Science Arts & Métiers (SAM)

is an open access repository that collects the work of Arts et Métiers Institute of
Technology researchers and makes it freely available over the web where possible.

This is an author-deposited version published in: https://sam.ensam.eu

Handle ID: .http://hdl.handle.net/10985/17836

To cite this version :

Mohamed Abdessamed AIT CHIKH, Idir BELAIDI, José PARIS, Michael DELIGANT, Farid
BAKIR, Sofiane KHELLADI - Efficiency of bio- and socio-inspired optimization algorithms for axial
turbomachinery design - Applied Soft Computing - Vol. 64, p.282-306 - 2018

Any correspondence concerning this service should be sent to the repository

Administrator : [email protected]
Efficiency of bio- and socio-inspired optimization algorithms for axial
turbomachinery design
Mohamed Abdessamed Ait Chikh a,∗ , Idir Belaidi a , Sofiane Khelladi b , José Paris c ,
Michael Deligant b , Farid Bakir b
a
Laboratoire d’Energétique, Mécanique et Ingénierie, Université de Boumerdes, Algeria
b
Laboratoire de Dynamique des Fluides, Arts et Métiers ParisTech, France
c
GMNI, Universidad da Coruña, Spain

a b s t r a c t

Turbomachinery design is a complex problem which requires a lot of experience. The procedure may be
speed up by the development of new numerical tools and optimization techniques. The latter rely on the
parameterization of the geometry, a model to assess the performance of a given geometry and the defi-
nition of an objective functions and constraints to compare solutions. In order to improve the reference
machine performance, two formulations including the off-design have been developed. The first one is
the maximization of the total nominal efficiency. The second one consists to maximize the operation
area under the efficiency curve. In this paper five optimization methods have been assessed for axial
pump design: Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Cuckoo Search (CS), Teach-
ing Learning Based Optimization (TLBO) and Sequential Linear Programming (SLP). Four non-intrusive
methods and the latter intrusive. Given an identical design point and set of constraints, each method
proposed an optimized geometry. Their computing time, the optimized geometry and its performances
(flow rate, head (H), efficiency (), net pressure suction head (NPSH) and power) are compared. Although
all methods would converge to similar results and geometry, it is not the case when increasing the
range and number of constraints. The discrepancy in geometries and the variety of results are presented
and discussed. The computational fluid dynamics (CFD) is used to validate the reference and optimized
machines performances in two main formulations. The most adapted approach is compared with some
existing approaches in literature.

1. Introduction local optimization methods. On such methods, the performances

of turbomachines were able to be improved by making geome-
Turbomachines are omnipresent in everyday domestic environ- try modifications. However, it is established that the fluid flow
ment and in industry. The applications range from air conditioning in turbomachines is viscous, three-dimensional (3D) and turbu-
to water distribution and from automotive to aeronautics and naval lent. Sorensen et al. [6,7] coupled the resolution of the momentum
engines. In the turbomachines design optimization, because of the and the energy differential equations (CFD) with a nonlinear opti-
large number of variables of design parameters, which is in this case mization method to optimize the performance of a given design.
always larger than the number of equations, this can be handled by Moreover, the use of CFD tools within 3D RANS solvers are
the classical design methods based on the exploitation of empirical widely employed with evolutionary optimization algorithms, such
laws like Coridier’s diagram [1,2]. as genetic algorithms [8,9], with gradient methods [10] such as
Most studies concerning turbomachines optimization prob- Simplex, with simulated annealing (SA) [11] and with differen-
lem focus on the maximization of efficiency. Some optimization tial evolution algorithm [12]. As drawback, the use of CFD for
approaches are based on the resolution of analytical and loss the optimization of turbomachines remains very expensive which
models [3,4], others on empirical approaches [5], but both using directed the researches to the exploitation of alternative meth-
ods that establish a relation between the design parameters and
the targeted performances. One can make reference to design of
experiment methods, response surface, kriging, neuron networks
∗ Corresponding author.
and so on. In using local method optimization with a complex
E-mail address: [email protected] (M.A. Ait Chikh).
problem. The missing of the global solution often happens. May grating the off-design”, the number of constraint is low and the
this type of method gives a good precision result when it’s cou- computational time does not represent a priority criterion, we pro-
pled with a global method as Genetic Algorithm [13]. Although pose in this paper an approach for the design optimization of an
the latter is considered as global method but the optimal solu- axial pump based on inverse method with some meta-heuristics
tion is not granted mainly with large dimension of the problem. algorithms based on Bio-inspired approaches namely Genetic Algo-
May be using the other method in parallel, tell us more about rithm (GA), Particle Swarm Optimization (PSO) and Cuckoo Search
optimal solution [14]. This work maybe considered among one of (CS), in one hand, and socio-inspired approaches namely Teach-
the rare study that compared several optimization methods for ing Learning Based Optimization (TLBO). These methods present
turbomachinery design namely GA,SA and Sequential Quadratic the advantage to be non-intrusive to the design codes. For each
Programming. However, the comparison study was not presented method, many runs are repeated for choosing the best solution.
as statistical significance. Recently, the PSO algorithm became In this paper, starting from a reference axial pump, we ana-
included in the turbomachinery design specifically in aerodynam- lyze the performance of several natural inspired optimization
ics type [15,16]. For more information about design optimization algorithms in two main formulations of the design problem by
methods in turbomachinery aerodynamics, see Ref. [17]. Concern- considering a very large number of geometrical and mechanical
ing the optimization strategy and objective parameters, Li [18] constraints handled by modified constant penalty technique for the
proposed a two stages approach to optimize the required Net Pos- reducing CPU time. The first formulation consists of maximizing
itive Section Head (NPSHr) of an axial pump. The optimization is the total nominal efficiency. Five scenarios of different complexity
carried out by applying the response surface method with the radial have been considered using free vortex (FV), constant vortex (CV)
equilibrium equation and the theory of actuator disks. Although and forced vortex (FCV).
this study have allowed to reduce the NPSHr by 37% under specific In the second formulation, we propose a new objective func-
required point, the operation pump under other conditions (volu- tion which consist in maximizing the surface constructed by the
metric flow rate range) was not predicted. Lin et al. [19] proposed total efficiency and the [−25%, +25%] range of nominal volume
a complete approach to minimize the losses of an axial fan and flow rate, with the constraint that the variation of the pump head
to reduce the overall computational time. This approach has four within this range does not exceed 15% of that of the specifications.
stages: (1) construction of 3D blade using the “Generated Machin- In this formulation, three scenarios are considered using constant
ing Method”, (2) analysis of the performances of the obtained vortex strategy. A SLP, which is considered as an intrusive, is also
geometry by CFD, (3) establishment of the relation between the used for first scenario in case one in order to compare the effi-
geometrical parameters and flow properties by neural networks ciency and the performance of the proposed algorithms to intrusive
method, and finally (4) search for the optimal geometry by complex approaches family algorithms. The computational fluid dynamics
optimization method. These four steps seem long and no reduction (CFD) is employed to predict the performance of the reference and
of process’s time was mentioned. In most cases, the objective of the optimized machines. Finally, after the behavior of each algorithm
design of an axial fan/pump aims at the optimization of the effi- was analyzed and the approaches were classified, the most adapted
ciency. Huang and Gau [20] opted for a different approach, such as approach is selected to compare its optimized machines with some
the chosen cost function leads to obtaining the desired of volumet- existing machines in literature.
ric flow rate of an axial fan, since the optimization of efficiency does
not guarantee the increase of the volumetric flow rate. They used
2. Design of axial pumps and fans
CFD to estimate the volumetric flow rate according to optimiza-
tion variables. Starting from an initial geometry, the updates of the
2.1. Inverse and direct approaches
design variables were made by Levenberg–Marquardt method to
minimize the objective function.
The inverse design is applied to define the geometry of the
Always concerning axial fans, Lee et al. [21] proposed a strategy
rotor from a priori known specifications (volume flow rate, head
for the optimization of an axial fan consisting in coupling experi-
and rotational velocity), and geometrical parameters (inlet/outlet
mental analysis with inverse design approach (using TURBOdesign
hub/tip radius and rotor blades number). The direct approach con-
code) for the choice of the decision variables. The response surface
sists on analyzing the performances of a given geometries, as for
method was then used for the optimization of the efficiency and the
instance those given by the inverse method. It permits to define
pressure rise. In another article, Lee et al. [22] presented a numeri-
the overall characteristics (performances) of the machine, in other
cal procedure of a low rotational speed axial fan optimization. The
terms, the evolution of head, efficiency, power, NPSH, etc. with flow
objective function was established from a polynomial interpolation
rate.
based on the response surface method and CFD. The method of gra-
The inverse approach uses laws based on a simplified radial
dient was then used to find out the optimal shape of the machine.
equilibrium (1D) model to define some coherent geometries,
The CPU time required was approximately 12 h with a Pentium-IV,
whereas direct method uses loss laws (incidence, friction, flow leak-
3.0 GHz processor. Song and Sun [23] developed an integral design
age, etc.) to estimate the performances of the defined geometry,
approach of a transonic fan. In order to define an objective function
see Refs. [24,25] for more details concerning the chosen approach.
and then optimize the efficiency and the pressure, they combined
Below some parameters which used in the present inverse design
global optimization method, experience design method, CFD and
approach:
finally kriging reconstruction. The total running time is about 71 h
with using Dell 7500 (with 2 Xeon 2.93 GHz processors and 24 GB
• The type of vortex is chosen arbitrarily with a hydraulic efficiency
RAM).
Most works that use modeling of the relationship between (H ) of 70%.
• The rotor outlet tangential velocity Cu2 (r) is variable and depends
design variable and machine’s performance are unconstrained or
with a low number of constraints, may be because the expensive on the radius (between hub and tip) according to Eqs. (1)–(3) for
computational time needed, also, the off-design still using only in free vortex, constant vortex and forced vortex respectively.
the verification of the performance and does not exploit in the K
optimization process. Cu2 (r) = (1)
r
Differently to the optimization methods described above, in
which the objective functions are conventional “without inte- Cu2 (r) = K (2)
 Fig. 1. Velocity triangles of the axial compressible rotor.

Cu2 (r) = Kr (3)

where K = g H/(ωH ) for free vortex, K = gH/(ωH (Ri2 + Re2 )/2)

2
for constant vortex and K = gH/(ωH ((Ri2 + Re2 )/2) ) for forced
vortex, g is the gravity, H is the head, ω is the angular velocity
and (Ri2 , Re2 ) are the hub and tip outlet radius respectively.
• The velocity triangles are calculated by applying Eulers’ equa-
tion for each section along the rotor hub and tip radius (radial
discretization) (Fig. 1).
• The stagger angle  is determined from the angle of incidence i
defined experimentally by Eqs. (4)–(6) (ˇ = ˇ1 − ˇ2 ) represent
the angular deviation. It is, in fact, given by minimizing losses
(continuity of pressure distribution on the blade pressure side)
Fig. 2. Scheme of an axial rotor.
for the case of a blade cascade of NACA65 airfoil (10% thickness).

 = ˇ1 − i (4)
2.2. Design descriptive parameters
ˇ + 0.94
i= + 2.07 (5) The definition of the geometry of turbomachines requires the
q(ˇ1 )
knowledge of a set of some geometrical parameters, every valuable
q(ˇ1 ) = 2.103 − 4.01910−7 × ˇ13.382 (6) variation of parameters describe a different geometrical shape, i.e.,
• The solidity a different machine with its own performances. These geometrical
( 2 )
at the outlet of the rotor in the hub and tip is
parameters will be associated with some other relevant mechanical
expressed according to the diffusion factor of Lieblein D, accord-
and fluid parameters.
ing to Eq. (7) and at the intermediate radius according to Eq. (8)
respectively.
2.2.1. Geometrical parameters
cos(ˇ1 ) × |Ca1 × tan(ˇ1 ) − Ca2 × tan(ˇ2 )| The geometrical parameters considered in the context of this
2 = (7)
2 × Ca1 × (D − 1 + Ca2 × cos(ˇ1 )/(Ca1 × cos(ˇ2 ))) study are: the slack distance between the tip of the rotor blades
• The chord at the rotor hub and tip are calculated using Eq. (8) and and the hull, the mean roughness of the rotor blades, the number
by linearization at the intermediate radius. of the rotor blades (Z), the hub and tip inlet radius (Ri1 , Re1 ), the hub
and tip outlet radius (Ri2 , Re2 ), the diffusion factors at the hub (Di )
2r
C = 2 (8) and the tip (De ) of the rotor blades, and the minimum value of the
Z maximum thickness of an airfoil section, see Fig. 2.
• The solidity ( 1 ) at the inlet of the rotor is expressed according
to the chord, blades number and the radius, according to Eq. (9). 2.2.2. Mechanical parameters
ZC The token fluid properties, are the fluid density (), and the kine-
1 = (9) matic viscosity of the fluid ( ). The mechanical parameters are: the
2r
(initial) theoretical volume flow rate (Qv ), rotational velocity (N),
• The camber coefficients are expressed according to incidence
head (H).
angle and the inlet solidity, and they are given by Eqs. (10) and
(11).
3. Formulation of the design optimization problem
i + 2.525
Cz∞0 = − 0.823 (10)
p( 1 ) In this section, a model adapted to optimization methods is pro-
posed, it allows to obtain an optimal geometry according to a given
−0.4242 1
p( 1 ) = 15.535 − 12.467e (11) objective and needs respecting some imposed constraints.
• Finally, to complete the empirical equations relative to the
NACA65 airfoil, the solidity and the camber are bounded due to 3.1. Parameters held constant during the process of optimization
geometrical and technological integrity limitation, and will be
used as constraints in the formulation of the optimization prob- During all the process of optimization some parameters will be
lems. considered constant, for instance:
 Fig. 3. Objective function of the two cases. Fig. 4. Comparison of the computed efficiency with a second order interpolating
polynomial and a fourth order interpolating polynomial for different values of the
volume flow rate.
• the slack distance between the rotor blades tip and the hull, the
roughness of the rotor blades, the maximal thickness of the rotor
ciency and the axis of volume flow rate as shown in Fig. 3. The
blades;
objective of this optimization approach is to widen the range of
• the fluid properties (density and viscosity),
variation of the efficiency in the neighborhood of the design point
• parameters connected to the mechanical and fluid properties,
(nominal flow condition), a numerical integration (based on trape-
being a part of a tender specifications (operating condition needs)
zoidal rule) is used. The flow rate integration range varies between
such as: the rotational speed, the head, etc. are fixed a priori;
−25% and +25% of the nominal volume flows rate (Qvnom), the
• computational parameters such as the number of volume flow
objective function is then given by:
rate point and radial sections discretization, etc. are also consid-
 1.25Qvnom
ered constant during all the process of optimization. Their value
has influence only on the accuracy of the results of the analysis max , F = (QV ), dQv (13)
0.75Qvnom
model. These two parameters can be however, modified for more
precision.
3.3.2. Intrusive method
The analysis formulation developed by the authors computes
As already mentioned previously, the modification of this set
the nominal flow rate by studying the response of the proposed
of parameter allows to obtain several geometrical configurations
design for a set of initial theoretical volume flow rates homo-
(several machines with their own corresponding performances).
geneously distributed in the interval [Qvmin, Qvmax] previously
defined. Consequently, a predefined number of database points of
3.2. Considered decision variables
the efficiency–volume flow rate function can be obtained and an
approximation to the nominal flow rate can be estimated.
The design variables relative to the proposed model of optimiza-
In this paper the authors propose the use of a second order inter-
tion are considered as real numbers except the number of blades
polating polynomial to obtain the nominal volume flow rate and the
which is taken as integer. Decision variables are: the theoretical vol-
related nominal efficiency. This approximation can be obviously
ume flow rate (Qv ), inlet hub and tip radius (Ri1 , Re1 ), outlet hub and
improved by increasing the order of the polynomial but, due to
tip radius (Ri2 , Re2 ), hub and tip diffusion factors (Di , De ), the num- practical considerations, second order has been considered by the
ber of blades (Z); however, for the local optimization method, the authors as adequate for this problem. Some numerical tests have
optimal number of rotor blades is determined after analyzing the been developed in order to verify this assumption.
optimum design of turbomachine with various rotor blade number. Fig. 4 shows the computed values of the efficiency of an initial
Practically, each design variable has a variation interval (max design of a turbomachine and two interpolating polynomials: a sec-
and min), usually named “side constraints”. This interval is defined ond order interpolating polynomial and a fourth order interpolating
according to the designer’s needs and for other theoretical or envi- polynomial. The complete database is defined by 25 points that cor-
ronmental constraints. Contrary to classical optimization methods, responds to 25 values of the flow rate homogeneously distributed in
the meta-heuristics methods using in this paper are based on popu- the interval [Qvmin, Qvmax]. The second order interpolating poly-
lation constituted of individual solutions, each individual is defined nomial is defined by three database points: the volume flow rate
by a set of decision variables. with maximum efficiency and the two neighbor volume flow rates
studied. The fourth order interpolating polynomial is defined by
3.3. Objective function five database points: the volume flow rate of the database with
maximum efficiency and the first and second neighbors. According
3.3.1. Non-intrusive methods to this figure, both approximations are adequate in an interval cen-
First case. In this first case, the optimization process con- tered at the volume flow rate that produces maximum efficiency,
sists on maximizing the total efficiency on the axial turbomachine where the nominal volume flow rate and nominal efficiency are
(pump/fan) by varying a set of decision variables. In fact, the effi- contained. The computed values, assumed as exact in this analysis,
ciency depends on many parameters, but generally, the overall are obtained by using a larger database with 100 discrete values
performance of the machine is described according to the volume of the volume flow rate homogeneously distributed in the interval
flow rate. A simple design on a range of [Qvmin , Qvmax ] discretized [Qvmin, Qvmax]. In practice, the relative error of the approximation
uniformly (Fig. 3) shows that the efficiency is a continuous func- around the nominal volume flow rate is adequate for this problem.
tion having a maximal point corresponding to a value of the specific Fig. 5 shows the computed values of the efficiency and two inter-
volume flow rate called: “nominal” volume flow rate. polating functions: a second order interpolating polynomial and a
fourth order interpolating polynomial. Both of them are obtained
max F = (nom ) (12)
with the same algorithms used in the first study example. According
Second case. In the second case we propose a new objective to this figure, both approximations are also adequate in an interval
function in the field of turbomachine optimization. The purpose of centered at the maximum value of the efficiency that contains the
this function is to maximize the area formed by the curve of effi- nominal volume flow rate and the related nominal efficiency.
 rate can be easily obtained by computing the second order Lagrange
interpolating polynomial as:
(Q vnom − Q vi )(Q vnom − Q vi+1 )
nom = i−1
Pi−1
(Q vnom − Q vi−1 )(Q vnom − Q vi+1 )
+ i (16)
Pi
(Q vnom − Q vi−1 )(Q vnom − Q vi )
+ i+1
Pi+1

