2012 - CALIBRATION OF GAS TURBINE BLADE TEMPERATURE PREDICTIONS USING SURROGATE MODELS

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Proceedings of ASME Turbo Expo 2012

GT2012
June 11-15, 2012, Copenhagen, Denmark

GT2012-69601

CALIBRATION OF GAS TURBINE BLADE TEMPERATURE PREDICTIONS USING


SURROGATE MODELS

John M. McFarland Grant O. Musgrove


Southwest Research Institute Southwest Research Institute
San Antonio, TX 78238 San Antonio, TX 78238

Sung Yong Chang David L. Ransom


Korea Electric Power Research Institute Southwest Research Institute
Daejeon, Korea 305-380 San Antonio, TX 78238

s radial span measured from hub


S radial span distance between hub and shroud
ABSTRACT t interdiffusion zone thermal exposure time (Eqn 1)
Actual gas turbine performance and component life at
T temperature; absolute temperature (Eqn 2)
specific engine installations is highly dependent on the actual Tavg inlet profile average temperature
operating conditions, since not all engines are operated in the
W interdiffusion zone width [µm] (Eqn 1)
same manner. Due to the variability in turbine operation, it
W0 initial interdiffusion zone width (Eqn 1)
may be prudent to evaluate the operation of hot section
components for turbine inlet conditions that are specific to a
Greek
single installation. However, determining the actual turbine
φ dimensionless absolute temperature, φ = T / Tavg
inlet conditions can be a difficult and expensive process that is
usually only done on test bed gas turbines. This paper θ vector of model calibration parameters
presents a method to determine turbine inlet conditions using a
model calibration approach. Two-stage CFD and thermal
Subscripts
0 inlet profile value at hub location
analyses are developed to predict blade temperature. By
varying the model inputs, computational predictions of blade 1 inlet profile value at shroud location
h radial span of inlet profile peak value
temperature are calibrated to blade interdiffusion zone
p radial distance of inlet profile peak value evaluated
thickness measurements. In order to speed up calculations,
surrogate models are used in place of the full-scale analysis from the hub
peak inlet profile peak value
codes during the calibration analysis. The result of the study
is a prediction of the turbine inlet profile necessary to obtain
the best agreement between predicted and measured blade INTRODUCTION
temperatures. During the design of the hot section of a gas turbine, a
range of operating conditions is taken into consideration for
achieving both maximum part life and maximum performance.
NOMENCLATURE It is difficult, however, to predict all operating conditions for a
A constant (Eqn 2) turbine fleet due to differences such as fuel composition and
kp interdiffusion zone parabolic growth rate constant operational history. Due to the variability in turbine operation,
(Eqn 1) it may be prudent to evaluate the operation of hot section
PRcool ratio of coolant source pressure to flow path static components for turbine inlet conditions that are specific to a
pressure single installation. Determining the actual turbine inlet
Q activation energy (Eqn 2) conditions can be a difficult and expensive process that is
R gas constant usually only done on specific test bed gas turbines.

