Jurnal Propulsi

Energy Conversion and Management 271 (2022) 116245
Contents lists available at ScienceDirect
Energy Conversion and Management

journal homepage: www.elsevier.com/locate/enconman
Optimizing the performance of a single-shaft micro gas turbine engine by

modifying its centrifugal compressor design
S.M. Hosseinimaab , A.M. Tousi *
Department of Aerospace Engineering, Amirkabir University of Technology, Tehran 159163-4311, Iran
A R T I C L E I N F O A B S T R A C T
Keywords: The purpose of this paper is optimizing the performance parameters of a single-shaft micro gas turbine with a
Single-shaft micro gas turbine power of about 150 kW at the design point by modifying its centrifugal compressor geometry. According to the
Centrifugal compressor physics of this problem, a hybrid optimization approach (including modern and numerical optimizers) was
Optimization process
employed. The optimization was based on the numerical simulation of each geometry using a three-dimensional
Engine performance parameters
Genetic algorithm
Reynolds Averaged Navier-Stokes solver. The geometric parameters of the centrifugal compressor impeller
Computational fluid dynamics (CFD) constituted the design variables and the Latin hypercube was employed for sampling. To reduce the cost of the
numerical calculation, the neural network was chosen for approximate modeling of the design space. The physics
of the optimization problem ahead was identified and a state-of-the-art optimization process was presented
accordingly. Furthermore, a detailed discussion was conducted on the design space transparency, convergence
criteria, and optimum design selection, the disadvantages of current optimization processes and the advantages
of the present one were explained. NASA CC3 compressor was scaled down for the baseline engine, which is a
150 kW single-shaft gas turbine. This engine was optimized using the proposed optimization process and a
significant improvement was achieved, where efficiency increased by 13.93% and power by 11.07%, and the
specific fuel consumption decreased by 12.15%. Furthermore, despite the constant impeller tip diameter and
rotational speed, in the optimum compressor, the mass flow increased by 2.22% and the pressure ratio by
29.62% compared to the baseline compressor, while the isentropic efficiency not only did not decrease but also
increased slightly (0.2%). However, the operating range decreased by 12.64%.
they will be inefficient, computationally expensive, and find a local

1. Introduction optimum closest to the initial value. However, modern optimization
algorithms such as genetic algorithms (GAs) have been adapted to solve
Gas turbines are widely used in power plants, oil and gas trans such problems and have shown their ability to fully search the solution
mission pipelines as mechanical drives, and transportation applications space to find the global optimum with a high probability [11]. Modern
including aircraft, marine, and automotive [1,2]. Also, they are optimizers include genetic algorithms [12], simulated annealing [13],
employed in recent applications including combined cycle [3], com particle swarm optimization [14], ant colony optimization [15], and etc.
bined heat and power (CHP) [4] and combined cooling, heating, and The modern and numerical optimization methods have inherent
power (CCHP) units [5], organic Rankin cycles (ORC) [6], fuel cells [7], advantages and disadvantages. Using a hybrid approach, one can
innovative power plant cycles [8] and etc. In all of these applications, combine the ability of exploration of modern algorithm with the ability
gas turbines play an essential role. Therefore, the design optimization of of exploitation of the numerical techniques. In other words, the ability of
gas turbine engines to maximize the thermodynamic efficiency or power modern methods to search the entire solution space and find the global
generation or minimize the specific fuel consumption or emissions is optimum can be combined with the ability of numerical methods to
crucial and will have a deep impact on the energy economy and the reach the local optimum quickly and directly from an initial value
environment. The real engineering optimization problems, such as the [16,17]. This significantly reduces the optimization calculations.
design optimization of gas turbines, have many complexities and the Accurate aerodynamic performance of turbomachines is achieved
design space is often multimodal and highly non-linear [9,10]. using 3D numerical simulation. Also, modern optimizers require func
If numerical optimization techniques are used for these problems tion evaluation of a large number of samples [18]. Thus, direct
* Corresponding author.
E-mail address: tousi@aut.ac.ir (A.M. Tousi).
https://doi.org/10.1016/j.enconman.2022.116245
Received 21 June 2022; Received in revised form 12 September 2022; Accepted 13 September 2022
Available online 12 October 2022
0196-8904/© 2022 Published by Elsevier Ltd.
S.M. Hosseinimaab and A.M. Tousi Energy Conversion and Management 271 (2022) 116245
Nomenclature η efficiency
ηeng thermodynamic cycle efficiency of the single-shaft gas
b2 meridional depth of impeller outlet turbine engine
cp specific heat at constant pressure (J/(kg.K)) π pressure ratio
LHS Latin hypercube sampling θ angular cylindrical coordinate
m meridional coordinate
ṁ mass flow rate (kg/s) Subscripts
M meridional distance a air
Mʹ radius normalized meridional distance c compressor
Ma Mach number cc combustion chamber
r radial cylindrical coordinate d design
R radius dif diffuser
SFC specific fuel consumption int intake
t nondimensional meridional coordinate LE leading edge
T temperature (K) m mechanical
TIT turbine inlet temperature (K) rel relative
Wnet net power output (W) TE trailing edge
Y+ normalized wall distance t turbine
z axial cylindrical coordinate ted turbine exit duct
0 total
Greek Letter
β blade camber line angle (blade angle)
connection of modern optimizers and 3D numerical simulator is The combination of the modern optimizer and the approximate
computationally expensive, time-consuming, and inefficient [19]. model has been widely used in research [25,26,28–31], but the above
Therefore, to reduce the cost of calculations, methods of approximate points has been neglected in many of them.
modeling of the design space were employed. Approximate methods Optimization is a process, not a function, and the optimizer is only a
(also called Surrogates or metamodels) greatly reduce computational function in the optimization process. Therefore, researchers who have a
costs but instead provide an approximate design space. process approach to optimization have proposed algorithms and flow
There are a number of techniques that have been employed by re charts for the optimization process [25,26,31–33].
searchers to build approximate models, such as the Kriging [20], the There are three important points in these algorithms. First, the
response surface model [21], artificial neural network (ANN) [22,23], minimum number of samples needed to form a database (at the begin
support vector regression (SVR) [24], and etc. Among these, the neural ning of the optimization process) that can make the design space
network is widely used in various sciences due to its ability to global transparent. The articles did not usually mention the number of samples
modeling, high flexibility, good accuracy, and having sparks of artificial needed, nor did they provide a criterion for ensuring the transparency of
intelligence. the design space based on the number of samples used to form the
Approximate models require a database, and the quality of their database.
prediction depends on the richness of the information in the database Second, the following two convergence criteria can be defined for
[25]. As a result, enriching the database with a minimum number of optimization algorithms that use approximate methods:
samples is crucial. How to select the samples for the database plays a
major role in the richness of the database information. • Convergence of objective functions obtained from CFD analysis
The utilization of random sampling to create a database is very • Convergence of results obtained from the approximate model and
common due to the simplicity of its execution [26]. However, this CFD analysis
method is not efficient and increases the computational cost. In an
appropriate sampling method, in any set of samples, the maximum Almost all the articles reviewed so far have chosen the second cri
possible information must be extracted from the entire range of each terion for convergence. However, true convergence occurs when the
design variable. The idea of Latin hypercube sampling [27] is to ensure exact value of the objective function calculated by CFD methods is not
the full coverage of the range of each design variable. In this method, improved. In this case, the results obtained from the approximate model
there is proper control on the distribution of the value of design vari may not yet converge to the CFD results.
ables throughout the design space, which is not the case in randomly The third problem is that in the optimization process it is assumed
selecting the value of variables. that the optimum geometry obtained by the optimizer based on the
There are two important points in the optimization process. First: an predictions of the neural networks has the best performance.
optimizer such as a genetic algorithm, assuming the settings are In many studies, the purpose of optimization is to optimize the
accomplished correctly, finds the global optimum of the problem given performance of engine components, and little attention has been paid to
to it. The important point is whether this problem really is the intended optimizing the performance of the engine (as an end product) by
problem or an accurate estimate of it? In other words, is the performance modifying the design of its components. Verstraete et al. [25], Ekradi
of the approximate method, sampling method, etc. such that the design and Madadi [26], Ju et al. [28], Guo et al. [29], He and Zheng [30],
space is properly estimated so that the optimizer can use it to find the Khalfallah and Ghenaiet [31], Guo et al. [34], Moussavi et al. [35], and
optimum answer? Li et al. [36] for example, optimized a centrifugal compressor impeller to
Second: how to achieve an optimum and reliable design with a improve its performance parameters. This was done by modifying the
reasonable number of samples, given the characteristics of the problem geometric parameters of the impeller.
ahead, that even with increasing the number of samples, a better answer However, Wang et al. [37] proposed a system-component coupled
cannot be obtained. optimization method of the supercritical CO2 (S-CO2) power cycle,
2
which is implemented by genetic algorithm based on the one- procedures. This means that each time they are executed on a large-scale
dimensional model of the S-CO2 centrifugal compressor and S-CO2 r problem (similar to the present paper’s problem), a new answer is ob
adial turbine. Then, three-dimensional modeling and CFD simulation tained. Of course, in the genetic algorithm, repetitive answers may be
of optimum compressor and turbine were implemented. As they stated, obtained after several runs. The severity of this depends on the physics
the relative errors of one-dimensional design and three-dimensional of the problem. However, due to the stochastic nature of neural network
simulation results were within 5%. The use of a one-dimensional training, it is very unlikely to achieve repetitive networks.
model greatly reduces the computational cost, but instead the error is A state-of-the-art design optimization process should be able to
in the same order or even more than the performance improvement of extract a reliable, robust, and reproducible process from stochastic
the compressor and turbine, and therefore, its results will not be reliable. procedures in the optimizers and neural networks that can lead to the
As can be seen, the vacancy of a reliable and reproducible optimi optimum solution with the least computational cost, time, and need for
zation process that can be used to successfully optimize turbomachines the designer’s experience.
or gas turbine engines performance is quite obvious. Previous optimi The accuracy of numerical calculations has a profound effect on the
zation processes do not provide any criteria for whether the approximate results of the optimization process, therefore the results of numerical
model is a correct estimate of the design space. Also, they do not guar aerodynamic simulation of NASA CC3 compressor with two turbulence
antee finding the global optimum design via a reliable method with models k-ε and k-ω SST were validated against the test rig results. The
minimum samples. In addition, in the flowcharts provided for the simulation results with both turbulence models were very close to the
optimization processes, the design space will not be uniformly trans test rig ones, however, the results of the k-ε model were slightly better.
parent, which makes the approximate model unable to provide a correct In the following, the specifications of the baseline engine were pre
estimate of the solution space. Furthermore, the convergence criteria sented. The baseline engine was optimized using the proposed optimi
and the assumption that the optimum geometry obtained by the opti zation process. The specifications of the optimum engine were compared
mizer based on the predictions of the neural networks has the best with the baseline engine. A significant improvement in the engine per
performance are other deficiencies of the previous flowcharts. More formance has been achieved, in this way that efficiency increased by
over, optimization to improve engine performance as an end product has about 14% and power by about 11% and the specific fuel consumption
received less attention. (SFC) decreased by about 12%. In addition, the geometry and perfor
The purpose of this paper is to maximize the efficiency of the ther mance parameters of optimum and baseline compressors were
modynamic cycle of a single-shaft micro gas turbine engine at the design compared. This engine will be used in a combined heat and power unit.
point by modifying its centrifugal compressor geometry, so that the Furthermore, a detailed discussion was conducted on the design
engine output power is not less than 150 kW and the design rotational space transparency, convergence criteria, and optimum design selection,
speed and the impeller tip diameter do not change. As a result, there is a the disadvantages of the current optimization processes [25,26,31–33]
constraint single-objective optimization problem. were explained and the advantages of the present one were expressed.
Each component of the engine alone has an optimum operating Correct extraction of the design point from the numerical simulation
point. When these components are considered as a whole called the results plays an essential role in the accuracy and success of the proposed
engine, then the optimum point of engine performance is determined optimization process. Therefore, a procedure including instructions on
based on the interaction of the components with each other and with the how to numerically simulate along with how to fit the curve to the
aim of optimizing engine performance. simulation results was presented in order to obtain the exact location of
A state-of-the-art optimization process was presented according to the design point.
the optimization problem of a single-shaft gas turbine engine. In addi The main objectives and novelties of this paper include the
tion to the optimization process itself, its components should be selected following:
according to the problem characteristics in order to perform properly
and efficiently. • Presenting a state-of-the-art optimization process along with its de
To do this, a hybrid optimization approach was employed including tails, according to the optimization problem of a single-shaft gas
the genetic algorithm as a modern optimizer for exploring the solution turbine engine.
space and the sequential quadratic programming (SQP) technique as a • Optimization to improve engine performance parameters instead of
direct numerical optimization method to exploit the answer [11]. Also, compressor performance alone.
the optimization was based on the numerical simulation of the impel • Optimizing a single-shaft gas turbine engine and obtaining a signif
lers’ geometries using a three-dimensional Reynolds Averaged Navier- icant improvement in performance.
Stokes solver. The constituent geometrical parameters of the centrifu • Employing a hybrid optimization.
gal impeller were chosen as design variables. Furthermore, neural net • Process approach to optimization versus functional approach.
works for approximate modeling and the Latin hypercube method for
sampling were employed. 2. Optimization process
The new optimization process is reliable, robust, reproducible, and
has the least need for designer experience. Furthermore, it acquires In this section, the optimization problem was defined and the
reliable and practical optimum design with minimum computational configuration of the baseline gas turbine engine was presented. Also, the
cost. The present optimization process makes the design space uniformly design variables and sampling method were introduced. Then, details of
transparent and creates an approximate model that provides a true and numerical simulation, neural network specifications, and hybrid opti
reliable estimate of the design space. In other words, this method is a mizer were presented. Finally, the flowchart of the optimization process
compromise between the minimum number of samples for transparency was presented (Fig. 5), which shows how closely the above items work
of the design space on the one hand and obtaining a reliable optimum cooperatively through the optimization process.
solution on the other hand. Moreover, for the new process, a new cri
terion for convergence of the optimization process was developed, and 2.1. Definition of the optimization problem
the best geometry was determined after the aerodynamic simulation of
the geometries found by the optimizer based on the predictions of the The optimization problem ahead is to optimize the performance
neural networks. parameters of a micro single-shaft gas turbine engine (Fig. 1) with an
As mentioned above, the new optimization process is based on nu output power of about 150 kW at the design point by modifying the
merical simulation, genetic algorithm, and neural networks. Among compressor geometry. This engine will be used in a CHP unit.
these, genetic algorithms and neural networks work based on stochastic Accordingly, the optimization problem is to maximize the efficiency
3
the impeller is formed. By parameterizing the above items and using all
or some of them as design variables, the optimization process can be
performed on a centrifugal impeller.
In the present paper, the distribution of camber lines, the position of
the leading edge and the pitch position of splitter, and the meridional
depth of the impeller outlet are considered as design variables, and the
hub and shroud curves in the meridional plane, thickness distribution,
and the number of blades are the same as the baseline geometry. Also,
the splitter camber and its thickness distribution are similar to the main
blade.
The blade camber lines at the hub and shroud can be defined by the
distribution of β(t), the angle between the meridional plane and the
blade camber (Fig. 2). The β distribution is usually plotted versus the
meridional distance along the hub and shroud curves, M. Here this
distribution is defined by fourth-order Bezier polynomials (Eq. (6)) with
five control points as follows [39]:
β(t) = β0 (1 − t)4 + 4β1 t(1 − t)3 + 6β2 t2 (1 − t)2 + 4β3 t3 (1 − t) + β4 t4 (6)

Fig. 1. Design stations of the single-shaft engine.
where t is the nondimensional meridional coordinatet ∈ [0, 1], (0 at the
of the thermodynamic cycle or to minimize the SFC of a single-shaft gas leading edge and 1 at the trailing edge) and β0 to β4 are the control
turbine engine so that the engine output power is not less than 150 kW points on the M-β plane. β0 and β4 are the blade angles at the leading and
and the design rotational speed and the impeller tip diameter do not trailing edges, respectively. This definition is used for both the hub and
change. Therefore, we are encountered with a constraint single- shroud. Also, M is defined as follows [40]:
objective optimization problem. √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∫ t0 ( ) ( )̅
According to the thermodynamic cycle of the single-shaft gas turbine
∂z(t) 2 ∂r(t) 2
M= + dt (7)
engines, maximizing the cycle efficiency or minimizing the SFC will 0 ∂t ∂t
have the same result [1]. Here the efficiency of the thermodynamic cycle
where r(t) and z(t) are the radial and axial coordinates on the meridional
is chosen as the objective function, i.e.
plane, respectively, andt0 ∈ [0, 1]. It is very common in the turboma
Objective Function = ηeng (1) chinery literature to use radius normalized meridional distance, Mʹ,
which is defined as follows [40]:
According to the numbering of the design stations in Fig. 1, we have.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
Wnet ∫ t0 (∂z(t)/∂t)2 + (∂r(t)/∂t)2
(8)
′
ηeng = ∫ T04 M = dt
0 r(t)
ṁ cp (T)dT
T03 (2) In the present paper, the distribution of β versus Mʹ is considered,
∫ T04 ∫ T03
1
Wnet = ṁ cp (T)dT − ṁa cp (T)dT
T05 ηm T02
In above equation, the air is considered as an ideal gas and the

temperature dependence of cp of air is considered as below form [38]:
Cp0 (T)/R = a1 T − 2 + a2 T − 1 + a3 + a4 T + a5 T 2 + a6 T 3 + a7 T 4 (3)
where C0p (T) is the molar heat capacity at constant pressure. Finally, the
optimization problem will be as follows:
Maximize f (X) = ηeng
subject to (4)
g(X)⩽0,
where.
X(l) ⩽X⩽X(u)
(5)
g(X) = Wnet,baseline − Wnet
In the above equations X is the vector of design variables and X(l) and
(u)
X denote the lower and upper bounds on X, respectively.
2.2. Design variables
The main blade of the three-dimensional centrifugal impeller is

defined by the hub and shroud curves on the meridional plane, the blade
camber lines in the hub and shroud, and the thickness distribution
perpendicular to the camber line. The definition of the splitter is the
same if it exists. Camber lines and thickness distribution of the splitter
can be similar to the main blade or defined separately.
Knowing the number of blades and the position of the splitter pitch, Fig. 2. Geometric definition of the blade camber line by β angle.
4
which is similar to Eq. (6). The circumferential position of the camber matched with one of the nLHS values on x2 which is selected at random
line, θ, (Fig. 2) is also defined as follows [40]: and without replacement, and so on up to xnx (nx =No. of design vari
ables). These collectively constitute the first sample, X1. A similar pro
Rdθ = dmtanβ (9)
cedure is followed for X2,…,XnLHS , which results in a Latin hypercube
The coordinates of the control points of Bezier curves are selected as sample [41].
the design variables that can be modified by the optimizer to optimize
Xk = [xk1 , xk2 , ..., xknx ], k = 1, ..., nLHS , (10)
the objective functions. The possible variations of control points of the
Bezier curve (red points) are shown by the arrow in Fig. 3. The control
points at the leading edge of the hub and shroud in all geometries are 2.4. Numerical simulation
equal to the βLE of the hub and shroud of baseline geometry, respec
tively. As a result, β-angle distribution in the hub and shroud is defined In the present paper, an in-house built CFD code was used to
by the backswept angle, βTE, and 6 other parameters (Fig. 3). In other numerically simulate the aerodynamic performance of the centrifugal
words, the β-angle distribution for hub and shroud will consist of 7 compressors. This code has been developed in the Dana Laboratory of
design variables. This results in 14 design variables for the blade camber the Faculty of Aerospace of the Amirkabir University of Technology.
line definition. This code solves the three-dimensional, steady-state, and compressible
The range of design variables of β distribution was determined so Reynolds-averaged Navier-Stokes (RANS) equations using the finite
that even in the worst case would not be negative. In other words, the volume method.
design variables are defined in such a way that all the geometries created The k-ε and k-ω SST (shear stress transport) models were utilized for
by the optimizer or sampling method are feasible. turbulence modeling. For any turbulence models, an appropriate indi
Meridional depth of impeller outlet, b2, is another design variable. vidual grid structure was produced, so that, in the SST model Y+ is less
Proportion to the change in outlet width, the shroud meridional curve is than 1 everywhere and for the k-ε model about 30 [42,43]. This makes it
modified. The streamwise position of the splitter blade leading edge at possible to take advantage of each model.
hub and shroud are also two design variables. These are defined as a At the inlet boundary, the standard total pressure and temperature
percentage of the hub and shroud curves on the meridional plane and were applied. Also, the inlet flow was completely axial. At the outlet
can vary between 25% and 35%. The splitter pitch position was also boundary, an average static pressure was imposed. The impeller and
selected as a design variable. As a result, the total number of design casing walls were assumed as no-slip and adiabatic boundary conditions.
variables will be 18. The periodic boundary conditions were imposed along the circumfer
ential direction. The boundary conditions are shown in Fig. 4. Air as the
working fluid was assumed to be an ideal gas. The identical boundary
2.3. Sampling method conditions were used in all geometries in the optimization process.
Numerical simulation convergence was investigated using physical
Metamodels should be able to make their predictions as closely as parameters. When the mass flow rate at the inlet and outlet boundaries
possible to the true values in all parts of the design space. The accuracy of the solution field was fixed, then convergence is achieved. The cal
of the metamodels predictions depends on the quality and quantity of culations were performed with an Intel I Core I i7 PC.
information in the database. As the quantity of information increases,
the database will be richer; however, the quality of the information plays
a significant role. Perhaps, high-quality information with low quantity
can create a richer database than low quality with high quantity infor
mation. The quality of information depends on the sampling method.
The Latin hypercube sampling (LHS) [27] is an efficient sampling
method that provides full coverage of the entire range of design vari
ables. In LHS the range of each variable (i.e., the xi) is divided into nLHS
intervals of equal probability, and one value is selected at random from
each interval.
Now, one of the nLHS values on x1 is selected at random and without
replacement (the probability of selection of each values is equal), and is
Fig. 3. Definition of the β-angle distribution by the Bezier curve. Fig. 4. Boundary conditions for the centrifugal compressor.
5
2.5. Approximate model The specifications of the CC3 impeller with the vaneless diffuser in the
design point are given in Table 1.
According to the above description, in this paper, artificial neural Numerical simulations were performed to predict the aerodynamic
network was employed for approximate modeling which was a multi performance of the CC3 impeller with a vaneless diffuser. Both k-ω SST
layer feed-forward perceptron network. The most important feature of a and k-ε models were employed to simulate turbulence. As mentioned,
multilayer feed-forward network is that it is able to learn any type of the grid structure produced for the SST model is such that Y+ is less than
mapping with any degree of complexity [44]. The weights of this NN are 1 everywhere and about 30 for the k-ε model [42,43].
gradually modified during training so that the NN will be able to classify The numerical simulation results are compared with the test rig re
the samples correctly. sults for the design speed in Fig. 6. Simulations with both turbulence
Training begins with providing samples to the neural network. The models give an acceptable prediction of the performance of the CC3
output errors are then calculated (root mean square errors) and are impeller with a vaneless diffuser. However, both models predicted
propagated backward through the hidden layers (errors back- slightly lower chocking mass flow rates.
propagation). Finally, the network weights are corrected based on the In Fig. 6 in the mass flow versus pressure ratio diagram, the simu
back-propagated errors. To implement this process, there are various lation results of the k-ε model are closer to the test rig results, while in
algorithms. In this paper, the generalized delta rule was used. the efficiency diagram, the SST model has better performance. The
Three individual artificial neural networks are trained to predict the simulation and the test rig results at the design point are compared in
values of centrifugal compressors’ performance parameters include Table 2. Since the meshing network of the k-ε model is almost half of the
mass flow rate, pressure ratio, and isentropic efficiency. This splitting SST model and the convergence of the solution occurs faster, so in this
into dedicated ANNs enhances the accuracy of predictions. paper, the k-ε model was used in numerical simulations.
A grid study was performed to ensure that the results are indepen
2.6. Optimizer dent of the computational grid. To do this, 5 grids were produced for the
impeller and the performance characteristics obtained from them
The optimization problem ahead is a nonlinear and constrained including mass flow rate, pressure ratio, and isentropic efficiency were
problem with a large number of local optimums. Also, it is not possible compared in Fig. 7. In this figure, the results of all grids were normalized
to explicitly define objective functions, design functions, and constraints by the results of the fifth grid and the difference between them and the
based on the design variables. fifth grid was shown. As can be seen, the difference between the pressure
Therefore, a hybrid optimization approach was employed. The ge ratio and the efficiency of the fourth and fifth grids is less than 0.04%
netic algorithm was selected as a modern optimizer for exploring the and for the mass flow rate is less than 0.15%. Therefore, it can be
solution space and the SQP technique was selected as a direct numerical concluded that from the fourth grid onwards, the dependency of the
optimization method to exploit the answer. In the hybrid process, the results on the grid resolution is negligible. A similar process was per
optimization was initiated by the genetic algorithm and a large formed to study the independence of the vaneless diffuser grid (results
convergence tolerance was chosen for it. are not presented here). Finally, a grid with about 1 million nodes was
Due to the high computational cost of modern methods, choosing a selected for the impeller (fourth grid) and 900,000 nodes for the vane
large convergence criterion results in rapid convergence and a signifi less diffuser.
cant reduction in computational cost. When the solution converged, the
answer was given as the initial value to the numerical optimization
3.2. Baseline gas turbine engine
methods to achieve the optimum value with a minimum evaluation of
the objective function. As mentioned, this significantly reduces the
The baseline engine is a single-shaft micro gas turbine (Fig. 1) that
optimization calculations.
produces about 150 kW of power and 0.9 kg/s of hot gas at a temper
The idea of using hybrid optimization has been formed according to
ature of 625˚C in a CHP unit. The compressor impeller of this engine is a
the physics of the real optimization problems that exist in research and
scaled-down of CC3 impeller. The performance specifications of other
in industries, the problems with large dimensions, solution space with
engine components are determined according to similar engines and are
multiple local optimums, and sometimes discrete. As mentioned in the
given in Table 3. Also, the performance specifications of the baseline
introduction section, if numerical optimization techniques are used for
these problems they will be inefficient, computationally expensive, and
find a local optimum closest to the initial value. Therefore, only modern Table 1
optimizers can be used. Design Point Specifications of CC3 impeller with vaneless diffuser [45].
Although modern optimizers have the ability to solve these prob Parameters Values
lems, at the same time, they have disadvantages that are tried to be Shaft speed 21,789 (rpm)
covered by numerical optimizers in hybrid optimization. In all problems Design mass flow rate 4.54(10) kg/s (lbm/s)
where modern optimizers can be used, the proposed hybrid optimizer Blade count full/splitter 15/15
Inlet blade height 64 (mm)
can also be used, because there is no special condition for its application.
Inlet tip diameter 210 (mm)
This hybridization is exactly based on the physics of real optimization Relative Ma @ inlet tip 0.9
problems. Exit diameter 431 (mm)
Meridional depth of Impeller 17 (mm)
3. Results and discussion outlet (b2)
Blade angle @ TE (backswept 50
angle)
3.1. Validation of the numerical simulation Exit tip speed 492 (m/s)
Splitter LE location %30 of the hub and shroud curves on the
In order to validate the numerical simulation of the aerodynamic meridional plane
Inlet pressure 101,325 (Pa)
performance of the centrifugal compressors, the NASA CC3 compressor
Inlet temperature 288.15 (K)
was chosen. The NASA CC3 compressor is originally an Allison Engine Measurement station radius ratio 1.18
Company design with the size of 1.66 kg/s (3.655lbm/s) that was scaled Pressure ratio @ measurement 4.175
up to a mass flow rate of 4.54 kg/s (l0 lbm/s). The performance map of station
the CC3 stage and impeller with vaneless diffuser obtained by test rig are Efficiency @ measurement 0.867
station
presented in [45] and impeller and diffuser geometries are given in [46].
6
Fig. 5. Flowchart of the optimization process.
engine are presented in Table 3. (especially channel diffusers) are commonly employed in stages with a
As shown in Table 3, a mass flow rate of 0.88 kg/s is required to reach pressure ratio of approximately 3 or 4 and greater [47]. The CC3 stage
150 kW power. Therefore, the CC3 impeller was scaled with a coefficient pressure ratio is about 4 and, as expected, has a channel diffuser.
of (0.88/4.54) 0.5. The results of the performance of the scaled impeller Second, modifying the compressor stage, i.e. impeller and diffuser,
with a vaneless diffuser, i.e. the baseline compressor, were obtained via requires considering the diffuser design variables in addition to the
CFD simulation and used in the cycle analysis. These results are shown in impeller. But this increases the number of design variables and conse
Table 3. quently increases the computations. In addition, it complicates the
There are two points. First, it should be noted that vanned diffuser optimization problem and makes it difficult to draw conclusions about
7
Table 3
Assumptions used in the analysis of the baseline engine thermodynamic cycle
and its results.
Parameter Value Parameter Value
TIT(K) 1188 ΔP0cc 0.07

πc 4.31 ΔP0ted 0.02
ṁa (kg/s) 0.88 ηdif 0.99
ΔP0int 0.0 ηm 0.99
ηc 0.9043 ηeng 0.2092
ηt 0.85 SFC(kgh/kW) 0.4156
ηcc 0.96 Wnet (kW) 152.07
compressor and a vaneless diffuser and the geometrical parameters of

the impeller constitute the design variables. This has no effect on the
optimization process and only reduces the computational cost.
3.3. Correct extraction of the design point from numerical simulation

results
In this section, a procedure for the correct extraction of the design

point from the numerical simulation results is presented. This plays an
essential role in the accuracy and success of the proposed optimization
process. To determine the design point in any geometry, it is necessary
to obtain its characteristics in the design speed. To do this, each ge
ometry in the design speed was simulated at 4 to 6 points. Then, a curve
Fig. 6. Comparison of the numerical simulation results with the performance that is in accordance with the physics of the characteristics of centrifugal
map of the NASA CC3 impeller with a vaneless diffuser. compressors was fitted on the simulated points. Subsequently, the exact
location of the design point and its specifications were extracted and
provided to the neural network. This procedure is described in more
Table 2 detail below.
Comparison of the numerical simulation and test rig results at the design point The performance map of the baseline compressor at the design speed
(ṁ = 10 (lbm/s)). along with the design point is shown in Fig. 8. The choking and the
ηc πc design point mass flow rate on the design speed are 0.9667 kg/s and
⃒ ⃒
Original 0.8667 4.175
⃒ ⃒
0.88 kg/s, respectively, thereforeṁa ⃒⃒ /ṁa ⃒⃒ = 0.88/0.9667 =
SST 0.8766 3.9856 d choked
difference (%) 1.14 − 4.54 0.9103. This ratio was used to find the mass flow rate of the design point
k-ε 0.8837 4.0966
in new geometries.
difference (%) 1.96 − 1.88
Applying a low static pressure at the outlet domain, the choking mass
flow is determined in each geometry. Knowing the design/choked mass
flow ratio, the design mass flow rate is obtained.
Then, the characteristic of the new geometry is obtained by
increasing the static pressure at the outlet at several other points, so that
the design point must be located between the obtained points. Subse
quently, using the Piecewise Cubic Hermite Interpolating Polynomial
Fig. 7. Grid independency study.
the results obtained.

Considering the above two points and without losing the generality
in the proposed optimization process, it is assumed that the compressor
stage of the baseline engine includes the scaled impeller of the CC3 Fig. 8. The performance map of the baseline compressor at the design speed
along with the design point.
8
(PCHIP) method [48,49], a curve is fitted on these points. Knowing the divided by their standard deviation.
design mass flow rate, the pressure ratio, and efficiency are also As mentioned, in this paper, hybrid optimization was used. The
obtained. settings related to the genetic algorithm in MATLAB are listed in Table 4.
Based on the authors’ experience, the results of this method of curve As can be seen from this table, the convergence tolerance of 10-4 was
fitting are closer to the physics of characteristics. For each geometry, chosen for the genetic algorithm. When the genetic algorithm
between 4 and 6 points are required to obtain the characteristics. One converged, the solution was given as the initial value to the SQP (or
point is related to the choking condition. Two very important points are fmincon in MATLAB), a numerical optimization method to quickly
related to the region of change of the characteristics diagram from obtain the exact location of the global optimum point with a minimum
vertical to relatively horizontal. The rest of the points are related to the evaluation of the objective function. The convergence tolerance of SQP
relatively horizontal part of the characteristics. was 10-6.
One of the disadvantages of modern optimizers is their high
computational cost. This is the most important factor for using hybrid
3.4. Fitting Bezier curve on β-angle distribution of the baseline geometry
optimization in this paper. So that, in the present problem, when the
modern optimizer (i.e. genetic algorithm) was used alone, the optimum
First, we fitted the 4th order Bezier curves, whose coordinates of the
solution was obtained after about 80,000 calls of the design function.
control points were considered as design variables, on the β-angle dis
Here almost 75% of the calls were to achieve tolerances higher than 10-4
tribution curve of the baseline geometry. The result is shown in Fig. 9.
(i.e. 10-6). This is while, when the solution obtained from the modern
The root mean square (RMS) error of fitting the Bezier curve on the hub
optimizer with a tolerance of 10− 4 was given to the numerical optimizer
is 0.0084 and on the shroud is 0.0115. It means that the 4th order Bezier
as an initial value, the optimum solution was obtained with only a few
curve can well express the β-angle distribution in the baseline geometry.
hundred calls of the design function. In other words, the hybridization of
Also, the range of design variables was adjusted so that the values of the
modern and numerical optimizers reduced the total number of calls of
design variables (which vary between 0 and 1) were 0.5 for the baseline
the design function to less than 20,000, i.e. just under 25%.
geometry (Table 8). Of course, except for the backswept angle in the hub
and shroud, i.e. variables 1 and 9.
3.5.1. What happens in optimization loops?
In the first loop of the optimization process, 4 sets of 18 samples were
3.5. Introducing the optimization process prepared. After numerical simulation, the results of their aerodynamic
performance were obtained. Then the values of their design variables as
Initially, a set of 18 samples was selected as test samples. These input and the results of their aerodynamic performance as output were
samples were then numerically simulated and their aerodynamic per given to the three neural networks. Each neural network was trained 30
formance results were obtained. As mentioned earlier, the values of the times. After training, the accuracy of each network in predicting the test
design variables for the sample sets were selected using the Latin hy samples was calculated. The network with the least errors was selected
percube method. (Fig. 5).
Three feed-forward perceptron neural networks consisting of one Then, 30 times optimization was performed by genetic algorithm
hidden layer with 4 × 18 + 1 = 73 neurons, and tangent sigmoid as using selected neural networks. In these 30 times optimizations, a
neuron activation function were prepared for mass flow, pressure ratio, number of the results were the same, and the optimizer found the same
and isentropic efficiency. The examinations of the authors of the present optimum points. Among the results, the geometry with the highest en
paper showed that increasing the number of hidden layers or the number gine efficiency was selected as the optimum geometry. The black dia
of neurons in the layers have no benefit to improving the performance of gram in Fig. 10 shows the optimum engine efficiency in each
neural networks and not only increases the computational cost but also optimization loop, which is obtained based on the prediction of neural
in some cases decreases the neural network performance. The neural networks. In addition, the optimum geometry in each loop is numeri
network inputs, which are the design variables, vary between 0 and 1. cally simulated and the engine efficiency obtained based on it is shown
To speed up the network training, neural network outputs including by the blue diagram in Fig. 10. In the next loops, 5 sets of samples were
mass flow, pressure ratio, and isentropic efficiency were normalized. For added to the previous sets each time.
this, the average of each parameter for the samples of that optimization
loop was calculated and subtracted from all values of that parameter. 3.5.2. A new flowchart for the optimization process
Then the values of all parameters (after subtracting the average) were Because of high computational cost in this problem, integrating an
approximate model with the optimizer is inevitable. On the other hand,
finding a reliable method or process that converges to the correct answer
using an approximate model is of great importance. In other words,
using an approximate model to reduce the computational cost should
not lead to a reduction in the quality of the results. As mentioned before,
the main issue is how to achieve an optimum and reliable design with a
reasonable number of samples, that even with increasing the number of
samples, a better answer cannot be obtained.
Researchers have developed a number of algorithms for the
Table 4
Details of the genetic algorithm.
Criteria Function/value
Population size 75
Number of generations No limit
Convergence criteria 10-4
Fitness scaling Rank
Selection Stochastic uniform
Mutation Constraint dependent
Fig. 9. Fitting the 4th order Bezier curve on β-angle distribution of base
Crossover Constraint dependent
line geometry.
9
while the actual issue is the convergence of the exact solutions (results of
numerical simulations) to an optimum geometry. This will be explained
further below.
The third problem is that in the optimization process it is assumed
that the optimum geometry obtained by the optimizer based on the
predictions of the neural networks has the best performance.
In this paper, the following proceedings are provided to resolve the
optimization processes problems mentioned above. To address the issue
of transparency in optimization flowchart of Verstraete et al. [25], we
made the entire design space transparent. For this purpose, in each
optimization loop, if there is no convergence, 5 sets of samples are added
to the database, so that these samples are selected by the Latin hyper
cube sampling from all the solution space.
By resolving the lack of transparency in the optimization flowchart of
Verstraete et al. [25], it is now possible to compare the performance of
the flowchart presented in this paper with it. The black and blue dia
grams in Fig. 10 are based on the Verstraete et al. [25] convergence
criterion, which states that the difference between the neural networks
prediction and the numerical simulation must ultimately satisfy the
Fig. 10. Convergence diagram of the optimization process. convergence criterion. By comparing the black and blue diagrams, it is
clear that despite the use of a large number of samples (34 sets of 18
optimization process that are almost identical. For example, Verstraete samples, i.e. 612 samples), there is still a large difference (3.3%) be
et al. [25] developed a flowchart (which was employed by many re tween the results obtained from the neural network and the numerical
searchers such as [26,29,50,51]) that greatly reduces the computational simulation. Furthermore, with increasing the number of samples, the
cost and the optimization process converges rapidly. In this flowchart, convergence process is too slow, especially after 14 sets of samples, and
about twice the number of design variables, sample geometries are to achieve convergence, large amounts of samples are needed, which
produced (23 design variables and 55 samples), and after numerical leads to an increase in computational cost.
simulations, their aerodynamic performance results are obtained and a Fig. 11 shows the convergence trend of neural network predictions.
database is created with them. Relevant neural networks are trained Mass flow and efficiency neural networks predict test samples with good
through this database. The genetic algorithm then finds the optimum accuracy from the beginning (less than 1% error in 4 sets of samples) and
geometry using these neural networks. This optimum geometry is this accuracy increases as the number of sets of samples increases (about
numerically simulated and its performance results are compared with 0.1% error in 14 sets). This process is initially fast and slows down as the
the results predicted by neural networks. If the difference between the number of sets of samples increases. This trend is also repeated for the
results is less than the convergence criterion, the optimization process is pressure ratio network, with the difference that the error is initially
converged. Otherwise, data is added to the database, the neural net about 10% and decreases rapidly to 3% as the sample sets increase to 14.
works are retrained and new optimizations are performed. This process From here, the trend of reduction of the prediction error is too slow and
continues until the convergence criterion is met. Verstraete et al. [25] a large number of extra samples are needed to achieve the error rate
stopped the optimization process after 50 iterations. In other words, in below 1. This will be inefficient and computationally expensive and
their flowchart, the optimization process converged by simulating a time-consuming (on the scale of the calculations of this paper). On the
total of 105 geometries for a problem with 23 design variables. other hand, it remains to be seen, given the physics of the present
At first glance, this flowchart is an ingenious method for rapid optimization problem, to what extent the accuracy of the neural network
convergence of the optimization process and achieves the optimum so predictions will affect the final results.
lution with a minimum number of samples. However, this method has Unlike the flowchart of Verstraete et al. [25] which assumes that the
major problems that make its use questionable. optimal geometry obtained by the optimizer based on the predictions of
The first problem is that in this process only a portion of the solution
space becomes transparent. Selecting samples only about twice the
number of design variables for the initial database provides a non-
transparent solution space in which the optimizer recognizes a point
as the optimum, and in the next loops, only the area around that point
becomes transparent. This is while the global optimum point may be
elsewhere in the solution space! So long as the entire solution space is
not uniformly transparent and the peaks and valleys are not recognized,
the optimizer cannot find the overall optimum point. This becomes even
more important when the solution space has a large number of local
optimums, and as the space becomes more transparent, new local opti
mums appear. Some of these local optimums may eventually be selected
as the global optimum.
The second problem is related to the convergence criterion of the
optimization process. Verstraete et al. [25] chose the equality of the
results predicted by the neural networks and the numerical simulations
as the convergence criterion of the optimization process. However, the
results obtained by numerical solutions may be converged, while the
neural network still does not accurately predict the values of peaks and
valleys. In their flowchart, the optimization process tries to ensure that
the neural networks can correctly determine the values of the peaks,
Fig. 11. Convergence trend of neural networks prediction.
10
the neural networks has the best performance, here all unique geome has a large number of local optimums. The optimization process must be
tries in 30 times optimization in each loop were numerically simulated, adapted to the optimization problem in such a way that it can get the
and the geometry with the maximum efficiency was chosen. These re right answers. Simply running an optimizer in this problem without
sults are shown by the red diagram in Fig. 10. This diagram shows well supervising its operation will not lead us to the right answers.
the convergence of optimization process, so that the difference between It means that in this problem there is not a global optimum whose
the results of 34 and 14 sets of samples is less than 0.3% (less than 0.07 objective function is significantly higher than other optimum points.
increases in efficiency). The green diagram in Fig. 10 is the neural Therefore, achieving the best optimum point (or the best answer) that is
network predictions corresponding to the red diagram. The difference acceptable in terms of engineering or application should be considered.
between this diagram and the numerical simulation results decreases As the number of sets of samples increases, the design space becomes a
more rapidly than the black diagram, however, its convergence still little more transparent and more precise each time. This leads to the
requires a large number of additional samples. finding of a number of new local optimum points. In the following, we
This indicates that the optimizer correctly identifies the location of encounter a situation where the objective functions of the optimum
the local optimums, but the estimations of the neural networks for the points are equal to the previous optimum points or their improvement is
values of peaks (in the present maximization problem) are not accurate. so small that it is not reasonable to spend computational cost and time
Although, as the number of samples for neural networks training in on them. As mentioned, the difference between the objective functions
creases, the accuracy of estimating the values of peaks increases, how of the optimal points in 34 sets and 14 sets of samples is less than 0.3%,
ever, this process is too slow (Fig. 10 and Fig. 11). while from 14 to 34 sets 360 samples are generated and numerically
Simply paying attention to reducing the difference between the re simulated.
sults predicted by the neural networks and the numerical simulation At the beginning of the optimization process, the physics and the
results to achieve convergence of the optimization process can be design space of the problem ahead were not known. During the opti
misleading. Because the geometries produced by the optimization pro mization process, the physics of the problem and the design space were
cess based on the numerical simulation have very close performance identified, and as the optimization process progressed, the optimization
(red diagram in Fig. 10) while their performance predicted by the neural process adapted to the problem.
networks are far from the numerical simulation (green diagram in
Fig. 10). As mentioned, neural networks require a large number of 3.5.3. Introducing the optimum engine and compressor
samples to accurately estimate the value of peaks. However, due to the The best 7th loop engine was selected as the optimum engine.
precise estimation of the location of the peaks by the existing neural However, as mentioned, according to the results of Fig. 10, the perfor
networks, it is possible to obtain the overall optimality without the need mance of the optimum engines from the 4th to the 7th loop is almost the
for extra samples by intelligently using numerical simulations. same. Table 6 compares the performance parameters of the optimum
During the optimization process, we encountered a phenomenon that and baseline engines. As can be seen from this table, a significant
in an optimization loop in the design space there were optimum points improvement in engine performance has been achieved.
whose objective functions were equal to or very close to the overall Table 7 compares the performance parameters of optimum and
optimum point, while their locations were different from the overall baseline compressors. Despite the constant impeller tip diameter and
optimum location. This is shown in Table 5 for the 4th optimization rotational speed, in the optimum compressor, the mass flow increased by
loop. Furthermore, in the 7th loop, an optimum point was obtained about 2.2% and the pressure ratio by about 30% compared to the
whose objective function was approximately equal to the overall opti baseline compressor, while the isentropic efficiency not only did not
mum of the 4th loop, but still, its location was different in the design decrease but also increased slightly (0.2%).
space (Table 5). Fig. 12 shows the geometry of the optimum and baseline impellers.
This is due to the physics of the optimization problem ahead, which As mentioned, the β-angle at the leading edge in the hub and the shroud
is equal to the baseline in all geometries, including optimum geometry.
Therefore, as can be seen from this figure, the inducers are the same in
Table 5 both geometries. Also, the number and thickness distribution of the
Comparing the values of design variables for three geometries with the almost main blades and splitters are the same. Due to the fewer backswept in
identical objective function. the optimum geometry, the curvature of the blades is clearly less in this
No. of Samples sets geometry. Moreover, in the optimum geometry, the meridional depth of
19 19 34
the impeller outlet is less. The leading edge of the splitters in the hub and
shroud have moved towards the trailing edge and the leading edge of the
ṁa (kg/s) 0.8927 0.8917 0.8845
main blades, respectively. Furthermore, the splitters have relatively
πc 5.5664 5.5847 5.5904
ηc 90.6805 90.5230 90.4838 approached the main blades from their suction side. More details were
ηeng 0.238031 0.238024 0.238042 given in Table 8 and Fig. 14. Also, the trailing edge in the baseline ge
Design Variable’s Number ometry has slightly leaned towards the pressure surface compared to the
1 0.0000 0.0000 0.0000 optimum geometry.
2 1.0000 1.0000 1.0000
3 0.0000 0.6582 0.3345
Table 8 presents the values of the design variables for the optimum
4 0.4271 1.0000 0.2044 and baseline geometries. As can be seen, in optimum geometry, the
5 0.6239 0.7665 0.2327 seven design variables have reached the lower (0) or upper (1) band of
6 0.7137 0.0000 1.0000 design variables range. This means that if the range was wider, these
7 0.0000 0.2493 0.0000
variables would continue to move towards the end of the range. How
8 0.0000 0.0000 0.0000
9 0.0000 0.0000 0.0000 ever, due to physical and geometric limitations, it was not possible to
10 1.0000 0.5651 1.0000
11 0.5340 1.0000 1.0000
12 0.2565 0.8783 1.0000 Table 6
13 0.9402 0.2234 1.0000 Comparison of performance parameters of optimum and baseline engines.
14 0.0000 1.0000 0.5461
Parameter Baseline Engine Optimized Engine Difference (%)
15 0.0583 0.4882 0.2610
16 0.5449 0.9947 0.9929 ηeng 0.209214 0.238353 13.9278
17 1.0000 1.0000 1.0000 SFC (kg/kW.h) 0.415611 0.365127 − 12.1469
18 1.0000 0.4505 0.0000 Wnet (kW) 152.07 168.90 11.0673
11
Table 7
Comparison of performance parameters of optimum and baseline compressors.
Parameter Baseline Compressor Optimized Compressor Difference (%)
ṁc (kg/s) 0.88 0.89957 2.2239

πc 4.31427 5.5922 29.6210
ηc 90.4323 90.61 0.1965
Table 8
The values of the design variables for the optimum and baseline geometries.
Design Variable’s Number Baseline Design Optimum Design
1 0.8232 0
2 0.5 1
3 0.5 0.2854
4 0.5 1
5 0.5 0.3459
6 0.5 0.0969
7 0.5 0.3883
Fig. 13. Comparison of the β-angle distribution of the optimum and base
8 0.5 0.4852
9 0.6092 0
line impellers.
10 0.5 0.4227
11 0.5 0
12 0.5 0.5377
13 0.5 0.4931
14 0.5 0.9322
15 0.5 0.2129
16 0.5 0.787
17 0.5 1
18 0.5 0
Fig. 14. Comparison of the meridional curves and the position of the leading
edge of the splitters of the optimum and baseline impellers.
[25] or Ekradi and Madadi [26], eliminate unacceptable geometries or

put low values instead of their objective functions, for example in the
case of a maximization problem. This causes the neural networks that
deals with a continuous design space to encounter some discontinuities.
These discontinuities disrupt neural networks learning and cause net
works training to be accompanied by many errors. To solve this prob
lem, the range of design variables should be determined so that even in
the worst conditions, generated geometries are acceptable. This was
done in this paper.
In Fig. 13, the β-angle distribution of the optimum and baseline
impellers in the hub and the shroud are compared, respectively. Also,
Fig. 12. Geometry of top: baseline impeller, bottom: optimum impeller.
the control points of the related fourth order Bezier curves are shown. As
mentioned, the β-angle at the leading edge in the hub and the shroud is
extend the range of some of these design variables.
equal to the baseline in all geometries. In the optimum impeller hub, the
For example, the backswept angle range was considered to be be
backswept angle is 25˚. Also, the β-angle variation in the shroud
tween 25˚and 55˚and normalized to 0 to 1 (design variable No. 1 for the
isβTE |Shroud = βTE |Hub ± 3◦ . So as the design variable of the backswept
hub and No. 9 for the shroud). This range is very wide and makes it
angle in the shroud (variable number 9) is 0 in the optimum geometry,
difficult to parameterize the β-angle distribution because unacceptable
therefore, the backswept angle in the shroud will be 22˚. In other words,
geometries are produced. Some researchers, such as Verstraete et al.
12
both variables related to the backswept angles have reached the lower at the design point is about 0.2% higher.
band of their range. Fig. 16 shows the contours of the relative Mach number at the exit of
Fig. 14 compares the meridional curves and the position of the the impeller of the optimum and baseline compressors. In the impellers
leading edge of the splitters of the optimum and baseline impellers. The with a higher backswept angle such as the baseline impeller, the cross-
optimum impeller meridian curve in the hub is similar to the baseline sectional area perpendicular to the flow is smaller (Fig. 12) and the
impeller and is modified in shroud due to the reduction of the meridi streamlines have less space to move. As a result, the possibility of flow
onal depth of the impeller outlet. According to Fig. 14, the leading edge separation reduces and the streamlines follow the channel space be
of the optimum impeller splitter in the hub has moved to the trailing tween the blades well.
edge until the end of the design variable 17 range. The reverse is true in On the contrary, in the impellers with a lower backswept angle, the
shroud and variable 18. Due to the design variable 16, which is related cross-sectional area perpendicular to the flow is larger (Fig. 12) and the
to the pitch position of the splitter, in the optimum geometry, the splitter streamlines have more space to distance from each other. As a result, the
has been somewhat approached the main blade from its suction surface. static pressure increases, and at the same time, the possibility of flow
After comparing the geometry of the optimum and baseline com separation increases, and the streamlines do not follow the channel
pressors, we now compare their performance. The performance maps of space between the blades well. Consequently, an impeller design such as
the optimum and baseline compressors in the design speed are shown in the optimum impeller that achieves a high-pressure ratio without flow
Fig. 15. In the optimum compressor, the choked mass flow rate increased separation while the tip diameter of impeller and design rotational speed
from 0.967 to 0.988. Also, the pressure ratio has increased significantly. remain constant will be highly successful and innovative.
In the problem ahead, this is the most important factor in increasing the The flow separation shows its effect in the relative Mach number
efficiency of the thermodynamic cycle of the single-shaft gas turbine contour at the impeller outlet as a region with Mach number close to
engine. The operating range of the compressor is defined as follows [52]. zero. This phenomenon usually exists in centrifugal impellers, and the
better the design of the impeller, the smaller the area of this region and
Operating Range = (ṁchoke − ṁsurge )/ṁsurge (11)
the energy loss in it. In both the baseline and optimum designs, there is
According to the above formula, the operating range has been no flow separation in the impeller and we only encounter regions where
reduced from 0.6279 for the baseline compressor to 0.5486 for the op the flow energy is lower than other regions, so that, the relative Mach
timum compressor. This was predictable, because in redesign processes number in this region in the baseline and optimum impellers, is about
when the tip diameter of impeller and design rotational speed remain 0.30 and 0.15 respectively (Fig. 16). This results in high isentropic ef
constant, the operating range usually decreases with increasing the ficiency for both designs.
pressure ratio [53]. However, here with a 29.62% increase in pressure Considering the increase in the flow cross-sectional area in the op
ratio at the design point, the operating range has decreased by 12.64%. timum geometry and the mass flow rate is almost the same in both ge
Moreover, in Fig. 15 the efficiency diagram in both compressors are very ometries, as expected, the average outlet relative Mach number of the
close to each other, however, the efficiency of the optimum compressor optimum impeller (0.4681) is lower than the baseline (0.5099). This
justifies part of the difference of Mach number values of 0.3 for the
baseline geometry and 0.15 for the optimum geometry in the low energy
regions. While the average outlet Mach number of the optimum and
baseline impeller is 0.4681 and 0.5099, respectively, and the Mach
number in their low energy regions is 0.15 and 0.3, the difference be
tween the highest and the lowest values in the outlet Mach number
contours of the optimum and baseline geometry is about 0.45 and 0.40,
respectively (Fig. 16). It means that the flow uniformity at the outlet is
slightly better in the baseline impeller. In addition, the absolute Mach
number of the impeller outlet increased from 0.8968 in the baseline
compressor to 1.0959 in the optimum compressor. These two factors
(uniformity and low absolute Mach number) cause a wider operating
range of the baseline geometry in the design speed.
This is the essence of centrifugal compressors, by reducing the
backswept angle, the pressure ratio and outlet absolute Mach number
increase and their operating range decreases, and their flow is highly
inclined to separation [47]. Because the optimum geometry has a high-
quality design, despite the large reduction of the backswept angle
compared to the baseline geometry, it still provides a high-pressure ratio
without flow separation and with high isentropic efficiency. However,
there is a slight decrease in the uniformity of outlet flow.
Since the goal of optimization is to improve the performance of the
engine and not the compressor, it should not be expected that the per
formance of the optimized compressor is necessarily better than the
baseline compressor in all performance parameters. Although, even in
the optimization with the aim of improving the performance of the
compressor itself, it is not possible to improve all the performance pa
rameters at the same time, because there is a compromise between the
performance parameters of the centrifugal compressor, which means
that with a constant impeller tip diameter and design rotational speed, it
is not possible to have a compressor that on the one hand, it should have
a high-pressure ratio, and on the other hand, it should have a wide
operating range and a low outlet absolute Mach number.
Fig. 15. Comparison of performance map of the optimum and baseline com Which design is optimum depends entirely on the application and the
pressors at the design speed. definition of the objective function based on it, perhaps in an application
13
Fig. 16. The contours of the relative Mach number at the exit of the impeller of the optimum and baseline geometries.
that requires a wide operating range against a high-pressure ratio, the optimum and baseline compressors were examined, and by comparing
baseline geometry will be optimum. the value of design variables, important lessons were learned. Further
more, a detailed discussion was conducted on the design space trans
4. Conclusions parency, convergence criteria, and optimum design selection, the
disadvantages of current optimization processes were explained and the
In this paper, the thermodynamic cycle efficiency of a single-shaft advantages of the present one were expressed. The optimized engine will
micro gas turbine engine was maximized at the design point by modi be used in a CHP unit.
fying its centrifugal compressor geometry, while the engine output The most important limitation in the present study is that any
power was not less than 150 kW, and the impeller tip diameter and the conclusion is obtained after a large number of numerical simulations, so
design rotational speed was fixed. So, there was a constraint single- the computational cost is too high. The philosophy employed for the
objective optimization problem. In the current optimization processes, single-shaft engines optimization process using compressor modification
there are serious problems in the design space transparency, conver can be extended to other components of the engine or more complex
gence criteria, and optimal design selection, and unfortunately, almost engines.
all researchers used such processes for optimization. Therefore, the va
cancy of a reliable and reproducible optimization process that can be CRediT authorship contribution statement
employed for the successful optimization of turbomachines or gas tur
bine engines’ performance is quite obvious. In addition, research has S.M. Hosseinimaab: Methodology, Software, Validation, Writing –
focused on optimizing engine components such as compressors and original draft, Writing – review & editing, Visualization. A.M. Tousi:
turbines, and optimization to improve engine performance as an end Conceptualization, Supervision, Project administration.
product has received less attention. Optimization is a process, not a
function, and the optimizer is only a function in this process. Simply
Declaration of Competing Interest
running an optimizer in the optimization problems like the problem
studied in this paper without supervising its operation will not lead us to
The authors declare that they have no known competing financial
the true answers.
interests or personal relationships that could have appeared to influence
A state-of-the-art optimization process was presented according to
the work reported in this paper.
the physics of the optimization problem of a single-shaft gas turbine
engine. In addition to the optimization process itself, its components
should be selected according to the problem characteristics in order to References
perform properly and efficiently. The new optimization process has the
[1] Cohen H, Rogers GFC, Saravanamuttoo HIH. Gas turbine theory. 4th ed. Harlow:
least need for the designer’s expertise and acquires a reliable and Longman; 1996.
practical optimum design with the minimum computational cost and a [2] Walsh PP, Fletcher P. Gas turbine performance. 2nd ed. Oxford: Blackwell Science;
number of samples. In addition, the presented optimization process is 2004.
[3] Li Y, Zhang G, Wang L, Yang Y. Part-load performance analysis of a combined cycle
reliable, robust, and reproducible and creates an approximate model with intermediate recuperated gas turbine. Energy Convers Manag 2020;205:
that provides a true and reliable estimate of the design space. 112346. https://doi.org/10.1016/j.enconman.2019.112346.
To examine the accuracy of numerical calculations, the results of the [4] Kiaee M, Tousi AM, Toudefallah M. Performance adaptation of a 100 kW
microturbine. Appl Therm Eng 2015;87:234–50. https://doi.org/10.1016/j.
numerical simulation of the NASA CC3 compressor were validated applthermaleng.2015.04.075.
against the test rig results. This compressor was scaled down for the [5] Chen Q, Han W, Zheng J, Sui J, Jin H. The exergy and energy level analysis of a
baseline engine (a 150 kW single-shaft gas turbine). combined cooling, heating and power system driven by a small scale gas turbine at
off design condition. Appl Therm Eng 2014;66(1):590–602. https://doi.org/
The baseline engine was optimized using the proposed optimization 10.1016/j.applthermaleng.2014.02.066.
process. The performance of the optimum and baseline compressors and [6] Cao Y, Dai Y. Comparative analysis on off-design performance of a gas turbine and
engines were compared and a significant improvement was observed, ORC combined cycle under different operation approaches. Energy Convers Manag
2017;135:84–100. https://doi.org/10.1016/j.enconman.2016.12.072.
where efficiency increased by 13.93% and power by 11.07%, and the
[7] Huang Y, Turan A. Mechanical equilibrium operation integrated modelling of
specific fuel consumption decreased by 12.15%. Despite the constant hybrid SOFC – GT systems: Design analyses and off-design optimization. Energy
impeller tip diameter and rotational speed, in the optimum compressor, 2020;208:118334. https://doi.org/10.1016/j.energy.2020.118334.
the mass flow increased by 2.22% and the pressure ratio by 29.62% [8] Asgari N, Khoshbakhti Saray R, Mirmasoumi S. Energy and exergy analyses of a
novel seasonal CCHP system driven by a gas turbine integrated with a biomass
compared to the baseline compressor, while the isentropic efficiency not gasification unit and a LiBr-water absorption chiller. Energy Convers Manag 2020;
only did not decrease but also increased slightly (0.2%). However, the 220:113096. https://doi.org/10.1016/j.enconman.2020.113096.
operating range decreased by 12.64%. In addition, the geometry of the [9] Nejat A, Mirzabeygi P, Shariat PM. Airfoil shape optimization using improved
multiobjective territorial particle swarm algorithm with the objective of improving
14
stall characteristics. Struct Multidiscip Optim 2013;49(6):953–67. https://doi.org/ [30] He X, Zheng X. Performance improvement of transonic centrifugal compressors by
10.1007/s00158-013-1025-3. optimization of complex three-dimensional features. Proc Inst Mech Eng G: J
[10] Arias-Montano A, Coello Coello C, Mezura-Montes E. Evolutionary algorithms Aerosp Eng 2017;231:2723–38. https://doi.org/10.1177/0954410016673395.
applied to multi-objective aerodynamic shape optimization. Comput Optim [31] Khalfallah S, Ghenaiet A. Radial basis function-based shape optimization of
Methods Algoritm 2011;356:211–40. https://doi.org/10.1007/978-3-642-20859- centrifugal impeller using sequential sampling. Proc Inst Mech Eng G: J Aerosp Eng
1_10. 2015;229(4):648–65. https://doi.org/10.1177/0954410014538421.
[11] Rao SS. Engineering optimization theory and practice. 4th ed. New Jersey: John [32] Song L, Guo Z, Li J, Feng Z. Research on metamodel-based global design
Wiley & sons; 2009. optimization and data mining methods. ASME J Eng Gas Turbines Power 2016;138
[12] Holland JH. Adaptation in natural and artificial systems. Ann Arbor: University of (9):092604. https://doi.org/10.1115/1.4032653.
Michigan Press; 1975. [33] Chahine C, Seume JR, Verstraete T. The influence of metamodeling techniques on
[13] Kirkpatrick S, Gelatt CD, Vecchi MP. Optimization by simulated annealing. Science the multidisciplinary design optimization of a radial compressor impeller. ASME
1983;220(4598):671–80. Turbo Expo: Power for Land, Sea, and Air 2012;1951:64. https://doi.org/10.1115/
[14] J, Kennedy RC, Eberhart Particle swarm optimization. IEEE International GT2012-68358.
Conference on Neural Networks 1995. [34] Guo S, Duan F, Tang H, Lim SC, Yip MS. Multi-objective optimization for
[15] Colorni A, Dorigo M, Maniezzo M. Distributed optimization by ant colonies. centrifugal compressor of mini turbojet engine. J Aero Sci Technol 2014;39:
Cambridge: MIT Press; 1992. 414–25. https://doi.org/10.1016/j.ast.2014.04.014.
[16] Li X, Zhao Y, Liu Z. A novel global optimization algorithm and data-mining [35] Moussavi SA, Hajilouy Benisi A, Durali M. Effect of splitter leading edge location
methods for turbomachinery design. Struct Multidiscip Optim 2019;60(2): on performance of an automotive turbocharger compressor. Energy 2017;123:
581–612. https://doi.org/10.1007/s00158-019-02227-5. 511–20. https://doi.org/10.1016/j.energy.2017.02.011.
[17] Aslimani N, Ellaia R. A new hybrid algorithm combining a new chaos optimization [36] Li X, Zhao Y, Liu Z, Chen H. The optimization of a centrifugal impeller based on a
approach with gradient descent for high dimensional optimization problems. Comp new multi-objective evolutionary strategy. ASME Turbo Expo: power for Land. Sea
Appl Math 2018;37:2460–88. https://doi.org/10.1007/s40314-017-0454-9. and Air 2016;56592. https://doi.org/10.1115/GT2016-56592.
[18] Zhang T, Wang Z, Huang W, Yan L. Parameterization and optimization of [37] Wang J, Guo Y, Zhou K, Xia J, Li Y, Zhao P, et al. Design and performance analysis
hypersonic- gliding vehicle configurations during conceptual design. Aero Sci of compressor and turbine in supercritical CO2 power cycle based on system-
Technol 2016;58:225–34. https://doi.org/10.1016/j.ast.2016.08.020. component coupled optimization. Energy Convers Manag 2020;221:113179.
[19] Ou M, Yan L, Huang W, Zhang T. Design exploration of combinational spike and https://doi.org/10.1016/j.enconman.2020.113179.
opposing jet concept in hypersonic flows based on CFD calculation and surrogate [38] BJ, McBride, MJ, Zehe, S. Gordon, Nasa glenn coefficients for calculating
model. Acta Astronaut 2019;155:287–301. https://doi.org/10.1016/j. thermodynamic properties of individual species, NASA Technical Publication,
actaastro.2018.12.012. 20020085330; 2002. https://ntrs.nasa.gov/citations/20020085330.
[20] Tang X, Gu N, Wang W, Wang Z, Peng R. Aerodynamic robustness optimization and [39] Casey MV. A computational geometry for the blades and internal flow channels of
design exploration of centrifugal compressor impeller under uncertainties. Int J centrifugal compressors. ASME J Eng Power 1983;105:288–95. https://doi.org/
Heat Mass Transfer 2021;180:121799. https://doi.org/10.1016/j. 10.1115/1.3227414.
ijheatmasstransfer.2021.121799. [40] Aungier RH. Centrifugal compressors a strategy for aerodynamic design and
[21] Roy D, Samanta S, Ghosh S. Performance optimization through response surface analysis. New York: ASME press; 2000.
methodology of an integrated biomass gasification based combined heat and power [41] Saltelli A, Chan K, Scott EM. Sensitivity analysis. Chichester: John Wiley & Sons
plant employing solid oxide fuel cell and externally fired gas turbine. Energy Ltd; 2001.
Convers Manag 2020;222:113182. https://doi.org/10.1016/j. [42] Launder BE, Spalding DB. Lectures in mathematical models of turbulence. London:
enconman.2020.113182. Academic Press; 1972.
[22] Wang Y, Liu T, Meng Y, Zhang D, Xie Y. Integrated optimization for design and [43] Menter FR. Two-equation eddy-viscosity turbulence models for engineering
operation of turbomachinery in a solar-based Brayton cycle based on deep learning applications. AIAA J 1994;32(8):1598–605. https://doi.org/10.2514/3.12149.
techniques. Energy 2022;252:123980. https://doi.org/10.1016/j. [44] Haykin S. Neural networks a comprehensive foundation. 2nd ed. Delhi: Pearson
energy.2022.123980. Education; 1999.
[23] Mengistu T, Ghaly W. Aerodynamic optimization of turbomachinery blades using [45] Skoch GJ, Prahst PS, Wernet MP, Wood JR, Strazisar AJ. Laser anemometer
evolutionary methods and ANN-based surrogate models. Optim Eng 2008;9: measurements of the flow field in a 4:1 pressure ratio centrifugal impeller. ASME
239–55. https://doi.org/10.1007/s11081-007-9031-1. Int Gas Turbine and Aeroengine Congress and Exhibition 1997;V001T03A049..
[24] Ju Y, Qin R, Kipouros T, Parks G, Zhang C. A high-dimensional design optimisation https://doi.org/10.1115/97-GT-342.
method for centrifugal impellers. Proc Inst Mech Eng A: J Power Energy 2016;230 [46] McKain TF, Holbrook GJ. Coordinates for a high performance 4:1 pressure ratio
(3):272–88. https://doi.org/10.1177/0957650915626274. centrifugal compressor. NASA Contract Report, 204134; 1982. https://ntrs.nasa.
[25] Verstraete T, Alsalihi Z, Van den Braembussche RA. Multidisciplinary optimization gov/citations/19970024917.
of a radial compressor for microgas turbine applications. J Turbomach 2010;132 [47] Japikse D. Centrifugal compressor design and performance. 1st ed. Vermont:
(3):031004. https://doi.org/10.1115/1.3144162. Concepts ETI; 1996.
[26] Ekradi K, Madadi A. Performance improvement of a transonic centrifugal [48] Fritsch FN, Carlson RE. Monotone piecewise cubic interpolation. SIAM J Numer
compressor impeller with splitter blade by three-dimensional optimization. Energy Anal 1980;17(2):238–46.
2020;201:117582. https://doi.org/10.1016/j.energy.2020.117582. [49] Kahaner D, Moler C, Nash S. Numerical methods and software. Englewood Cliffs,
[27] McKay MD, Beckman RJ, Conover WJ. A comparison of three methods of selecting N.J.: Prentice Hall; 1988.
values of input variables in the analysis of output from a computer code. [50] Persico G, Rodriguez-Fernandez P, Romei A. High-fidelity shape optimization of
Technometrics 1979;21:239–45. non-conventional turbomachinery by surrogate evolutionary strategies.
[28] Ju YP, Zhang CH, Chi XL. Optimization of centrifugal impellers for uniform J Turbomach 2019;141(8):081010. https://doi.org/10.1115/1.4043252.
discharge flow and wide operating range. J Propul Power 2012;28:888–99. [51] Cho S-Y, Ahn K-Y, Lee Y-D, Kim Y-C. Optimal design of a centrifugal compressor
https://doi.org/10.2514/1.B34193. impeller using evolutionary algorithms. Math Probl Eng 2012;2012:1–22.
[29] Guo Z, Song L, Zhou Z, Li J, Feng Z. Multi-objective aerodynamic optimization [52] Ibaraki S, Van den Braembussche R, Verstraete T, Alsalihi Z, Sugimoto K, Tomita I.
design and data mining of a highly pressure ratio centrifugal impeller. ASME J Eng In: 11th International Conference on Turbochargers and Turbocharging. Elsevier;
Gas Turbines Power 2015;137(9):092602. https://doi.org/10.1115/1.4029882. 2014. p. 65–77.
[53] Krain H. Review of centrifugal compressor’s application and development. J
Turbomach 2005;127(1):25–34.https://doi.org/10.1115/1.1791280.
15

Jurnal Propulsi

Uploaded by

Copyright:

Available Formats

Jurnal Propulsi

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Jurnal Propulsi

Uploaded by

Copyright:

Available Formats

Energy Conversion and Management 271 (2022) 116245

Contents lists available at ScienceDirect

Energy Conversion and Management

Optimizing the performance of a single-shaft micro gas turbine engine by

they will be inefficient, computationally expensive, and find a local

β(t) = β0 (1 − t)4 + 4β1 t(1 − t)3 + 6β2 t2 (1 − t)2 + 4β3 t3 (1 − t) + β4 t4 (6)

In above equation, the air is considered as an ideal gas and the

Cp0 (T)/R = a1 T − 2 + a2 T − 1 + a3 + a4 T + a5 T 2 + a6 T 3 + a7 T 4 (3)

2.2. Design variables

The main blade of the three-dimensional centrifugal impeller is

Fig. 5. Flowchart of the optimization process.

TIT(K) 1188 ΔP0cc 0.07

compressor and a vaneless diffuser and the geometrical parameters of

3.3. Correct extraction of the design point from numerical simulation

In this section, a procedure for the correct extraction of the design

Fig. 7. Grid independency study.

the results obtained.

ṁc (kg/s) 0.88 0.89957 2.2239

[25] or Ekradi and Madadi [26], eliminate unacceptable geometries or

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.