Fig. 5. Comparison of the computed efficiency with a second order interpolating

Thus, the objective function of the optimization problem consists
polynomial and a fourth order interpolating polynomial for different values of the of maximizing nom defined according to Eq. (16). In practice, opti-
volume flow rate. mization problems are usually stated as minimization problems.
Thus, the objective function consists of minimizing:
Table 1
Nominal volume flow rate and nominal efficiency for a initial design of a turbo-
max , F = −nom (17)
machine and comparison with the second order polynomial interpolation and the
On the other hand, the optimization of turbomachines also requires
fourth order polynomial interpolation.
some specifications to be satisfied. In practice, these specifications
Initial design Qvnom (m3 /h) nom (%) are treated in an optimization problem as design constraints.
Computed values 776.9440 60.15808
Second order polynomial approx 776.8716 60.15801 3.4. The retained constraints
Relative error 9.31 × 10−4 1.20 × 10−6
Fourth order polynomial approx 776.9515 60.15809
Relative error 9.65 × 10−5 3.86 × 10−8
3.4.1. Side constraints
In the search for the optimal geometrical shape certain viola-
tions of the constraints must be avoided. The constraints of the
Table 2 design variable or the side constraints allow to establish inequali-
Nominal volume flow rate and nominal efficiency for an improved design of a turbo- ties of the search spaces described below. These spaces represent
machine and comparison with the second order polynomial interpolation and the
the first constraint that must be treated, since in the practice, all
fourth order polynomial interpolation.
the design variables must be limited by two values (min and max).
Improved design Qvnom (m3 /h) nom (%) In the model of optimization all the design variables vary between
Computed values 719.9224 68.68955 two predefined positive extremes. We note:
Second order polynomial approx 720.0684 68.68900
Relative error 2.03 × 10−4 8.00 × 10−6 xj,Min ≤ xj ≤ xj,Max 1≤j≤D (18)
Fourth order polynomial approx 719.90792 68.68955
Relative error 2.01 × 10−5 1.39 × 10−8 Being that D is the number of variables or the dimensions of design
optimization problems. This type of constraint is handled differ-
ently than the other types i.e. without the integration in the fitness
The computed values, assumed as exact in this analysis, are also of the individual.
obtained by using a larger database with 100 discrete values of the
initial volume flow rate homogeneously distributed in the interval 3.4.2. Geometric constraints
[Qvmin, Qvmax]. The first inclusive geometrical constraint is the hub radius
Figs. 4 and 5 and Tables 1 and 2 show that the proposed second constraints. Because of the random nature of the meta-heuristics
order interpolating polynomial gives accurate approximations of algorithms, it can fall in cases where the hub radius at the entrance
the nominal volume flow rate and nominal efficiency. of the rotor is larger than that at the exit, which could cause a per-
The objective function of this problem is the maximization of the turbed flow. A constraint of the entrance and exit hub radius must
nominal efficiency. Thus, we propose to use the second order inter- be respected as follows:
polating polynomial as the mathematical expression that allows to Ri1 − Ri2 ≤ 0 (19)
compute the objective function. In practice, the nominal volume
flow rate can be obtained as: On the other hand, the constraint of the rotor thickness has been
added, which represents the distance between the entrance and
Si−1 Si Si+1

Pi−1 i−1
+ 
Pi i
+ 
Pi+1 i+1 the exit of the turbomachine. This parameter is calculated in every
Q vnom =   (14) radius situated between hub and tip of the rotor. Every section of
i−1 i i+1
2 Pi−1
+ Pi
+ Pi+1 the discretization has a different airfoil geometry (the stagger angle,
the chord length, the camber, the thickness of airfoil. etc.), all these
being: geometries of airfoil along the radial direction define the blade of
 the turbomachine. In each section, the thickness of the rotor and
Pi = (Q vi − Q vj ), j ∈ {i − 1, i, i + 1} the chord form a stagger angle. The thickness is the projection of

j=
/ i the chord on the longitudinal axis. The design of the turbomachine
(15)
Si = Q vj , j ∈ {i − 1, i, i + 1} must satisfy that a predefined maximal thickness should not be
exceeded. This constraint is represented mathematically as follows:
j=
/ i
lr cos(r ) − thmax ≤ 0, r = 1, . . ., Nrd (20)
and j the efficiency of the axial turbomachine related to volume
flow rate Qvi being. The volume flow rate Qvi corresponds to the where that lr and  r represent respectively the chord and the stag-
value of the database with the largest value of efficiency. The lower ger angle to each section r; thmax the maximum rotor thickness
and upper neighbor points of Qvi that defines the second order allowed, Nrd is the number of discretization sections.
interpolating polynomial are identified as Q vi−1 and Q vi+1 , respec- This model would need local constraints due to empirical rea-
tively. The nominal efficiency related to the nominal volume flow sons linked directly with the airfoil shape of the chosen blade.
Consequently, a maximum camber value was imposed for each penalizing the solutions located in the unfeasible region by using a
section of the radial discretization. penalty constant.
r
Cz∞o − 2.7 ≤ 0 (21) 
m
fitness = F + Ci ıi (26)
The same reasons for the solidity in the entrance r1 And at the exit i=1
r2 of the rotor. 
ıi = 1 if constraint is not respected
r1 − 1.5 ≤ 0 and r2 − 1.5 ≤ 0 (22) where
ıi = 0 else
To ensure geometric relevance and to avoid any kind of inadequate
shapes of the optimized rotor blade. The inlet flow angle must be where m is the number of constraints, C the constant of penalty
greater than the exit flow angle on all radially discretized sections. depending on the objective function F.
In this paper, we propose a strategy that handles the constraints
ˇ1r − ˇ2r > 0 (23) in a way that it decreases the computational time as much as
possible. The evaluation of the individual begins with the inverse
Finally a constraint of minimum nominal value of volume flow
design to determine the geometrical parameters of the machine,
rate was considered to guarantee the specifications of the turbo-
just before passing to the direct method for analyzing the perfor-
machine.
mances, one verifies the geometrical constraints (camber, solidity,
blade thickness, etc.). The fitness of the individual is initialized by
3.4.3. Mechanical constraint null value, if there is at least a single violation of the constraints,
To complete the constraints, the nominal volume flow rate must the individual will be penalized by a constant of penalty equal to
be limited by a lower value. By the fact that the problem of opti- 1. In this case, it is not necessary to analyze the machine by the
mization consists of reducing the nominal volume flow rate in order direct method. If the opposite occurs, i.e. no violation of the con-
to increase the total efficiency. straints is found, one passes to the direct method to determine the
main fitness which is the nominal total efficiency. The nominal con-
Q vnom − Q̂ v ≥ 0 (24) straint of minimum volume flow rate is then handled with the same
method of penalty. The elimination of the direct method for indi-
Knowing that Q̂ v is the allowed minimum nominal volume flow
viduals who do not verify the geometrical constraints allows to
rate.
decrease dramatically the computational time.
The head constraint is added specially for second formulation
problem:
4.2. Intrusive methods
|H(0.75Qv) − H(1.25Qv)| ≤ 0.15H(Q vnom ) (25)
The handling of the constraints in the intrusive gradient-based
methods is strongly linked to the design method itself, this phase
4. Constraints handling
will be detailed in the algorithmic development section.

4.1. Non-intrusive methods

5. Optimization algorithms

A review of the literature [26–28] shows that the constraint

Four bio- and socio-inspired meta-heuristics are used in this
handling methods can be classified as:
work:

• Methods based on preserving feasibility of solutions; • bio-inspired: Genetic Algorithm (GA), Particle Swarm Optimiza-
• Methods based on penalty functions; tion (PSO), Cuckoo Search (CS);
• Methods making distinction between feasible and unfeasible • socio-inspired: Teaching Learning Based Optimization (TLBO).
solutions;
• Methods based on decoders; The proposed set of methods are iterative and non-intrusive, the
• Hybrid methods. variables are modified in each iteration.

The method of penalty is the oldest one, its difficulty is in choos- 5.1. Genetic Algorithm
ing the most effective penalty coefficient according to the problem
setup. Deb [28] proposed a technique without using the coefficient 5.1.1. General principle
of penalty, which is based on the principle of feasible and unfeasible The well known Genetic Algorithms are a part of the evolution-
solutions. ary family based on the theory of evolution and natural selection
In our case (single-objective optimization), as mentioned above, proposed by Darwin. The method, developed by Holland (since
the total efficiency needs to be improved in the interval [0,1]. Like 1960), simulates mathematically the operators of crossover, muta-
the other performances parameters, where their maximum and tion and selection [30]. GA was made popular by Goldberg [31]. An
minimum values can be estimated, the coefficient of penalty must individual (set of variables) is represented by a chromosome and
be satisfied in order to penalize the objective function (the fitness a gene represents a variable formed by a string of 0 and 1. Binary
moves outside its theoretical interval). form of representing variable is the most used.
The variable constraints (side constraints) are treated differ- In this work, real form of variables is used, the flexibility and the
ently. The constraints do not influence on the fitness. That means performances of the algorithm in terms of speed of execution are
that there is no penalization of the objective function. Neverthe- better [32].
less, the constraints must be respected during the evolution of the
population. 5.1.2. GA operators
Other types of constraints may be handled by using the tech- 5.1.2.1. Selection. We chose the selection by binary tournament.
nique of static penalty function [29]. This technique consists of Two individuals are randomly pulled x1 and x2 , and the one who
has the best fitness will be in the next generation (t + 1), the other 16: - Generate the geometry (inverse design)
one will be rejected. 17: - Check the geometrical constraints with the penalty method
18: if no violation found then
if (fitness(x1t ) > fitness(x2t )) xt+1 = x1t ; (27) 19: Analyze the performances (direct design)
20: Check the mechanical constraints (nominal volume flow
t+1
else x = x2t ; (28) rate) with the penalty method (compute the total fitness)
21: end if
where, 22: end for
23: return The best solution is obtained
1 ≤ t ≤ iteration, max

5.1.2.2. Crossover. The intermediate crossover, proposed by Kaya 5.2. Particle Swarm Optimization
et al. [33], which allows to create two “children” from two randomly
pulled “parents” is used, this crossover is controlled by a ratio. 5.2.1. General principle
Particle Swarm Optimization (PSO) method was developed by
x1t+1 = x1t + rand × ratio × (x2t − x1t ) (29)
Kennedy and Eberhart in 1995 [35] it is based on the simulation of
x2t+1 = x2t + rand × ratio × (x2t − x1t ) (30) the swarm behavior of birds or fishes. A particle or an individual
represents a bird/fish who is analogically a set of variable in an
where rand is a random number between [0,1] and the ratio is a optimization problem. When the individual moves from a position
constant between [0,1], it can be larger than 1 if there is a problem to another, it is affected by three main factors:
of premature convergence; ratio = 1.2.

5.1.2.3. Mutation. A Gaussian mutation [34] is preferred with a 1. the attraction toward the leader team,
probability of Pm = 0.01. This method adds a random normal dis- 2. the attraction toward its best position so far,
tribution for each variable. 3. staying in the actual position.
t+1 t
xid = xid + S × randn × (xd, min − xd, max ) (31)
 1 − shrink × (t + 1) 
(t+1) (t) (t) (t)
S = scale × (32) Vid = w × Vid + C1 × rand × (Pbest id − xid ) +
tmax
(t) (t)
with, C2 × rand × (Gbest d − xid ) (34)
1 ≤ i ≤ n, (Population, size)

1 ≤ d ≤ D, (number, of , design, variables) (t+1) (t) (t+1)

xid = xid + Vid (35)
xd min , xd max are respectively the minimum and the maximum
values of variable xd . Knowing that scale is a parameter which deter- Knowing that Pbest is the best position passed by the individual,
mines a standard deviation of the generated random number, its Gbest (the best global individual), is the best position of all best
value is between [0,1], shrink is a number between [0.5,1.0]; in our positions of the generation, V is the velocity of the decision variable,
case we took scale = 0.1 and shrink = 0.5. C1 = C2 = 2.0 are cognitive coefficient and social coefficient respec-
tively, w is the inertia weight, it can be fixed to 0.5 or varied between
5.1.3. Implementation 0.4 and 0.9 during iterations. The inertia weight can be given by,
Each decision variable xid represents a design variable:
t −t

xid ∈ [Qv , N, Ri1 , Re1 , Ri2 , Re2 , De , Di , Z] (33) w(t) = wmin +
max
× (wmax − wmin ) (36)
tmax
Algorithm 1 shows GA optimization process of an axial pump.

Algorithm 1. Design approach using Genetic Algorithm where wmin and wmax are the minimums and the maximums values
of the inertia weight. This formulation, called Linearly Decreasing
1: Start Weight Particle Swarm Optimization (LDW-PSO), is proposed by
2: Input parameters: Shi and Eberhart [36]. A literature review shows a lot of approachs
- Pa, side constraints, maximum iteration, population size,
proposing improvements like in [37] who developed an automatic
number of dimensions (variable)
- Design constant: range of volume flow rate, the head H, restart of the calculation (multi-start PSO) with an initial popula-
rotational speed N, the density and the viscosity of fluid, the tion created randomly when a stagnation of fitness variation during
roughness, the slack distance between the tip of the rotor the iterations occur. Wei et al. [38] add elitism and mutation mech-
blades and the hull, the maximum thickness of rotor, type of anisms to the PSO. In this paper, we propose to use standard PSO
vortex, number of the volume flow rate and the section
3: Initialize the population randomly
with w = 0.5 and an unlimited velocity.
4: Evaluate the population opulation:
5: - Generate the geometry (inverse design)
6: - Check the geometrical constraints with the penalty method 5.2.2. Algorithmic implementation
7: if no violation found then
For each xid , Pbestid and Gbestid such as:
8: - Analyze the performances (direct design)
9: - Check the mechanical constraints (nominal volume flow ⎧ ⎫
rate) with the penalty method (compute the total fitness) ⎪ x
⎨ id ⎪
⎬
10: end if
11: for iteration:=1 to Max iteration do pbestid ∈ [Qv , N, Ri1 , Re1 , Ri2 , Re2 , De , Di , Z] (37)
12: Selection Eq. (27) or (28). ⎪
⎩ ⎪
⎭
13: Crossover Eqs. (29) and (30) gbestd
14: Mutation Eqs. (31) and (32)
15: Evaluate the population:
we propose the PSO Algorithm 2,
Algorithm 2. Design approach using Particle Swarm Optimization 2 = 1 (44)
algorithm
where is the normal distribution of an average 0 and a variance
1: Start  2 , is the normal distribution of an average 0 and a variance 1.
2: Input parameters: The phase of changing the nest for the aim of finding a new solu-
- w, C1 , C2 , space search, max iteration, population size,
tion with biased/selective random walks [41] is given as follows:
number of dimensions (variable)
- Design constant: range of volume flow rate, the head H, Two solutions are randomly pulled, xp and xq and a new nest
rotational speed N, the density and the viscosity of fluid, the from these two solutions is then looked for, therefore
roughness, the slack distance between the tip of the rotor
blades and the hull, the maximum thickness of rotor, type of stepsize = rand × (xpt − xqt ) (45)
vortex, number of the volume flow rate and the section
3: Initialize the population randomly (swarm, Pbest, velocity v) new xit = xit + stepsize (46)
4: Evaluate the population:
5: - Generate the geometry (inverse design) 5.3.2. Algorithmic implementation
6: - Check the geometrical constraints with the penalty method
7: if no violation found then
Every nest represents a set of decision variables xid . The code in
8: - Analyze the performances (direct design) Algorithm 3 shows the process of design optimization of an axial
9: - Check the mechanical constraints (nominal volume flow pump by CS.
rate) with the penalty method (compute the total fitness)
10: end if Algorithm 3. Design approach using Cuckoo Search Algorithm
11: Obtain the global best (gbest)
12: for iteration:=1 to Max iteration do 1: Start
13: Fly Eqs. (34) and (35) 2: Input parameters:
14: Evaluate the population: -Pa, space search, max iteration, population size, number of
15: - Generate the geometry (inverse design) dimensions (variable)
16: - Check the geometrical constraints with the penalty method - Design constant: range of volume flow rate, the head H,
17: if no violation found then rotational speed N, the density and the viscosity of fluid, the
18: - Analyze the performances (direct design) roughness, the slack distance between the tip of the rotor
19: - Check the mechanical constraints (nominal volume flow blades and the hull, the maximum thickness of rotor, type of
rate) with the penalty method (compute the total fitness) vortex, number of the volume flow rate and the section
20: end if 3: Initialize the population randomly
21: Update the Pbest and Gbest 4: Evaluate the population:
22: end for 5: - Generate the geometry (inverse design)
23: return The best solution (Gbest) 6: - Check the geometrical constraints with the penalty method
7: if no violation found then
8: - Analyze the performances (direct design)
5.3. Cuckoo Search Algorithm 9: - Check the mechanical constraints (nominal volume flow
rate) with the penalty method (compute the total fitness)
5.3.1. General principle 10: end if
Following the example of the previous algorithms, the Cuckoo 11: for iteration:=1 to Max iteration do
12: Phase get a nest Eqs. (38)–(43)
Search (CS) is a natural-inspired method formulated by Yang and 13: Evaluate the population:
Deb [39]. CS is based on the behavior of cuckoo, which is character- 14: - Generate the geometry (inverse design)
ized by the practice of “parasitism of brood”, combined to the Levy 15: - Check the geometrical constraints with the penalty method
flight (random walk), which is another behavior of some birds. In 16: if no violation found then
17: - Analyze the performances (direct design)
this algorithm, an individual (proposed solution or the set of vari-
18: - Check the mechanical constraints (nominal volume flow
ables) is considered as a “nest”. A nest which has a good quality rate) with the penalty method (compute the total fitness)
eggs represent an individual of good fitness. 19: end if
The standard CS algorithm [40] has three main stages: 20: Keep the best solution
21: Phase Empty nest Eqs. (45) and (46)
22: Evaluate the population:
1. each cuckoo lays an egg at a time and drops it in a randomly 23: - Generate the geometry (inverse design)
chosen nest, 24: - Check the geometrical constraints with the penalty method
2. the best nest with a high quality of eggs continue existing in the 25: if no violation found then
next generation, 26: - Analyze the performances (direct design)
27: - Check the mechanical constraints (nominal volume flow
3. there is a probability Pa (between 0 and 1), that the host bird
rate) with the penalty method (compute the total fitness)
discovers the cuckoo eggs, thus he can get rid of it either by 28: end if
abandoning its nest or by reconstructing another. 29: Keep the best solution
30: end for
31: return The best solution
The phase of choosing a nest is described as follows:
(t+1) (t)
xi = xi + ˛ ⊕ Lévy(ˇ) (38) 5.4. Teaching-Learning based Optimization Algorithm
(t)
˛ = ˛0 × (best − xi ) (39) 5.4.1. General principle
(t) (t) The previously presented algorithms are natural or bio-inspired.
˛0 × (best − xi ) ⊕ Lévy(ˇ) ≈ 0.01 1
(best − xi ) (40) Differently, the Teaching Learning based Optimization algorithm
|v| ˇ (TLBO) proposed by Rao et al. [42] is based on the influence of a
≈ N(0,  2 ) (41) professor on students scholar performances and the influence of
students on each other. Therefore, this is why this method is consid-
v ≈ N(0, v2 ) (42) ered as socio-inspired. A population is presented by students, each
   1 student represents an individual, that is mean a solution which
ˇ
(1 + ˇ) sin(ˇ/2) contains the variables, referred to as subjects. The teacher is the
2
 =   (43) best individual (reference) which can latter be changed during the
[(1 + ˇ)/2] × ˇ × 2(ˇ−1)/2
iterations. Opposed to other algorithms like GA, PSO and CS, where
the choice of the parameters remains always dependent on several 10: end if
regulation tests, TLBO do not have any regulation parameter. TLBO 11: for iteration:=1 to Max iteration do
was tested with several constrained benchmark functions, where 12: The teaching phase Eq. (47).
13: Evaluate the population:
it showed good performances [43].
14: - Generate the geometry (inverse design).
TLBO is characterized by two phases: teaching and learning. 15: - Check the geometrical constraints with the penalty method
16: if no violation found then
5.4.1.1. Teaching phase. In this phase, the teacher which is the indi- 17: - Analyze the performances (direct design)
vidual who has the best fitness tries to raise or to improve the 18: - Check the mechanical constraints (nominal volume flow
rate) with the penalty method (compute the total fitness)
average of marks of his students (the individuals of the rest of the
19: end if
population) by transmitting its knowledge. A good teacher forms 20: Keep the best solution
good students. Xi = {xi1 , xi2 , xi3 , . . ., xiD }, Xi: individual (student), 21: The Learning phase Eq. (49)
xi1 ,xi2 ,xi3 : the decision variables (the subjects) i = {1, 2, 3, 4, 5, . . ., 22: Evaluate the population:
23: - Generate the geometry (inverse design).
N}, N is the number of student (population size).
24: - Check the geometrical constraints with the penalty method
(t+1) (t) t 25: if no violation found then
xi = xi + rand × (xTeacher − TF M t ) (47)
26: - Analyze the performances (direct design)
Knowing that Mt is a vector mean of each subject. 27: - Check the mechanical constraints (nominal volume flow
rate) with the penalty method (compute the total fitness)
t
M = [mt1 , mt2 , mt3 , mt4 , . . ., mtD ] (48) 28: end if
29: Keep the best solution and update the teacher
TF is not a parameter of TLBO algorithm. Several tests with various 30: end for
values of TF showed that the algorithm is more successful when 31: return The best solution (teacher)

TF take a value between 1 and 2. The best manner to improve the

algorithm is to set TF as a random variable between 1 and 2 during 5.5. Sequential Linear Programming and Simplex
the iterations [44].
A mathematical model of the optimization problem has been
5.4.1.2. Learning phase. In this phase, repeatedly two students introduced. According to that, a set of design variables (x) has
of the population are pulled randomly, they change knowledge been defined according to the stated optimization problem. The
between each other by a mutual interaction. Every student learn optimization methodology proposed states an iterative procedure
new “things” from the other that has more knowledge (having a where the design variables are modified at each iteration as:
good fitness). This phenomenon was mathematically described by xk = xk−1 + xk , k ≥ 1; (50)
Rao et al. [42] and expresses as:
⎧ being k the index of the iteration. Thus, the main goal of the
⎪ (t) (t) (t)
⎪ xi + rand × (xi − xj ) if optimization algorithm consists of finding the value of xk that
⎪
⎪
⎪
⎪ , xi
(t) minimizes the objective function and satisfies the constraints
⎪
⎪
⎪
⎨ , is imposed according to the definitions proposed in Sections 3.3 and
(t+1) , better 3.4. In this paper, the authors propose an optimization algorithm
xi = (49)
⎪
⎪ , than based on the use of linearized approximations. Thus, the objec-
⎪
⎪
⎪
⎪
(t)
, xj tive function can be approximated by using the first order Taylor
⎪
⎪ expansions as:
⎪
⎩ (t) (t) (t) 
xi + rand × (xj − xi ) otherwise
k k−1 ∂F 
F(x ) ≈ F(x )+ xk (51)
∂x xk−1
5.4.2. Algorithmic implementation
In this algorithm, each subject xid represents one design variable, Thus, the minimization of the first order approximation can be
a set of variable gives a student i.e. a solution, the teacher is always treated as the minimization of the right hand side term since F(xk−1 )
the best solution provided by the TLBO. In both phases, during the is constant in iteration k. On the other hand, the constraints of the
iterations, it is always necessary to keep the good student and to optimization model must be also satisfied. Consequently,
reject the one who has a low fitness.
gj (xk ) ≤ 0, j = 1, . . ., m (52)
The code in Algorithm 4 shows the design optimization process
of an axial pump by TLBO. where m is the total number of constraints. In practice, design con-
straints are usually nonlinear but, following the same idea used
Algorithm 4. Design approach using Teaching-Learning-Based
in the approximation of the objective function, first order Taylor
Optimization Algorithm
expansions are used as:

∂gj 
1: Start
k k−1
2: Input parameters: gj (x ) ≈ gj (x )+ xk , j = 1, . . ., m (53)
- Space search, max iteration, population size, number of ∂x xk−1
dimensions(variable)
- Design constant: range of volume flow rate, the head H, If second and higher order terms of the Taylor expansions are
rotational speed N, the density and the viscosity of fluid, the not considered, the verification of the constraints proposed can be
roughness, the slack distance between the tip of the rotor
stated as:
blades and the hull, the maximum thickness of rotor, type of 
vortex, number of the volume flow rate and the section ∂gj 
xk ≥ −gj (xk−1 ) (54)
∂x xk−1
3: Initialize the population randomly
4: Evaluate the population:
5: - Generate the geometry (inverse design)
The linearized problem can be solved by including slack variables
6: - Check the geometrical constraints with the penalty method
7: if no violation found then in the design constraints as:

∂gj 
8: - Analyze the performances (direct design)
9: - Check the mechanical constraints (nominal volume flow xk + hj = −gj (xk−1 ) (55)
rate) with the penalty method (compute the total fitness) ∂x xk−1
Table 3 Table 4
Specifications of the pump. Hypotheses and specifications of inverse and direct design.

Characterstic Pump mono-rotor Theoretical parameters Pump mono-rotor

H (m) 6 Roughness 2.0e−5 (m)

Qvmin (m3 /h) 360 Type of vortex FV-CV-FCV
Qvmax (m3 /h) 1800 Slack distance between the tip and the hull 5.0e−4 (m)
N (rpm) 1500 Radial equilibrium Simplified
Density of fluid (kg/m3 ) 1000 Fairing Without
max thickness of the airfoil rotor 0.012
Number of sections 20
where hj ≥ 0 are the slack variables. This algorithm is based on the Number of the volumetric flow rate 30
validity of the first order approximation. Consequently, the side
constraints of (xk ) must be restricted to small values close to zero
in order to satisfy the validity of the first order approximation. The Table 5
definition of the side constraints is usually stated as a centered Space of search for the designs variables.

interval around zero as: Side Qv N Ri1 Re1 Ri2 Re2 Di De Z

k
constraints
−x0 ≤ x ≤ x0 (56)
Min 0.1 1000 0.03 0.105 0.03 0.105 0.3 0.3 3
where x0 is usually defined as: Max 0.5 1900 0.100 0.150 0.100 0.150 0.7 0.7 20
Unit m3 /s rpm m m m m – – –
x0 = p(xmax − xmin ) (57)

being p a predefined percentage usually defined in the interval

[1%, 5%]. This percentage can be slightly modified during the opti-
mization process if needed. In addition, the definition of the side
constraints must satisfy:

xmin − xk−1 ≤ xk ≤ xmax − xk−1 (58)

to verify the side constraints of the updated design variables of the

problem. This definition of side constraints is usually called moving
limits in the literature [45–48] since the interval defined by the side
constraints of the linearized problem are usually centered around
the design variables, which are modified during the optimization
problem. The linearized optimization problem, including the side
constraints of the design variables, can be stated as:

∂F 
Minimize xk
∂x xk−1
such 
∂gj 
, that xk + hj = −gj (xk−1 )
∂x xk−1 (59)
j = 1, . . ., m Fig. 6. Reference rotor CAD.
verifying :
−x0 ≤ xk ≤ x0
truncation of the approximation series used and due to the proper-
xmin − xk−1 ≤ xk ≤ xmax − xk−1
ties of the resulting algorithms about round-off errors propagation.
being F the objective function defined according to Section 3.3.2. In addition, they usually require a large number of analyses of the
This linearized optimization problem can be directly solved by problem with different input data, which involves in practice con-
applying the Simplex algorithm proposed by G. Dantzig [49]. This siderable computing effort, specially in CPU time. In this paper, the
algorithm obtains the most adequate value of xk at each iteration. authors have developed analytically the first order derivatives of
Thus, the design variables of the problem can be updated according the objective function proposed in (16) and (17) and the first order
to Section 3.2. The complete iterative optimization process pro- derivatives of the constraints proposed in Section 3.4. The authors
posed in this paper to solve the optimization of axial turbomachines have also computed analytically the derivatives with respect to all
is usually called in the bibliography Sequential Linear Programming the variables of the problem in order to facilitate the definition of
(SLP) [45–48]. other different optimization problems. Thus, other different opti-
mization problems according to different practical considerations
5.5.1. Sensitivity analysis can be easily stated if needed by adequately selecting the design
The optimization algorithm proposed in Section 5.5 requires variables and redefining the objective functions and the design
the computation of first order derivatives of the objective function constraints.
and constraints with respect to the design variables of the prob-
lem. These derivatives can be computed by numerical methods like
finite differences. However, this kind of methods are not advisable 6. Optimization of axial pumps
if the numerical formulation for the analysis of the problem is avail-
able. In this case, the use of analytical methods of derivation is the The axial pump considered as reference in this paper consists of
most effective technique. Numerical methods of derivation are not a non shrouded mono rotor, presenting the specifications given in
usually advisable since they introduce sources of error due to the Tables 3 and 4 (Fig. 6).
 Fig. 7. Case 1–scenario 1: (a) reference rotor CAD, (b) TLBO rotor CAD and (c) SLP rotor CAD.

Table 6
Summary of cases and scenarios optimization studies.

Case 1 2  1.25Qnom 
Objective max (nom ) max (Qv ) dQv
0.75Qnom

Scenario 1 2 3 4 5 1 2 3

N (rpm) 1500 1500 1500 1500 [1000,1900] 1500 1500 [1000,1900]

Qv (m3 /h) [540,1800] [540,1800] [540,1800] [540,1800] 1080 [540,1800] [540,1800] 1080
Ri1 (m) [0.03,0.1] 0.075 0.075 0.075 0.075 [0.03,0.1] 0.075 0.075
Re1 (m) [0.105,0.15] 0.140 [0.126,0.154] 0.140 [0.126,0.154] [0.105,0.15] 0.140 [0.126,0.154]
Ri2 (m) [0.03,0.1] 0.075 0.075 0.075 0.075 [0.03,0.1] 0.075 0.075
Re2 (m) [0.105,0.15] 0.140 [0.126,0.154] 0.140 [0.126,0.154] [0.105,0.15] 0.140 [0.126,0.154]
Di [0.3,0.7] [0.3,0.7] [0.3,0.7] [0.3,0.7] [0.3,0.7] [0.3,0.7] [0.3,0.7] [0.3,0.7]
De [0.3,0.7] [0.3,0.7] [0.3,0.7] [0.3,0.7] [0.3,0.7] [0.3,0.7] [0.3,0.7] [0.3,0.7]
Z [3,20] [3,20] [3,20] [3,20] [3,20] [3,20] [3,20] [3,20]
Vortex FV FV FV CV-FCV FV-CV-FCV CV CV CV

Table 7
Results of different methods: case 1–scenario 1.

F Reference TLBO CS

best worst ave std best worst ave std

Efficiency (%) 60.15637 70.0543 70.0514 70.05325 8.7591e−04 70.0474 70.0038 70.02349 1.6508e−02

F PSO GA SLP

best worst ave std best worst ave std

Efficiency (%) 70.0546 70.0398 70.05176 5.9939e−03 69.4062 68.6234 69.05351 2.9060e−01 68.82087

6.1. Data relative to the algorithmic implementation allowed. The same computational devices are used as well (CPU
Intel Core 2 Duo E5400, 3 GB RAM, 2.9 GHz).
For all presented optimization algorithms, the same variable
constraints (side constraints) are considered. The values of these
6.2. Results and discussions
variables are given in Table 5.
Table 6 shows all the studied cases and scenarios.
6.2.1. Case one: maximization of total efficiency
The maximum number of function evaluation is fixed as 60,000
6.2.1.1. Scenario one. This first Scenario contains eight decision
for all the algorithms and the run was repeated for 10 times for
variables, namely volume flow rate, number of blades, internal and
obtaining the best solution, the worst solution, average (ave) and
external radius at the inlet and the outlet of the rotor, diffusion fac-
standard deviation (std). The population size and number of itera-
tors at the hub and the tip of the rotor, using free vortex strategy.
tions were selected according to the maximum function evaluation
The optimal pump variables with the obtained nominal efficien-

Table 8
Comparison of the optimal solutions: case 1–scenario 1.

Methods Qv (m3 /s) Ri1 (m) Re1 (m) Ri2 (m) Re2 (m) Di De Z

Reference 0.3 0.075 0.14 0.075 0.14 0.68 0.32 4

TLBO 0.269755 0.0300002 0.149148 0.0674282 0.15 0.699999 0.347475 3
CS 0.27028 0.0302294 0.148827 0.0677522 0.149998 0.699927 0.343204 3
PSO 0.269788 0.0300009 0.149043 0.0673856 0.15 0.699992 0.346976 3
GA 0.282407 0.0399075 0.147405 0.0731936 0.149789 0.585075 0.33793 4
SLP 0.281981 0.0483276 0.141835 0.0739864 0.15 0.7 0.373599 4
 Fig. 9. Comparison of best Fitness’s variation with the iterations.
Fig. 8. Comparison of the computational time: case 1–scenario 1.

cies obtained by the proposed optimization methods are given in able, which implies the necessity of an iterative calculation for the
Tables 7 and 8: additional variable. CS and PSO require relatively the same com-
Compared to the results of the reference machine, we record a putational time, TLBO records a little decrease of computational
clear increase of the total efficiency: about 10% of increase for TLBO, time and finally GA that is considered as faster algorithm to the
PSO and CS, and about 9% and 8% of increase is obtained by GA and previous algorithms. Modified penalty technique can decrease the
SLP respectively. computational time with an indeterminate rate because the ran-
We notice also that for all the explored methods, the opti- dom nature of used algorithms, after many tests, the value of this
mal exterior radius, at the inlet and at the outlet of the blade, is rate is between 14% and 50%.
close to its maximal value (Fig. 7). The obtained results confirm On the other hand, the fast convergence of the meta heuristics
that the efficiency increases with the decrease of the volume flow algorithms constitutes another advantage compared to the local
rate. Moreover, the diffusion factor at the rotor’s hub is raised, but methods (Fig. 9). Despite the fact that they are initiated with a
remains as low as possible at the tip. random population, contrary to the SLP, which is initiated with a
The average computational time is compared for the various reference solution, we notice that PSO and TLBO converge respec-
optimization methods and represented in the following histogram tively around 40th and 70th iterations, whereas the GA and CS
(Fig. 8). exceeds half the number of iteration before convergence.
It is shown that the SLP algorithm is by far, the heaviest because, Fig. 25a shows the variation of the efficiency as a function of the
besides, this local method, is restricted by a number decision vari- volume flow rate. The optimized machines which have the higher

Fig. 10. Case 1–scenario 1: performances.

 Fig. 11. Comparison of the performance prediction of proposed approach with CFD: case 1–scenario 1.

Fig. 12. Case 1–scenario 2: A) Reference rotor CAD, B) CS rotor CAD, C) GA rotor CAD.

Table 9
Results of different methods: case 1–scenario 2.

F TLBO CS

best worst ave std best worst ave std

Efficiency (%) 62.0627 62.0619 62.06248 3.3599e−04 62.0626 62.0624 62.06247 6.7495e−05

F PSO GA

best worst ave std best worst ave std

Efficiency (%) 62.0627 62.0621 62.06242 2.7809e−04 62.056 61.99784 62.0230 2.2565e−02

Table 10
Comparison of the optimal solutions: case 1–scenario 2.

Methods Qv (m3 /s) Di De Z

TLBO 0.284529 0.699999 0.397594 9

CS 0.284537 0.699955 0.397678 8
PSO 0.284529 0.699999 0.397594 9
GA 0.284657 0.699962 0.398492 9

Table 11
Results of different methods: case 1–scenario 3.

F TLBO CS

best worst ave std best worst ave std

Efficiency (%) 69.144 69.1395 69.13995 1.4230e−03 69.1393 69.1374 69.13855 7.6485e−04

F PSO GA

best worst ave std best worst ave std

Efficiency (%) 69.144 69.1438 69.14398 6.3246e−05 68.5954 67.6322 68.11708 2.9908e−01

nominal efficiency present also a higher off-design efficiency com- on all the range of volume flow rate. i.e. the cavitation criterion is
pared to the reference pump. respected.
In this paper, we are exclusively interested on the improvement Another positive point of optimized pumps, its consumed power
of the total nominal efficiency, the cavitation criterion was not con- is less than reference pump for all volumetric flow rate range
sidered. In Fig. 10c and d, it is noticed that for all machines, the (Fig. 10e and f).
required net pressure suction head (NPSHC) is inferior to the NPSH
 Table 12
Comparison of the optimal solutions: case 1–scenario 3.

Methods Qv (m3 /s) Re1 (m) Re2 (m) Di De Z

TLBO 0.269755 0.153302 0.154 0.7 0.492199 11

CS 0.250453 0.153287 0.154 0.699995 0.491273 11
PSO 0.282407 0.153298 0.154 07 0.492113 11
GA 0.258843 0.150389 0.152934 0.6992 0.510518 11

Tables 9 and 10 and Fig. 13 show the convergence of the

Fig. 13. Comparison of best fitness’s variation with the iterations: case 1–scenario algorithms to nearly the same solution, with an increase in total
2.
efficiency by almost 2% relative to the reference machine. The
performance curves presented in Fig. 14 show that the obtained
The CFD method using Ansys-Fluent is used to predict the per- machines are similar. Further, Fig. 14e and f shows that the power
formance of two first best optimized machines (i.e. PSO and TLBO) consumed by these machines is less than the power consumed by
as well as the reference machine. A steady flow is imposed using the reference machine. Figs. 14c and d shows that the cavitation cri-
Moving Reference frame with periodic boundary condition. The terion is satisfied. Finally, the geometries are relevant in terms of
turbulence model K − ω SST is chosen for solving three-dimensional design as shown in Fig. 12. Observe the difference in the rotor axial
Reynolds-averaged Navier–Stokes equations. For the design point distance between the optimized machines where the constraints
neighborhood the results of CFD are very acceptable, is so closed are respected and the reference machine where they are not.
to the analysis of the proposed approaches (Fig. 11). For the refer-
ence machine, the average relative error is about 2.11% and 13.60% 6.2.1.3. Scenario three. In the third scenario, only the hub radius
for efficiency and head respectively. For TLBO’s machine, the aver- are kept constant, the external radius vary in a range of −10% to
age relative error is about 2.86% and 1.96% for efficiency and head +10% of the external radius of the reference machine.
respectively. For the PSO’s machine, the average relative error is Tables 11 and 12 and Fig. 16 show that the rate of increase in
about 2.96% and 3.45% for efficiency and head respectively. total efficiency is about 9% for the TLBO, PSO and CS, and less then
9% for GA. Notice that the radius tend to their maximum value in
6.2.1.2. Scenario two. In the second scenario, the inlet and outlet the search space as well as the diffusion factor at the hub. Notice
radius are kept constant. For the following cases and starting from also that there is also an arguably high number of blades compared
this scenario, the SLP method is not explored. to the other scenarios.

Fig. 14. Case 1–scenario 2: performances.

 Fig. 15. Case 1–scenario 3: (a) reference rotor CAD, (b) PSO rotor CAD and (c) GA rotor CAD.

(Fig. 17d and c respectively), which verifies the cavitation criterion.

The curves of torque and power present a coincidence between
the PSO, CS and TLBO as shown in Fig. 17f and e. Finally for all
algorithms, the power is always lower than that of the reference
machine. The geometries given by PSO and GA are comparable as
shown in Fig. 15.

6.2.1.4. Scenario four. The fourth scenario is similar to the second

one concerning the decision variables except that optimization is
Fig. 16. Comparison of best fitness’s variation with the iterations: case 1–scenario done for two types of vortex, namely constant and forced (Fig. 18).
3. Tables 13 and 14 and Fig. 19 show that the use of forced and
constant vortex can slightly improve the performance, particularly
Except for GA, Fig. 17 shows homogeneity and coincidence of the forced vortex. For the diffusion factor, the results show that the
the efficiency curve (Fig. 17a), the head (Fig. 17b), NPSH and NPSHC obtained machine can operate with a marginally high values, either

Fig. 17. Case 1–scenario 3: performances.

 Fig. 18. Case 1–scenario 4: (a) reference rotor CAD, (b) CS-CV rotor CAD and (c) CS-FCV rotor CAD.

Table 13
Results of different methods: case 1–scenario 4.

F TLBO CS

best worst ave std best worst ave std

Efficiency (%) CV 63.3594 63.3527 63.35793 1.9282e−03 63.3595 63.3524 63.3565 3.4351e−03
FCV 63.6206 63.614 63.61993 2.0838e−03 63.6201 63.614 63.61604 2.7048e−03

F PSO GA

best worst ave std best worst ave std

Efficiency (%) CV 63.3594 63.3527 63.35739 2.5049e−03 63.3595 63.3149 63.33529 1.1766e−02
FCV 63.6206 63.614 63.61925 2.3206e−03 63.6132 63.6108 63.61202 8.5219e−04

Table 14 6.2.1.5. Scenario five. In the fifth scenario, the flow rate is kept con-
Comparison of the optimal solutions: case 1–scenario 4.
stant and the rotational speed variable and the hub radius are also
Methods Qv (m3 /s) Di De Z fixed. The outer radius is variables within a range of −10% to 10%
TLBO CV 0.294284 0.698088 0.516567 11 of the tip radius of the reference machine. The machines are opti-
FCV 0.296991 0.55635 0.684712 9 mized for the three types of vortex (free, forced and constant). In
this scenario, presented in Table 15 and Fig. 21, the efficiency can be
CS CV 0.294174 0.698246 0.515346 4
FCV 0.296967 0.556322 0.684121 6 improved by 14% and 13%. Notice that the machines operate with
the maximum rotational speed Table 16. Moreover, all tip radius
PSO CV 0.29425 0.69814 0.51619 11
FCV 0.297017 0.556317 0.685256 9
and are close to same solution that is the maximum allowed value.
In this scenario, the relevant solutions are present comparable per-
GA CV 0.294325 0.699532 0.51403 4
formances as shown in Fig. 22.
FCV 0.297048 0.541793 0.690215 4

6.2.2. Case two: maximization the operation area under the

efficiency curve
In this case we propose a new objective function for the design of
a turbomachine, it consists on maximizing the area under the per-
formance curves, in our case it is the efficiency curves. A numerical
integration was used on a flow range of −25% to 25% of the nominal
rate. A constraint is added to the head, which is the head variation
within this range, and that should not exceed 15% of the head of
the nominal point. Only constant vortex has been considered in
this case.
Fig. 19. Comparison of best fitness’s variation with the iterations for different vor-
tex: case 1–scenario 4.
6.2.2.1. Scenario one. In this scenario all the parameters are consid-
ered variables. Tables 17 and 18 and Fig. 23 show an increase of this
with forced or constant vortex. Fig. 20a illustrates the improvement area by 60.590% and 60.935% with TLBO and PSO respectively and
in the efficiency between the different types of vortex. Notice that by 60.558% and 59.963% with CS and GA respectively. It is found that
the optimized machines provide a little different charges depend- this objective can increase the design efficiency. Fig. 24a shows the
ing on the vortex type, also the forced vortex and constant vortex difference in the neighborhood area of nominal flow rate between
gives the highest and lowest heads respectively than the free vor- the reference machine and the optimized machine. The constraint
tex, see Fig. 20b. of variation limitation of the head is well respected as presented in
Fig. 20c and d shows that the cavitation criterion is respected. Fig. 24b. Fig. 24e and f shows that this objective, in which the effi-
Fig. 20e and f confirms this observation for the power consump- ciency variation range is wider, causes higher power consumption,
tion. In the other side, at nominal point, all the optimized machines and on a proportionate basis with the flow rate.
present less power compared to the reference machine for all con- Using the same CFD condition of the first case, in this second
sidered types of vortex. case and for the design point neighborhood of TLBO’s machine, the
 Fig. 20. Case 1–scenario 4: performances.

Table 15
Results of different methods: case 1–scenario 5.

F TLBO CS

best worst ave std best worst ave std

Efficiency (%) FV 73.4246 73.4246 73.4246 1.4980e−14 73.4246 73.4246 73.4246 1.4980e−14
CV 74.6334 74.6334 74.6334 1.4980e−14 74.6334 74.6334 74.6334 1.4980e−14
FCV 74.734 74.734 74.734 1.4980e−14 74.734 74.734 74.734 1.4980e−14

F PSO GA

best worst ave std best worst ave std

Efficiency (%) FV 73.4246 73.4246 73.4246 1.4980e−14 73.3955 73.3018 73.35107 3.0256e−02
CV 74.6334 73.5961 74.52967 3.2802e−01 74.63295 74.6177 74.6247 5.7797e−03
FCV 74.734 73.8391 74.64451 2.8299e−01 74.7305 74.7157 74.72102 4.1635e−03

results of CFD are acceptable, the average relative error is about

1.93% and 2.274% for efficiency and head respectively.

6.2.2.2. Scenario two. This scenario has the same considerations

with the first scenario except that all the radius are kept constant
(Fig. 26). Table 19 and Fig. 27 show that the efficiency varia-
tion range may be extended even with a small variation about
an increase of 7.66%. Almost all algorithms converge to the same
solution and all the obtained solutions are geometrically relevant
(Table 20). As for the previous scenario, Fig. 28 shows that the maxi-
Fig. 21. Comparison of best fitness’s variation with the iterations for different vor- mization of the efficiency range does not always lead to a maximum
tex: case 1–scenario 5.
nominal efficiency.
 Fig. 22. Case 1–scenario 5: performances.

Table 16
Comparison of the optimal solutions: case 1–scenario 5.

Methods N (rpm) Re1 (m) Re2 (m) Di De Z

TLBO FV 1900 0.154 0.154 0.7 0.3 4

CV 1900 0.154 0.154 0.684502 0.371043 3
FCV 1900 0.154 0.154 0.426198 0.513194 3

CS FV 1900 0.154 0.154 0.7 0.3 5

CV 1900 0.154 0.154 0.684502 0.371043 3
FCV 1900 0.154 0.154 0.426198 0.513194 3

PSO FV 1900 0.154 0.154 0.7 0.3 7

CV 1900 0.154 0.154 0.684502 0.371043 3
FCV 1900 0.154 0.154 0.426198 0.513194 3

GA FV 1900 0.153843 0.153921 0.699947 0.30039 5

CV 1900 0.153999 0.153996 0.683323 0.371847 5
FCV 1900 0.153843 0.153921 0.699947 0.30039 5

Table 17
Results of different methods: case 2–scenario 1.

F Reference TLBO CS

best worst ave std best worst ave std

Area 0.05931 0.095246 0.0947645 0.09509566 1.4530e−04 0.0952269 0.0926147 0.09468864 7.5626e−04

F PSO GA

best worst ave std best worst ave std

Area 0.0954504 0.0763508 0.09299933 5.9877e−03 0.0948743 0.0934098 0.09450688 4.3977e−04

Table 18
Comparison of the optimal solutions: case 2–scenario 1.

Methods Qv (m3 /s) Ri1 (m) Re1 (m) Ri2 (m) Re2 (m) Di De Z

Reference 0.3 0.075 0.14 0.075 0.14 0.68 0.32 4

TLBO 0.441313 0.030153 0.140272 0.0348952 0.149991 0.691658 0.593822 5
CS 0.442526 0.030381 0.144615 0.0351488 0.149994 0.646998 0.638884 4
PSO 0.441199 0.03 0.145582 0.0364082 0.15 0.699095 0.6172 4
GA 0.445684 0.0317479 0.144948 0.0376835 0.149819 0.638295 0.621847 6

Table 19
Results of different methods: case 2–scenario 2.

F TLBO CS

best worst ave std best worst ave std

Area 0.0638551 0.0638526 0.06385324 9.9017e−07 0.0638527 0.0638386 0.06384904 4.1358e−06

F PSO GA

best worst ave std best worst ave std

Area 0.0638551 0.0638471 0.06385336 2.3477e−06 0.0638449 0.0637933 0.06382547 1.8558e−05

Table 23 summarizes the performance ranking of the solutions

found and the stability of the algorithms used. For each scenario,
the algorithms were classified from 1 to 4 according to their best
fitness, worst fitness, average fitness and standard deviation. After
that, the sum of the rank was computed for each algorithm and
for each statistical parameter in order to find the nearness of the
algorithm to the ideal one (that presents the first rank for all sce-
narios). Finally, the average rank of each method was calculated to
define the overall rank. Based on this, TLBO present the best stabil-
ity (the nearest one), followed by CS and PSO, and finally GA. We
Fig. 23. Comparison of best Fitness’s variation with the iterations: case 2–scenario note that the ranking is correspond controlling parameters of the
1. algorithm which the TLBO does not need any parameters flowing
by CS that has fewer parameters (only one, the empty nest rate Pa),
Table 20 in the third rank we found PSO that uses three parameters (inertia
Comparison of the optimal solutions: case 2–scenario 2. weight, cognitive and social parameters) and final GA which works
Methods Qv (m3 /s) Di De Z with more than three parameters (crossover rate and mutation rate,
scale, ratio and shrink).
TLBO 0.426242 0.671216 0.478967 9
CS 0.426466 0.657295 0.480458 8
PSO 0.42624 0.671392 0.478946 9 6.4. Comparison of the proposed approach with some existing
GA 0.426003 0.650969 0.486162 4
machines

To demonstrate the effectiveness of the proposed framework,

6.2.2.3. Scenario three. In this scenario only the flow and the hub
some baseline pumps (pump A [50], pump B and pump C [25])
radius are kept constant. The tip radius varies within a range
from literature are compared with optimized machines using TLBO
of −10% to + 10% of the tip radius of the reference machine.
under same specifications. The strategy of comparison is based on
Tables 21 and 22 and Fig. 29 show that the algorithms converge to
two scenarios, the first (S1) is to fix the maximum outer radius
the approximate solution, and the rate of increase is lower than the
constraint equal to the value of the reference pump’s radius, the
first scenario and higher than the second one (9.56%). Concerning
second one (S2), is to extend maximum side constraint radius with
the performance, a high rotational speed can increase the nominal
+10% of the reference value. For both scenarios and for pump A, the
efficiency and the area with a lower power consumption, as shown
objectives are the maximization of nominal total efficiency (main
in Fig. 30.
formulation in pervious case 1 (C1)) and area of total efficiency-flow
rate (main formulation in pervious case 2 (C2)). The maximization
6.3. Recapitulation of algorithms behavior of the nominal hydraulic efficiency is considered for pump B and C.
Table 24 shows the optimal geometrical parameters and their
TLBO, CS and PSO record a higher and approximate solution, and operating conditions compared with the reference pump A. In the
the error between each other is between 10−4 and 10−6 . Compared first objective and for scenario S1, the improvement of nominal
with previous methods, almost for all cases, GA was recorded hav- efficiency is very small (about 0.01%) but the respecting of the head
ing the worst behavior whether of the way of the convergence or in point of design is clear. With the second scenario, the differ-
the solution (namely with more decision variable) may be because ence between optimized pump and the reference one is about 7%.
it needs more function evaluations or the parameters of this algo- In addition, the optimized machines have a good distribution of
rithm do not well adjusted which is considered as an inconvenient, the performance curves, as shown in Fig. 31, the curves quality is
because there is no rule to determine this parameter, they are still recorded for all flow rate range with the efficiency, and for neigh-
strongly linked to the optimization problem. SLP is limited with borhoods design point flow rate with the head curve. For the second
hard problem like that considered in this paper because of its imple- objective, for both scenarios the difference between the optimized
mentation complexity and its local solution. pumps area and reference pump area is clear, it is about 8% for first
 Fig. 24. Case 2–scenario 1: performances.

Fig. 25. Comparison of the performance prediction of proposed approach with CFD: case 2–scenario 1.

Fig. 26. Case 2–scenario 2: (a) reference rotor CAD, (b) PSO rotor CAD and (c) TLBO rotor CAD.
Fig. 27. Comparison of best fitness’s variation with the iterations: case 2–scenario Fig. 29. Comparison of best fitness’s variation with the iterations: case 2–scenario
2. 3.

scenario and about 27.6% for the second. In this comparison of the Numerically, these comparisons prove the effectiveness of the
two objectives, results are very close with a little improvement on proposed approach in incompressible axial turbomachine with any
the left of design point for the second objective. operating condition (head, rotational speed, volume flow rate).
Another comparison with pump B of Ref. [25] is shown in
Table 25. The hydraulic efficiency improvement is about 7% and 7. Conclusion
18% for the scenarios S1 and S2 respectively. Fig. 32 displays a good
distribution of hydraulic efficiency with great precision in design In this paper an assessment study of some meta-heuristics
point, (i.e. 0.6 m3 /s) and wide range. applied to the design optimization of incompressible axial turbo-
Last comparison is carried out for pump C of latter reference. machines is performed. A design strategy based on the inverse and
The pump is of the rotor–stator type and we focus only on rotor direct approaches was developed using bio and socio-inspired algo-
optimization, i.e. during the design optimization, the stator param- rithms as well as a local search algorithm. The aim is to improve the
eters design is kept constant with the same value of that of the volume flow rate efficiency while respecting a large number of con-
reference pump. Despite that, the proposed approach recorded the straints, it is improved about 10 ± 0.288% in case 1–scenario 1 and
improvement of the efficiency with about 6.74% and 8.33% for the about 14% in case 1–scenario 5. Numerically, the results in the sec-
scenarios S1 and S2 respectively, see Table 26. The wide operating ond formulation show that the range of efficiency/flow rate can be
range is remarkable in Fig. 33. The design could be further improved increased widely (about 60 ± 0.669% in case 2–scenario 1) with a
inclunding the stator parameters in variables design. little increase of nominal efficiency of reference machine. Besides,

Fig. 28. Case 2–scenario 2: performances.

Table 21
Results of different methods: case 2–scenario 3.

F TLBO CS

best worst ave std best worst ave std

Area 0.0649398 0.0642566 0.0644795 2.8594e−04 0.0648519 0.064356 0.06456078 1.4932e−04

F PSO GA

best worst ave std best worst ave std

Area 0.0649804 0.0648374 0.06494251 4.2905e−05 0.0646773 0.0641161 0.06442932 1.9858e−04

Table 22
Comparison of the optimal solutions: case 2–scenario 3.

Methods N (rpm) Re1 (m) Re2 (m) Di De Z

TLBO 1849 0.1523672 0.153993 0.698896 0.501064 4

CS 1856 0.152128 0.1538 0.694437 0.500728 4
PSO 1899 0.152123 0.153882 0.699758 0.470388 3
GA 1841 0.149659 0.151783 0.696984 0.495498 5

Table 23
Performance and stability ranking of algorithms for all cases studied.

Methods Fitness Case 1 Case 2 Sum Rank Average rank Final rank

S1 S2 S3 S4 S5 S1 S2 S3

FV FV FV CV FCV FV CV FCV CV CV CV

TLBO best 2 1 1 3 1 1 1 1 2 1 2 16 2 1.25 1

worst 1 3 2 1 1 1 1 1 1 1 3 16 1
ave 1 1 2 1 1 1 1 1 1 2 3 15 1
std 1 3 3 1 2 1 1 1 1 1 4 19 1

CS best 3 3 3 1 3 1 1 1 3 3 3 25 3 2.25 2
worst 3 1 3 3 1 1 1 1 3 3 2 22 2
ave 3 2 3 3 3 1 1 1 2 3 2 24 2
std 3 1 2 3 4 1 1 1 3 3 2 24 2

PSO best 1 1 1 3 1 1 1 1 1 1 1 13 1 2.5 3

worst 2 2 1 1 1 1 4 4 4 2 1 23 3
ave 2 3 1 2 2 1 4 4 4 1 1 25 3
std 2 2 1 2 3 1 4 4 4 2 1 26 3

GA best 4 4 4 1 4 4 4 4 4 4 4 41 4 4 4
worst 4 4 4 4 4 4 3 3 2 4 4 40 4
ave 4 4 4 4 4 4 3 3 3 4 4 41 4
std 4 4 4 4 1 4 3 3 2 4 3 36 4

Table 24
Comparison of the optimal solution: pump A [50], N = 1400 rpm, Qv = 1000 m3 /h, H = 8 m.

Methods Efficiency (%) Area H (m) Ri1 (m) Re1 (m) Ri2 (m) Re2 (m) Di De Z

Reference 65 0.0695 7.812 0.0705 0.157 0.0705 0.157 – – 6

TLBO C1 S1 65.78 – 8.00 0.0300249 0.149684 0.0613356 0.157 0.699 0.677 3
TLBO C2 S1 – 0.08867 8.00 0.0308582 0.149755 0.062324 0.157 0.699999 0.681804
TLBO C1 S2 72.13 – 8.00 0.0344461 0.169401 0.0667512 0.172699 0.699 0.638 3
TLBO C2 S2 – 0.09637 8.00 0.0416876 0.168948 0.0721562 0.172755 0.699989 0.6204773 3

Table 25
Comparison of the optimal solution: pump B [25], N = 1480 rpm, Qv = 2160 m3 /h, H = 6 m.

Methods Hydraulic efficiency (%) Ri1 (m) Re1 (m) Ri2 (m) Re2 (m) Di De Z

Reference 75 0.68 0.184 0.68 0.184 0.26 0.34 5

TLBO S1 82.90 0.03 0.182813 0.03 0.184 0.214246 0.699 3
TLBO S2 93.55 0.03 0.201887 0.03 0.2024 0.213066 0.699 3

Table 26
Comparison of the optimal solution: pump C [25], N = 980 rpm, Qv = 2160 m3 /h, H = 6 m.

Methods Hydraulic efficiency (%) Ri1 (m) Re1 (m) Ri2 (m) Re2 (m) Di De Z

Reference 86 0.105 0.217 0.105 0.217 0.5 0.3 5

TLBO S1 92.74 0.0750967 0.217 0.0750967 0.217 0.406172 0.7 8
TLBO S2 94.33 0.0336081 0.238697 0.0415088 0.238581 0.513954 0.580032 5
 Fig. 30. Case 2–scenario 3: performances.

Fig. 31. Pump A: performances.

Fig. 32. Pump B: performances.

an improvement of the efficiency was highlighted with regard to figuration is when all radius are considered variables, this is due to
the radius at the rotor inlet and outlet, also expressed when the the extension of side constraints.
exterior radius is high and the interior radius is low. The best con- A comparative study allowed us to highlight the robustness,
the flexibility and the implementation ease of the meta-heuristics
 Fig. 33. Pump C: performances.

in terms of efficiency improvement of turbomachines. The TLBO [16] D. Yanhui, W. Wenhua, F. Zhaolin, C. Ti, An introduction of aerodynamic shape
algorithm can be used efficiently for turbomachinery design optimization platform for compressor blade, in: ASME Turbo Expo 2016:
Turbomachinery Technical Conference and Exposition, American Society of
that includes a high number of several constraint types. For Mechanical Engineers, 2016, p. V02CT39A031.
constraint handling, the modified penalty technique can dramat- [17] Z. Li, X. Zheng, Review of design optimization methods for turbomachinery
ically decrease the computational time. With a large number of aerodynamics, Prog. Aerosp. Sci. 93 (2017) 1–23.
[18] W.-G. Li, NPSHr optimization of axial-flow pumps, J. Fluids Eng. 130 (7) (2008)
constraints, the optimization algorithms behave differently. The 074504.
algorithm without the controlling parameters is the most prof- [19] B.-J. Lin, C.-I. Hung, E. Tang, An optimal design of axial-flow fan blades by the
itable in this kind of optimization problem. It is proved also that machining method and an artificial neural network, Proc. Inst. Mech. Eng. Part
C: J. Mech. Eng. Sci. 216 (3) (2002) 367–376.
the TLBO, CS and PSO are more suitable than GA and SLP for the
[20] C.-H. Huang, C.-W. Gau, An optimal design for axial-flow fan blade: theoretical
problem treated in this work. and experimental studies, J. Mech. Sci. Technol. 26 (2) (2012) 427–436.
As an outlook, the design can be more improved with the con- [21] K.-Y. Lee, Y.-S. Choi, Y.-L. Kim, J.-H. Yun, