models are acknowl- X
H X
N
^y = ^f (X) = f2 wh f1 vih xi + bh + b2 ð1Þ
edged as the artificial intelligence (AI) metamodels, h=1 i=1
which allow the behavior of self-learning and can
smartly extract the mechanisms that underlie the com- where X is a N-dimensional input vector with xi (i = 1,
plex systems. However, it remains unknown yet which 2,., N) as its elements; ^y is the model output for the
of them has the best approximation capability in gen- approximation result; f1 ( ) and f2 ( ) represent the
eral cases and whether they could be used for input-hidden and hidden-output transfer functions,
Ju et al. 3
P
X
RBF ^f X, ap , a =
p ap ap k(X, Xp ) + b ð7Þ
9,10 p=1
The mathematical form of a RBF can be written as
X
H where ap , ap , and b are the obtained coefficients and
^y = ^f (X) = wi f(jjX Xi jj) + b ð2Þ k(X, Xp ) is the kernel function measuring the distance
i=1 between the new input X and the pth sample point Xp .
Table 1 lists some common kernel functions for the
where jjX Xi jj represents the Euclidean distance
SVR. The predetermined parameters of SVR are the
between the input vector X and the central point vector
penalty coefficient C in equation (5) and the kernel
Xi of the ith hidden neuron, f( ) is a RBF, H is the
function variable d or s shown in Table 1.
hidden neuron number which may be equal or smaller
than the sample size, wi is the weight connecting the ith
hidden neuron and the output neuron, and b is the bias
Use of GA
in the output layer. Therefore, the RBF actually
expresses a weighted linear combination of H RBFs. In The approximation performances of AI metamodels
this study, the most commonly used Gaussian function are closely associated with the predetermined para-
is chosen as the RBF, which reads9 meters, which in most studies were specified or adjusted
! randomly. This may be less convincing or even mislead-
jjX Xi jj2 ing in that method A with a good set of parameters is
f(jjX Xi jj) = exp ð3Þ likely to achieve better approximations than method B
2si 2
with a poor set of parameters even if method B is gen-
where si is the spread factor and directly affects the erally more accurate than method A. With this in mind,
smoothness of the approximation result. Li et al.24 employed the hold-out validation plus grid
During the learning procedure of a RBF model, the search scheme to optimize the predetermined para-
predetermined parameters are the spread factors in meters of metamodels. However, this scheme is time-
equation (3) and the initial weights in equation (2). consuming and only limited to discrete parameter val-
ues. By contrast, modern evolutionary algorithms, for
example, GA, can be more efficient and versatile for
SVR finding out the global optimums.
With the concept of e-insensitive loss function, the
mathematical form of a SVR is11
Table 1. Common kernel functions for the SVR model.11
^y = ^f (X) = WF(X) + b ð4Þ
Kernel function Mathematical expression
where F( ) represents the nonlinear transformation
Linear k(X, X9) = (X X9)
imposed on the input vectors. The coefficients W and b
Polynomial k(X, X9) = ((X X9) + 1)d
can be obtained by solving the following convex quad- Gaussian 2
Small-scale 2 – 9 20 1000
3 – 11 25 1000
Large-scale 5 11 30 55 1500
8 17 40 60 2000
10 19 50 100 2000
In this work, GA was adopted to optimize the prede- (see Table 2). In summary, totally 6 3 2 + 4 3 3 = 24
termined model parameters through minimizing errors training sample sets were generated for approximation.
between the metamodel outputs and the actual outputs, Here, the terms ‘‘scarce,’’‘‘small,’’ and ‘‘large’’ did not
the motivations behind which are twofold: first, it elimi- denote fixed values but were defined with respect to the
nates the uncontrollable randomness in the modeling problem scale, that is, the factor q in the uniform DOE
process and thus enhances the metamodel training effi- table. Besides the training sampling sets, additional 10
ciency and robustness; second, GA helps to maximize testing sample sets were randomly generated within the
the approximation accuracy of each metamodel, which variation ranges to verify the approximation
thus strengthens the fairness of comparison in the fol- performance.
lowing section.
Here, it has to be remarked that the GA parameters
were tuned for different metamodels and different test Performance measures
cases. Detailed implementations of GA in metamodel To evaluate the approximation performance of the
optimization can be found in Ju and Zhang.19 metamodels, a group of measures have been developed,
such as accuracy, robustness, efficiency, transparency,
and simplicity.21 Among them, accuracy and robust-
Numerical assessment of metamodeling ness were the primary focus of this study because they
performance play essential roles in determining whether a metamo-
del is acceptable or not for approximation.