1 Copyright © 2012 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


This paper demonstrates the use of a model calibration COMPUTATIONAL APPROACH
approach to determine inlet conditions based on blade A brief overview of the computational tools and approach
temperature measurements. Turbine blade temperature is used to perform the calibration study is given here. The
modeled using two-stage CFD and blade thermal analyses. calibration approach can be broken into the following steps:
The model inputs are then systematically adjusted in order to
maximize the agreement between the analysis predictions and • Develop analysis models to calculate blade temperatures
the blade temperature measurements. However, the total run for a set of input parameters.
time for the sequence of analysis models used to predict blade • Use the NESSUS software package to coordinate the
temperature is on the order of 10 hours for one case, rendering analysis models, define the inputs, and collect the training
direct use of the analysis models within an iterative search data.
algorithm prohibitive due to computational cost. Thus, • Use custom Python scripts to collect and organize the
surrogate models, also referred to as Response Surface results generated by NESSUS based on the 106
Approximations (RSA’s), are employed in order to develop a temperature response locations of interest.
fast-running approximation of the full analysis model in order • Use the CENTAUR library to generate response surface
to facilitate the calibration study. models from the training data and perform the calibration
One of the advantages of using surrogate models is their analysis. CENTAUR (Collection of Engineering Tools
flexibility: once constructed, the surrogate model may be used for Analyzing Uncertainty and Reliability) is a modular
for different studies, such as sensitivity analysis, optimization, library of probabilistic algorithms that facilitates the rapid
calibration, and uncertainty quantification. In fact, a large development of custom solutions. CENTAUR also
amount of recent work in the turbomachinery community has provides many of the algorithms used by NESSUS.
addressed the use of surrogate models for these types of
studies. The majority of these studies have focused on design NESSUS (Numerical Evaluation of Stochastic Structures
optimization (see, for example [1,2,3,4]). Multi-objective Under Stress) [18] is a general purpose software tool for
design optimization using response surface approximations probabilistic analysis and model integration. NESSUS
has also been of particular interest (see [5,6,7,8,9,10]). provides mechanisms for mapping input variables into text-
The use of surrogate models for model calibration studies based numerical model input files, defining files that must be
has not been as widespread in the gas turbine industry. Wang copied between sequential evaluations, and quickly restarting
et al. [11] gives an overview of the current state of the art, an analysis in the case of computer system failure or analysis
with an emphasis on the Kennedy and O’Hagan framework model errors.
[12], which is a comprehensive Bayesian approach that makes For this study, NESSUS is used to exercise the analysis
use of Gaussian Process (a.k.a. Kriging) surrogate modeling. models and collect the training data, as shown in Figure 1. The
Recent work by Lin et al. [13] demonstrates a surrogate-based NESSUS model coordinates a two-stage CFD simulation
calibration and bias detection case study on gas turbine using ANSYS CFX [19], a finite element thermal model of the
performance metrics. Other work in the gas turbine literature first stage blade using ANSYS Mechanical [20], and a finite
has also demonstrated the use of surrogate models for design element thermal model of the second stage blade using
space exploration [14,15] and probabilistic analysis [16]. ANSYS Mechanical. The model includes definitions of how
One of the challenges faced in this calibration study is input variables map into the numerical model input files, as
that the temperature measurements are given at over 100 well as how output files from the CFD analysis must be copied
locations across two blades. Thus, if the model is to be and used as input files for the thermal analyses. The Latin
calibrated against all data, the surrogate model must be Hypercube Sampling (LHS) probabilistic method in NESSUS
capable of representing the model response for all of these is used to define the design of computer experiments for the
locations. In other words, the surrogate model must deal with input variables.
highly multivariate output. This is a typical challenge when
dealing with CFD applications, and a common approach for
developing surrogate models has been to represent the high-
dimensional output space in terms of a reduced set of INPUT:
orthogonal variables using an approach such as Principal Temperature profile
Components Analysis (PCA) or Proper Orthogonal Coolant pressure ratio
Decomposition (POD) (for example, [17]). However, the NESSUS
approach demonstrated here is to develop individual surrogate CFD Simulation
models for each of the model outputs. While this increases the
computational burden somewhat because of the need to work
with a large number of surrogate models, it has the advantages Thermal Simulation
of not losing information due to order reduction as well as the
opportunity to identify the response surface goodness of fit as Figure 1. The NESSUS software is used to execute the
a function of the output variables. deterministic models, define input data, and evaluate the
output data.

2 Copyright © 2012 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


NESSUS creates an organized set of directories the results are independent of the mesh size. All solutions use
containing the data associated with each set of numerical the k-ω turbulence model with automatic wall treatment,
analyses. Custom Python scripts are then used to parse the whereby the convergence criteria is 10-6 residual values. The
directory structure and organize the training data based on flow field results from the CFD simulation are used to
each of the 106 temperature response quantities of interest. calculate the heat transfer coefficients on the blade airfoil,
The training data are then processed using custom scripts platform, and tip surfaces using widely accepted empirical
written with the CENTAUR library. These scripts use correlations for surface heat transfer coefficient.
CENTAUR to create all surrogate models and conduct the Exit boundary
calibration analysis. Mass flow rate

TURBINE MODELING
Blade temperatures are predicted for the first two stages
of a Westinghouse W501F gas turbine using ANSYS
Mechanical [20], a commercially available finite element
solver. Thermal models of the first and second stage blades
are used to solve for the steady-state temperature distribution
in the blade material. Because the finite element thermal
models are also used for stress predictions in Kim et al. [29], a
mesh size of one million nodes is required to produce mesh
3,600 rpm
independent results of stress. A large mesh size is required Inlet boundary
because complex geometry features such as the squealer tip, Pressure and Temperature
fillets, and internal cooling passages are included in the model. Figure 3. A steady-state CFD model of the first two turbine
The fine mesh size is applied in the platform and airfoil stages is used with empirical heat transfer correlations to
sections of the blade, whereas the blade root has a coarse mesh generate the external boundary conditions for the thermal
because it was not a focus of the study, as shown in Figure 2. blade models.

For the airfoil and platform surfaces, a correlation for a


turbulent flat plate [21] is used to calculate the heat transfer
coefficient along the surfaces. The stagnation region on the
blade leading edge is approximated as the plate leading edge
because the boundary layer on the airfoil begins to grow from
the stagnation point. A correlation from Glassman [22] is
used to calculate the heat transfer coefficient on the leading
edge region by approximating the leading edge as a cylinder in
cross flow. The results of both the flat plate and the cylinder-
in-cross flow correlations are blended to form a continuous
curve. Heat transfer coefficients are calculated on the blade
tip using the correlation developed in Metzger et al. [23]. The
blade tip leakage flow is calculated from the pressure ratio
between the pressure and suction sides of the blade near the
Figure 2. A mesh independency study confirms that the tip. The flow rate is calculated using the pressure ratio along
one million node mesh sizes for the first (left) and second with loss coefficients to approximate the contraction and
(right) stage blades is required for independent results. expansion losses across the grooved blade tip.
Driving temperatures applied to the thermal models are
Inputs to the thermal models are the internal and external taken directly from the flow field solution, In the case of the
heat transfer boundary conditions in the form of convection first stage blade, the thermal boundary conditions are adjusted
coefficients and driving temperatures. The external thermal to account for the effects of film cooling on the airfoil surface
boundary conditions are calculated from the flow field results [24,25]. The effects of thermal barrier coating (TBC) are also
of a two-stage CFD simulation of the turbine, as shown in included by combining the thermal resistance of the external
Figure 3. The steady-state simulation is solved using ANSYS convection and the TBC conduction. The coating thermal
CFX [19], a commercially available CFD solver, for specified conductivity is approximated as 2 W/m2/K [26].
turbine inlet temperature and pressure profiles. The model The internal thermal boundary conditions are calculated
utilizes rotationally periodic interfaces at the circumferential using empirical correlations found in the literature for the pin
faces and mixing plane interfaces in between components. fin [27] and ribbed cooling geometries [28]. Internal cooling
The mesh is generated using ANSYS ICEM v12.0 [19] to flow rates are calculated outside of the thermal model using a
create a hexahedral mesh with an O-grid around the airfoils. one-dimensional flow network, whereby the flow rates can be
The CFD model is a 1 million element mesh of the flow field expressed as a function of the ratio of the coolant-to-flowpath
through the first two stages of the turbine with y+ values less pressure (PRcool).
than 20. A mesh sensitivity study is completed to ensure that

3 Copyright © 2012 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


Model Input Variables coatings. The β-phase in the coating acts as a reservoir for
The objective of the model calibration study is to aluminum and provides aluminum to form protective oxide
determine an appropriate inlet temperature profile and coolant scale at the outer surface. Aluminum diffuses during service to
flow based on observed blade temperature. The average inlet re-form continuous scale when the protective oxide layer
temperature value and the inlet pressure profile are held exfoliates. This results in depletion of aluminum in the
constant. To perform the calibration study, the inlet coating, which promotes transformation of β into γ. Aluminum
temperature profile is represented parametrically in terms of a in the coating also diffuses into the substrate and nickel
small set of variables as shown in the schematic of Figure 4. diffuses from the substrate into the coating. Interdiffusion of
For this study, φpeak is not treated as a variable but is instead these elements during service is responsible for the
determined based on the constraint that the average inlet enlargement of the interdiffusion zone. The coating finally
temperature has a fixed, known value. The average fails to provide protection of the blade when the aluminum
temperature of the inlet profile is represented as the area under content in the coating drops below 3 wt.% or β-phase
the profile curve, and is known from gas turbine operating transforms into γ. The interdiffusion zone growth depends on
data. As a result, four variables are used to control the inlet service time and temperature since it is a diffusion controlled
temperature profile: φ0, φ1, Sp, and Sh. process.
The interdiffusion zone growth process can be expressed
φ1
by a parabolic relation:
1
(W − W0 )2 = k pt (1)

φpeak The parabolic growth rate constant, kp, is related to the


s/S sh exposure temperature by the Arrhenius equation, since the
interdiffusion zone growth is a diffusion controlled process:

k p = (W − W0 ) t = A exp(− Q RT )
2
sp (2)

To estimate operating metal the constants need to be


determined for the blade coating/substrate system of interest.
0 φ The accuracy of operating temperature estimation depends on
φ0 the initial zone width, W0.
Figure 4. Parametric Dimensionless Temperature (φ)
Profile Model. The Y-axis Represents Non-dimensional Temperature data
Radial Span The temperature data provided by the project sponsor are
based on measurements of interdiffusion zone thickness as
In addition to the inlet temperature profile, the coolant described above and are plotted below in Figure 5 and Figure
pressure ratio (PRcool) is a parameter used to determine the 6. Positive chord represents the pressure side of the airfoil and
coolant mass flow in the blade. Coolant pressure ratio is negative chord represents the airfoil suction side. The blade
considered as a model parameter because the actual mass flow leading edge is located at chord location 0 and the trailing
in the blade is a function of the coolant supply pressure and edge at chord locations ±1.
the gas path pressure. Thus, a total of five inputs are Typically, the operating metal temperature of a blade
considered as variables for this study. varies from location to location at a given airfoil height. The
blade (50 to 60% airfoil) mid height is exposed to the highest
TEMPERATURE MEASUREMENTS temperature and the blade near the hub or blade platform is
The objective of the model calibration study is to identify exposed to the lowest temperature. As a result, the
the turbine inlet conditions so that the blade temperature microstructure of the coating and the interdiffusion zone width
predictions agree with experimentally determined values. The near the blade platform is considered to represent the original
model is calibrated based on experimental measurements condition. In contrast, the interdiffusion data shown in Figure
obtained by the project sponsor. These measurements are of 5 and Figure 6 do not show a significant variation between the
the interdiffusion zone thickness values, from which blade hub, mid or tip locations. The maximum estimated
temperature values are determined. temperatures at the hub and mid height location were
comparable (differing by only φ = 0.03), suggesting that the
Experimental temperature determination coating at the hub location was also significantly degraded.
Degradation of the coating during long-term service at Thus, future work may be done to verify the temperature
elevated temperatures is a result of the transformation of the measurements, but the values are still used here as a
β-phase ((Ni and/or Cr)Al) in the coating into γ and an demonstration of the calibration analysis.
increase in the interdiffusion zone width below the MCrAlY

4 Copyright © 2012 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


Figure 5. Observed Temperature Distribution on Blade 1

Figure 6. Observed Temperature Distribution on Blade 2

interdiffusion zone thickness measurements, as described in


the previous section. This includes 106 total measurements
MODEL CALIBRATION ANALYSIS AND RESULTS across 18 locations on the tip, mid, and hub sections of two
The objective of the model calibration study is to adjust blades (measurements are missing at two locations for the tip
certain model parameters, referred to as the model calibration of Blade 1).
parameters, in order to achieve agreement between model
The model calibration analysis is formulated as a
predictions and experimentally observed data. In this case, the nonlinear least-squares problem. The objective is to identify
model is calibrated against the observed blade temperature
the values of the calibration parameters, θ, that minimize the
distribution, which is determined based on interdiffusion zone
sum of squared errors function, which measures the amount of
thickness measurements. For this study, the model calibration
mismatch between model prediction and experimental data:
parameters are those which determine the shape of the inlet n
temperature profile, as well as the coolant-to-flow path
pressure ratio, PRcool, as described above.
S(θ ) = ∑ [y
i =1
i − G (θ, x i )] ,
2
(3)
The bounds associated with the calibration parameters
are given in Table 1. where ( y1 ,  , y n ) denote the n available measurements and
G denotes corresponding model predictions, which are a
Table 1. Calibration Parameters and Bounds function of the calibration parameters, θ, and independent
Parameter Lower Bound Upper Bound variables, x. In this case x is just a conceptual indicator
variable so that there is a correspondence between the spatial
φ0 0.59 1.0 locations of the predictions and observations.
φ1 0.59 (2φavg-φ0*Sp) / (2-Sp) Because evaluation of the model predictions G(θ,x) is
Sh 0 1-Sp-0.001 extremely time-consuming, a surrogate-based approach is
Sp 0.05 0.95 used. For each input vector θ, evaluation of G involves a
series of 3 analysis models, requiring approximately 10 hours
PRcool 1.05 1.20
of computation time on an 8-core processor. The 3 analysis
models are the CFD model, thermal model of blade 1, and
The experimental data to which the model is to be thermal model of blade 2. Surrogate models (also known as
calibrated are the temperature values determined based on the

5 Copyright © 2012 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


response surface approximations) are constructed to
approximate the relationship between temperature and the S h = 0 + QS h × 1 − S p − 0.001 − 0 (4)
  
model calibration parameters. A separate surrogate model,
Thus, a uniformly distributed Latin Hypercube design is
Ĝi(θ), is constructed for each blade temperature location of constructed on a set of auxiliary variables each having a range
interest, resulting in a total of 108 surrogate models. (0,1). Based on the bounds given in Table 1, the auxiliary
The surrogate models are trained by exercising the true variables are then converted into calibration parameter values.
model, G, with various combinations of the calibration inputs, A total of 25 training points are randomly generated using this
θ, and observing the response values. The selection of the set design. An additional training point is also collected at the
of inputs used to collect the training data is known as the nominal values of the inputs (each input is set halfway
Design of Computer Experiments. The Latin Hypercube between the lower and upper bounds given in Table 1), which
Sampling (LHS) design [30] is commonly used for collection gives a total of 26 training points. After specification of the
of training data (for example [6,4,8,15]) because it possesses design, the turbine blade model is exercised for each of the 26
space-filling properties that help achieve good coverage of the cases, which involves execution of the CFD model as well as
design space [31]. the blade 1 and blade 2 thermal models.
The Latin Hypercube design is constructed by applying a The temperature profiles predicted by each of the 26
uniform probability distribution to each input, based on the training point runs are plotted in Figure 7 and Figure 8. Notice
given lower and upper bounds. One challenge is that in some that, as discussed above, the interdiffusion data show much
cases, the bounds for one input depend on the values of one or less temperature variation across the blade than is predicted by
more other inputs. The approach used here is to introduce a the models. Based on these plots, it is clear that it will not be
set of auxiliary variables that take values from (0,1). The possible, by adjusting only the parameters in Table 1, to
auxiliary variables define the location of the corresponding achieve a flat temperature profile that matches closely with the
input between its minimum and maximum value. Consider, interdiffusion data. Nevertheless, we can still search for the
for example, Sh, whose upper bound depends on the value of optimal solution, which is the one that minimizes the error
Sp. We introduce an auxiliary variable, say QSh and define Sh given by Eq. (3).
as the minimum value plus a fraction of the range (upper
minus lower).

Figure 7. Blade 1 temperature profiles predicted by the 26 calibration training point runs. For each curve, the input parameters
given in Table 1 were randomly chosen. The dots represent the experimental data.

Figure 8. Blade 2 temperature profiles predicted by the 26 calibration training point runs. each curve, the input parameters
given in Table 1 were randomly chosen. The dots represent the experimental data.

6 Copyright © 2012 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


A variety of different surrogate modeling approaches One such model is the Gaussian process (GP) model, also
have been demonstrated in the literature, ranging from simple known as a Kriging model [32]. The GP model is capable of
polynomial regression models [6,7,13] to sophisticated neural representing more complex response functions, and has the
networks [14,3,4,8] and Gaussian Process (a.k.a Kriging) advantage that it is not restricted to a particular functional
models [5,1,2,3,9,10]. For this study, we first consider linear form, like regression models. The form of the GP surface used
regression models for each of the Ĝi(θ). If the linear models in this work is:
represent the training data well, there would be no advantage m
to using a more sophisticated approach. The form of the linear Ĝ i,GP (θ ) = β i + ∑κ i,j (θ )(Yi,j − β i ) , (6)
model is given by: j=1
5
where m is the number of training points (26), βi is a
Gˆ i ,linear (θ ) = β i , 0 + ∑ β i , jθ j (5)
j =1 coefficient for each model, Yi,j is the jth training point
response value for the ith model, and κi,j(θ) is a weighting
where for the ith surrogate model, β i , j , j = 0,  ,5 are function based on the spatial correlation function. Further
coefficients that are estimated based on the training data. To details of the modeling approach used here can be found in
assess the goodness of fit, the R-squared value (also known as references [33,34].
the coefficient of determination), is computed for each One of the desirable features of the GP model is that,
surrogate model. The R-squared value is equal to the fraction unlike regression models, the response surface directly
of the variance in the training data outputs that is explained by interpolates all of the training points. This makes model
the surrogate model. For regression models, the R-squared assessment slightly more challenging, though. For this study,
value will always increase as higher order terms are added, so model assessment for the GP model is based on the “leave-
over fitting is sometimes a concern. However, for this case, one-out” cross-validation residuals [35]. The leave-one-out
over fitting is not a concern because the number of training residuals are obtained by computing the difference between
points (26) is significantly greater than the number of the observed and predicted response at each training point,
coefficients (6). Figure 9 shows a plot of the R-squared values while each training point is removed from the model one at a
for all 108 surrogate models (although there are only 106 time. These residuals are then used to compute an R-squared
experimental data points, since measurement data were value, which is plotted for each GP surrogate model in Figure
missing for two of the locations). The models are arranged on 10. The R-squared value computed in this fashion should be
the horizontal axis based on their spatial location in terms of slightly conservative, since one training point is removed
the “tip”, “mid”, and “hub” radial span locations of the blade. when computing each residual. As expected, the GP models
Based on Figure 9, we see that the R-squared values for fit the training data much more accurately than the linear
the linear models range from approximately 0.7 to 0.95. The regression models. However, the fit on the Blade 2 mid-
hub sections for both blades show the largest R-squared section could be improved with additional training points to
values, indicating that the temperature responses are improve the accuracy of the surrogates. Nevertheless, the fit
approximately linear in the input variables in these sections of of the surrogate models to the blade 2 mid section are likely
the blade. However, the R-squared values are lower in other not significantly affecting the calibration results when
sections, particularly on the mid-section of Blade 2, where it is comparing the surrogate model results with other blade
apparent that the linear models provide a fairly poor fit to the locations.
training data. These results suggest that consideration of a
more sophisticated surrogate model is warranted.

Mid Tip

Hub
Hub

Mid Blade 2 Mid


Tip

Blade 1 Blade 2
Figure 10. R-squared show good fit except for the Blade 2
mid section when calculated based on “leave-one-out”
Figure 9. R-squared values based on linear surrogate residuals for the Gaussian process surrogate models fit to
models fit to the 26 training points in five variables. the 26 training points in five variables.

7 Copyright © 2012 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


Once the surrogate models have been constructed, 1.0
virtually any numerical optimization routine can be used to
solve for the optimal calibration parameters by minimizing Eq. 0.8
(3), since the surrogate models are not expensive to evaluate.
For this study, two different optimization algorithms are 0.6
considered: (1) the Levenberg-Marquardt algorithm [36], s/S
which is a specialized least-squares solver, and (2) the 0.4
DIRECT algorithm [37], which is a general-purpose global
optimization routine. The Levenberg-Marquardt algorithm is
0.2
gradient-based, and so is capable of efficiently finding a good
solution, if given a good starting point. DIRECT is not
gradient-based and is more expensive (requires more function 0.0
0.90 0.95 1.00 1.05 1.10
evaluations), but it is less susceptible to finding a local φ
minimum instead of the global minimum. In each case, the
Figure 11. Calibrated Inlet Profile
calibration algorithm is constrained to search over the
parameter bounds identified in Table 1. This is necessary Comparing the predicted blade temperatures at the
because exploration beyond these bounds would result in calibrated inlet profile to the temperatures estimated using the
surrogate model extrapolation, for which the accuracy of the interdiffusion zone thickness measurements show a significant
surrogate models has not been assessed. difference. The differences lie in the temperature magnitude
In the end, the best solution is found by first running the and the distribution along the blade surface. As suggested
DIRECT algorithm to get an approximate globally-optimal earlier, verification of the temperatures based on the
solution. This solution is then used as the starting point for the interdiffusion zone thickness measurements is suggested due
Levenberg-Marquardt algorithm, which is able to improve the to the flat distribution of values across the blade surface in
solution slightly. The calibration solution is given in Table 2 both the chord-wise and radial directions. In contrast, the
and Figure 11. Figure 12 and Figure 13 plot the temperature predicted blade temperatures for the calibrated condition are
profile predicted for the calibrated solution (based on the shown to vary in both the chord-wise and radial directions.
surrogate models) against the experimental data. Total Nevertheless, the importance of having careful data to
computation time to arrive at the calibration solution is driven calibrate to is highlighted by this work.
by the cost of collecting the 26 training points, at roughly 10 It is important to note that the agreement between the
hours each, for a total cost on the order of 260 hours. surrogate model and the interdiffusion zone temperatures
could be improved by increasing the design space of the input
Table 2. Calibration Solution
parameters. The design space, however, is not increased
Parameter Value because the upper and lower bounds on the input parameters
φ0 0.955 are set to ensure a realistic temperature profile. For example,
the average value of the temperature profile could be included
φ1 0.954 as an input variable to allow the magnitude of the surrogate
Sh 0.253 model predictions more closely match the interdiffusion zone
Sp 0.05 temperatures. However, the calibrated average temperature
PRcool 1.05 would be much higher than the known operating temperature
of the gas turbine, which was held constant during the
calibration.

Figure 12. Blade 1 temperature profile predicted for calibrated solution (green curve) and experimental data (blue points).

8 Copyright © 2012 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


Figure 13. Blade 2 temperature profile predicted for calibrated solution (green curve) and experimental data (blue points).

it possible to identify specific regions on the blade where the


surrogates are not fitting as well.
CONCLUSIONS
The final result of the analysis is the calibrated inlet
Hot section gas turbine design is normally conducted
temperature profile. For this particular study, there were
with respect to a range of turbine operating conditions.
significant differences between the predicted and measured
Because of differences in turbine operation, it may be
blade temperature distributions, and the calibration analysis
advantageous to evaluate the operation of hot section
was limited to finding the best possible solution within the
components for inlet conditions that are specific to a particular
constraints defined for the problem. Nevertheless, the
installation. However, measurement of turbine inlet
surrogate-based approach shows promise, and it would even
conditions can be a difficult and expensive process.
be possible to re-calibrate the inlet profile using a new set of
The goal of this paper is to present an approach for
temperature data, without having to perform any additional
determining inlet conditions using analysis models and
analysis runs with the full models. New temperature data is
estimates of blade temperature from interdiffusion zone
suggested with verification of the calculation procedure from
thickness measurements. By using a model calibration
the interdiffusion zone thickness measurements. Future work
approach, model inputs are adjusted to optimize agreement
from this study could also include an expansion of the input
between the predicted and measured blade temperature values.
parameters to study the uncertainty effects of the empirical
One of the primary challenges with using such an approach is
models.
the computational expense associated with evaluating the
analysis model. Each individual analysis combines CFD and
heat transfer calculations to require 10 hours of run time. ACKNOWLEDGMENTS
Because the model calibration analysis requires a large The authors would like to thank Korea Electric Power
number of simulation runs with different combinations of the Research Institute for funding and sponsoring this research
inputs, direct use of the simulation within the calibration study effort including Woosung Choi, Bum Shin Kim, and Beom
is not feasible. Soo Kim.
To overcome the problem with computational cost, fast-
running surrogate models (also known as response surface
approximations) are developed. A total of 26 analyses are REFERENCES
made with the full simulations, and the resulting “training [1] Song, P., Sun, J., Wang, K., Zhilong, H., 2011,
data” are used to fit the surrogate models. “Development of an Optimization Design Method for
One of the challenges often faced when using surrogate Turbomachinery by Incorporating the Cooperative
models for CFD applications is that the simulations typically Coevolution Genetic Algorithm and Adaptive
have a large number of outputs that are of interest. In this Approximate Model,” GT2011-45411.
work the calibration study is performed with respect to [2] Carnie, G., Wang, Y., Qin, N., Shahpar, S., 2011,
temperature measurements at 106 blade locations, meaning the “Design Optimisation of Casing Grooves Using the
surrogate models must describe all 106 temperature responses. Zipper Layer Meshing Method,” GT2011-45483.
The issue with high-dimensional output is often addressed [3] Verstraete, T., Coletti, F., Bulle, J., Vanderwielen, T.,
using a dimension-reduction approach such as principal Arts, T., 2011, “Optimization of a U-Bend for Minimal
components analysis. This study demonstrates a different Pressure Loss in Internal Cooling Channels – Part I,
approach in which individual surrogate models are developed Numerical Method,” GT2011-46541.
for each response quantity of interest. One of the advantages [4] Kim, J., Choi, J., Kim, K., 2009, “Design Optimization
of this approach is that the goodness of fit of the surrogate of a Centrifugal Compressor Impeller Using Radial
models can be analyzed with respect to blade location, making Basis Neural Network Method,” GT2009-59666.

9 Copyright © 2012 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


[5] Lee, K. Kim, S., Kim, K., 2011, “Multi-Objective [22] Glassman, A., Turbine Design and Application (NASA
Optimization of Film-Cooling Holes Considering Heat SP-290), Vol. 3, 1975.
Transfer and Aerodynamic Loss,” GT2011-45402. [23] Metzger, D.E., Bunker, R.S., Chyu, M.K., 1989,
[6] Kim, J., Choi, K., Kim, K., 2011, “Optimization of a “Cavity Heat Transfer on a Transverse Grooved Wall in
Transonic Axial Compressor Considering Interaction of a Narrow Flow Channel,” Journal of Heat Transfer, 111
Blade and Casing Treatment to Improve Operating (73).
Stability,” GT2011-45404. [24] Baldauf, S., Scheurlen, M., Schulz, A., Wittig, S.,
[7] Iyengar, V., Sankar, L., 2011, “Comprehensive “Correlation of Film-Cooling Effectiveness From
Application of a First Principles Based Methodology Thermographic Measurements at Enginelike
for Design of Axial Compressor Configurations,” Conditions,” 2002, Journal of Turbomachinery, 124.
GT2011-45936. [25] Baldauf, S., Scheurlen, M., Schulz, A., Wittig, S., “Heat
[8] Arabnia, M., Ghaly, W., 2009, “A Strategy for Multi- Flux Reduction From Film Cooling and Correlation of
Point Shape Optimization of Turbine Stages in Three- Heat Transfer Coefficients From Thermographic
Dimensional Flow,” GT2009-59708. Measurements at Enginelike Conditions,” 2002, Journal
[9] Keskin, A., Swoboda, M., Flassig, P.M., Dutta, A.K., of Turbomachinery, 124.
Bestle, D., 2008, “Accelerated Industrial Blade Design [26] Su, Y.J., Trice, R.W., Faber, K.T., Wang, H., Porter,
Based on Multi-Objective Optimization Using W.D., “Thermal Conductivity, Phase Stability, and
Surrogate Model Methodology,” GT2008-50506. Oxidation Resistance of Y3Al5O12 (YAG)/Y2O3-
[10] Becker, K., Lawerenz, M., Voß, C., Mönig, R., 2008, ZrO2 (YSZ) Thermal-Barrier Coatings,” Oxidation of
“Multi-Objective Optimization in Axial Compressor Metals, 61 (314), 2004, pp 253-271.
Design Using a Linked CFD-Solver,” GT2008-51131. [27] Metzger, D.E., Shepard, W.B., Haley, S.W., 1986,
[11] Wang, L., Fang, X., Subramaniyan, A., Jothiprasad, G., “Row Resolved Heat Transfer Variations in Pin Fin
Gardner, M., Kale, A., Akkram, S., Beeson, D., Wiggs, Arrays Including Effects of Non-Uniform Arrays and
G., Nelson, J., 2011, “Challenges in Uncertainty, Flow Convergence,” 86-GT-132.
Calibration, Validation and Predictability of [28] Han, J.C., 1988, “Heat Transfer and Friction
Engineering Analysis Models,” GT2011-46554. Characteristics in Rectangular Channels with Rib
[12] Kennedy, M.C., O’Hagan, A., 2001, “Bayesian Turbulators,” Journal of Heat Transfer, 110.
calibration of computer models”, J. Royal Stat. Soc., [29] Kim, B.S., Kim, B.S., Choi, W., Musgrove, G.O.,
Ser. B, 63, pp. 425-464. McFarland, J., Fierro, F., Ransom, D.L., 2012, “Gas
[13] Lin, T., Mendoza, E., Kestner, B., 2011, “Model-Based Turbine Blade Stress and Temperature Sensitivity to
Data Reconciliation and Bias Detection for Heavy-Duty Turbine Inlet Profile and Cooling Flow,” GT2012-
Industrial Gas Turbines Performance Diagnosis,” 69603.
GT2011-45943. [30] McKay, M., Beckman, R., Conover, W., 1979, “A
[14] Schmitz, A., Aulich, M., Nicke, E., 2011, “Novel Comparison of Three Methods for Selecting Values of
Approach for Loss and Flow-Turning Prediction Using Input Variables in the Analysis of Output from a
Optimized Surrogate Models in Two-Dimensional Computer Code,” Technometrics 21 (2), pp. 239-245.
Compressor Design,” GT2011-45086. [31] Santner, T., Williams, B., Notz, W., The Design and
[15] Garrison, L., Walter, S., 2009, “Multidisciplinary Analysis of Computer Experiments, Springer Verlag,
Design Space Characterization of a Prediffuser, Strut, New York, 2003.
and Frame,” GT2009-60159. [32] Stein, M., Interpolation of spatial data: some theory
[16] Cui, W., Wang, J., 2011, “Probabilistic Analysis of Gas for Kriging, Springer-Verlag, New York, 1999.
Turbine Disk Multi-Crack Propagation,” GT2011- [33] McFarland, J., “Uncertainty Analysis for Computer
45439. Simulations through Validation and Calibration,” PhD
[17] Lang, Y., Malacina, A., Biegler, L., Munteanu, S., Dissertation, Vanderbilt University, 2008.
Madsen, J., Zitney, S., 2009, “Reduced Order Model [34] Bichon B., Eldred, M., Swiler, L., Mahadevan, S.,
Based on Principal Component Analysis for Process McFarland, J., “Efficient global reliability analysis for
Simulation and Optimization,” Energy & Fuels 23, pp. nonlinear implicit performance functions,” AIAA
1695-1706. Journal 46(10), pp. 2459-2468, 2008.
[18] Thacker, B.H., Riha, D.S., Fitch, S.H.K., Huyse, L.J., [35] Martin, J., Simpson, T., “Use of Kriging models to
Pleming, J.B., 2006, “Probabilistic engineering analysis approximate deterministic computer models,” AIAA
using the NESSUS software,” Structural Safety, 28, pp. Journal 43(4), pp. 853-863, 2005.
83-107. [36] Seber, G.F., Wild, C.J., Nonlinear Regression, John
[19] ANSYS, Inc., Version 12.0, 2011 (ANSYS Inc.: New Wiley & Sons, Inc., Hoboken, New Jersey, 2003.
Hampshire). [37] Jones, D., Perttunen, C., Stuckman, B., “Lipschitzian
[20] ANSYS, Inc., Version 13.0, 2011 (ANSYS Inc.: New optimization without the Lipschitz constant,” Journal of
Hampshire). Optimization Theory and Application 79(1), pp. 157-
[21] Incropera, F.P., Dewitt, D.P., Introduction to Heat 181, 1993.
Transfer, 4th ed., John Wiley & Sons, New York, 2002.

10 Copyright © 2012 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

You might also like