Descriptions of test problems
So far, the number of independent design variables is
generally limited below 10 in the simulation-based and Accuracy. Accuracy is defined as a metamodel’s capacity
metamodel-assisted design optimization of turboma- of predicting the system response over the pre-specified
chinery.13,16–18 A total of 10 analytic functions with dif- design space. To take a comprehensive measure of a
ferent degrees of nonlinearity and problem scales22,32,33 metamodel’s accuracy performance, three error metrics
were first chosen in this section for a comprehensive described in equations (8)–(10) were investigated, that
evaluation of the three AI metamodeling techniques. is, root mean square error (RMSE), averaged absolute
Among them, the first six functions have not more than error (AAE), and maximum absolute error (MAE)21
three independent variables and were classified as vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N
small-scale problems, while the other functions with 5, u1 X
8 or 10 variables were large-scale problems in this RMSE = t (^y yi )2 ð8Þ
N i=1 i
study. The mathematical expressions of these functions
are listed in Appendix 1.
1X N
The training sample sets were generated by means of AAE = j^y yi j ð9Þ
N i=1 i
the uniform design of experiment (DOE), which is
capable of providing fewer training samples with better MAE = maxj^yi yi j ð10Þ
uniformity and representativeness.34 In a uniform DOE
table Un(tq), n denotes the number of experiments or The above errors should be calculated from testing
the sample size while q and t represent the number of samples rather than training samples. Among them, N
factors (i.e. variables) and their levels, respectively. denotes the number of testing samples, yi is the actual
Accordingly, different training sampling sets were gen- (analytic) output of the ith testing sample, while ^yi is
erated for different functions, that is, small and large the corresponding metamodel output. Both RMSE and
sample sets for small-scale problems, while scarce, AAE provide the overall indications of the difference
small, and large sample sets for large-scale problems between metamodel outputs and actual data, while
Ju et al. 5
Normalization 1
RMSE =
RMSE 3 100
P
RMSEunder each sample size of the test function
ð11Þ
Normalization 2 Figure 1. Overall performance comparison among the three
metamodels: (a) mean values and (b) standard deviations.
RMSE =
RMSE 3 100
P Comparison results
RMSE under all sample sizes of the test function
ð12Þ Overall performance. The overall performances of the
GA-based AI (GA-AI) metamodeling techniques were
calculated as follows: for each metamodel,
wherePRMSE denotes the normalized value of RMSE Normalization 1 (equation (11)) was first implemented
and RMSE is the sum of RMSE of the three meta- to calculate its RMSE , AAE , and MAE for each test
models. Two other normalized errors, that is, AAE function under a certain sample size; then, the mean
and MAE , were defined similarly. Through normaliza- values and the standard deviations of RMSE , AAE ,
tion, the error metrics are closely associated with the and MAE were calculated through averaging across
concept of percentage, share the same order of magni- all the 24 approximation problems.
tude for different test functions, and can be added up The corresponding results are illustrated in Figure 1.
to calculate the averaged accuracies of a specific Compared with the other two metamodels, the GA-
metamodel based SVR (GA-SVR) achieved both lower mean val-
ues and lower standard deviations of the RMSE ,
m(RMSE ) = mean(RMSE ) ð13Þ
AAE , and MAE , indicating that GA-SVR has the
Two other mean values, m(AAE ) and m(MAE ), best overall performance and could be the best choice
were defined similarly. By such definition, lower values for metamodeling, especially when there is no knowl-
of these measures represent higher averaged accuracy edge about the problem scales and sample sizes. Except
of a metamodel for different approximation problems. for the s(MAE ), the GA-based ANN (GA-ANN)
generally performed the next best, followed by the GA-
based RBF (GA-RBF).
Robustness. Robustness measures a metamodel’s capac-
ity of achieving similar accuracies for different approxi-
mation problems. Mathematically, here it was defined Impact of problem scale. To examine each GA-AI meta-
as the standard deviation of error metrics across all the model’s performances in approximating problems with
problems of interest different scales, Normalization 1 (equation (11)) was
adopted, while a different averaging strategy from the
s(RMSE ) = Standard deviation(RMSE ) ð14Þ
above was used: for each metamodel, the mean values
Two other standard deviations, s(AAE ) and and standard deviations of errors were calculated first
s(MAE ), were defined similarly. Accordingly, a more for small-scale functions through averaging across a
robust and problem-independent metamodel should total of 6 3 2 = 12 approximation problems (i.e. the
have lower values of s(RMSE ), s(AAE ), and small and large sample sets for each of the first six func-
s(MAE ). tions); and then, the average was performed for large-
6 Advances in Mechanical Engineering
Figure 2. Performance comparison among the three metamodels for problems with different scales.
scale functions across the 4 3 3 = 12 approximation Figure 3 summarizes the performances of the three
problems (i.e. the scarce, small and large sample sets for GA-AI metamodeling techniques with different sample
each of the last four functions). The comparison results sets for small-scale problems. GA-SVR generally per-
are given in Figure 2. formed the best in spite of a slight inferiority in
Except for the m(MAE ) of small-scale problems, s(AAE ) under the large sample size. Moreover, all the
GA-SVR was generally better than the other two meta- three metamodels were found more accurate and
models. As for GA-ANN and GA-RBF, however, it is robust under the large sample size than under the small
difficult to determine which one is better: GA-ANN sample size, which could be attributed to the relatively
was more accurate but less robust than GA-RBF for rich information involved in the large sample sets. The
small-scale problems, while for large-scale problems, impact of sample size on the accuracy and robustness
GA-ANN showed less accurate but much more robust was the smallest for GA-RBF and the largest for GA-
performances. ANN.
Compared with the small-scale problems, the large-
Impact of training sample size. To further investigate the scale problems were less influenced by sample sizes (see
impact of training sample size on the performance of Figure 4). This is interesting and encouraging due to
metamodels, particular attention was paid to the the absence of the great expense for getting the large
approximation results under different training sample sample information. The averaged accuracies of GA-
sets. First, the values of RMSE*, AAE*, and MAE* SVR were almost equivalent for the three kinds of sam-
were obtained using Normalization 2 (equation (12)), ple sets, while significantly better than those of the
by which the best combination of the metamodel and other two metamodels. GA-RBF achieved the second
sample size could be obtained for each test function. best accuracy performance, followed by GA-ANN. In
Second, for each group of the small- and large-scale terms of the robustness performance, GA-SVR still
test functions, more detailed average was performed performed excellently for all the sample sets. GA-ANN
across approximation problems with the same type of was the most robust metamodel for the scarce and
sample set. small sample sets but its performance deteriorated
Ju et al. 7
Figure 3. Performance comparison among the three metamodels for small-scale problems under different sample sizes.
Figure 4. Performance comparison among the three metamodels for large-scale problems under different sample sizes.
8 Advances in Mechanical Engineering
Table 3. Different parameter combinations and numerical results of the uncertainty analysis in Case 1.
1 O O O O 22.25% 22.26%
2 O O 0 – 21.62% 21.60%
3 0 – O O 20.58% 20.58%
4 O 0.15 0 – 21.60% 21.60%
5 0 – O 0.15 20.59% 20.58%
6 0.21 O 0 – 21.61% 21.60%
7 0 – 0.21 O 20.58% 20.58%
MCS: Monte Carlo simulation; CFD: computational fluid dynamics; SS: suction surface; PS: pressure surface.
curve to characterize the nonuniformity of the surface prediction error of the constructed GA-SVR metamo-
roughness distribution around the airfoil LE. The inter- del was 0.93%.
polating points of the B-spline curve were supposed to Once the GA-SVR model was built, it was integrated
be randomly varied around the mean value, and any with the MCS method to conduct the uncertainty anal-
negative values on the curve were reassigned to be zero ysis. To further investigate how roughness uncertainties
to simulate the abrupt transitions between the rough affect the Clmax of the airfoil, several analyses were con-
and smooth regions in real conditions. The roughness ducted with different combinations of uncertain inputs.
parameters of both models had the same variation As shown in Table 3, the parameters marked with ‘‘O’’
ranges: the roughness height ks was from 0 to 0.42 mm were considered as the active uncertainties, while the
and the dimensionless covering length x/c varied from others were inactive with fixed deterministic values. For
0 to 0.3. the variations of Clmax relative to the smooth airfoil,
that is, DClmax , the related deterministic results were cal-
culated using the CFD method by setting each uncer-
Uncertainty analysis and results tain parameter to its mean level (0.21 mm for ks and
0.15 mm for x/c). From Table 3, it can be seen that the
In order to alleviate the huge computational cost in the
MCS results of m(DClmax ) agreed well with the CFD
uncertainty analysis by MCS, the GA-SVR metamodels
deterministic results of the DClmax , implying the high
presented in the previous sections were constructed to
accuracy of the constructed GA-SVR metamodel.
fast predict the Clmax of the airfoil with different surface
The calculated probability distributions of DClmax
roughness distributions. The two roughness uncertainty
for both multiple and single uncertainties are shown in
quantifications mentioned above were considered,
Figure 10. In Figure 10(a), the DClmax probability dis-
respectively, as follows.
tribution curve subjected to PS roughness uncertainties
was much slender and located on the right of the prob-
Case 1: Gaussian probability distribution. The first uncer- ability distribution curve subjected to SS roughness
tainty analysis concerned the roughness uncertainties uncertainties. This suggests that the SS roughness dis-
shown in Figure 8. A total of four uncertain inputs tributions do have the dominant impact on DClmax and
were considered, that is, ks and x/c for each of the SS further explain why the SS flow behavior is more sensi-
and pressure surface (PS) of the wind turbine airfoil. tive to the surface roughness as illustrated in Figure 8.
Therefore, this case was classified as a small-scale case. From Figure 10(b), it can be seen that for either SS or
A GA-SVR model was constructed to link the four PS of the airfoil, the roughness height ks and covering
roughness uncertain parameters and the Clmax of the length x/c had almost the same impacts on the DClmax .
According to the above results, the SS roughness
airfoil. According to the uniform DOE table U30(304),
distributions are the primary uncertainties that affect
a family of 30 roughness profiles were obtained for the
the aerodynamic performance and should be focused
airfoil as the training samples, which is in line with the
on in the future design and analysis of the wind turbine
suggestion obtained from Figure 3 that the GA-SVR
airfoil.
combined with a large sample set could achieve the best
approximation performance for small-scale problems.
Additional 10 roughness profiles were randomly gener- Case 2: B-spline distribution. In this case, the uncertainty
ated for the airfoil to be the testing samples. For each analysis was based on the second model shown in
of the training and testing samples, Clmax was calcu- Figure 9, which highlights the nonuniformity and ran-
lated by means of CFD. As a result, the averaged domness of the surface roughness distributed around
Ju et al. 11
drill in drilling composite materials. Int J Adv Manuf Tech 29. Pechlivanoglou G, Fuehr S, Nayeri CN, et al. The effect
2008; 37: 1061–1068. of distributed roughness on the power performance of
23. Kim BS, Lee YB and Choi DH. Comparison study on wind turbines. In: Proceedings of the ASME turbo expo
the accuracy of metamodeling technique for non-convex 2010: power for land, sea, and air, Glasgow, 14–18 June
functions. J Mech Sci Technol 2009; 23: 1175–1181. 2010, paper no. GT2010-23258. New York: ASME.
24. Li YF, Ng SH, Xie M, et al. A systematic comparison of 30. Huang CW, Yang K, Liu Q, et al. A study on perfor-
metamodeling techniques for simulation optimization in mance influences of airfoil aerodynamic parameters and
decision support systems. Appl Soft Comput 2010; 10: evaluation indicators for the roughness sensitivity on
1257–1273. wind turbine blade. Sci China Technol Sci 2011; 54:
25. Paiva RM, Carvalho ARD, Crawford C, et al. Compari- 2993–2998.
son of surrogate models in a multidisciplinary optimiza- 31. Sagol E, Reggio M and Ilinca A. Issues concerning
tion framework for wing design. AIAA J 2010; 48: roughness on wind turbine blades. Renew Sust Energ Rev
995–1005. 2013; 23: 514–525.
26. Madsen HA. Aerodynamics of a horizontal-axis wind tur- 32. Hock W and Schittkowski K. Test examples for nonlinear
bine in natural conditions. Risø-M 2903, 1991, http:// programming codes. New York: Springer-Verlag, 1981.
orbit.dtu.dk/en/publications/aerodynamics-of-a-horizon- 33. Brainin FK. A widely convergent method for finding
talaxis-wind-turbine-in-natural-conditions(1f4904c3- multiple solutions of simultaneous nonlinear equations.
819d-400a-9715-fcb3ffe74650).html IBM J Res Dev 1972; 16: 504–521.
27. Dalili N, Edrisy A and Carriveau R. A review of surface 34. Simpson TW, Lin DKJ and Chen W. Sampling strategies
engineering issues critical to wind turbine performance. for computer experiments: design and analysis. Int J
Renew Sust Energ Rev 2009; 13: 428–438. Reliab Appl 2001; 2: 209–240.
28. Ferrer E and Munduate X. CFD predictions of transition 35. Selig MS and McGranahan BD. Wind tunnel aerody-
and distributed roughness over a wind turbine airfoil. In: namic tests of six airfoils for use on small wind turbines.
Proceedings of the 47th AIAA aerospace sciences meeting NREL/SR-500-34515, 2004, http://m-selig.ae.illinois.edu/
including the new horizons forum and aerospace exposition, pubs/SeligMcGranahan-2004-NREL-SR-500-34515-
Orlando, FL, 5–8 January 2009, paper no. 2009-269. SixNRLAirfoils.pdf
Reston, VA: AIAA.
Appendix 1
Test functions
10
X
3 16 2 16
f (x) = + sin xj 1 + sin xj 1
8. j=1
10 15 15
1 xj 1
X 8
f (x) = 2xj 4 + (6 xj )
9. j=1
0 xj 10
0 1
X10 B xj C
B C
f (x) = xj Btj + ln 10 C
j=1
@ P A
xj
10. j=1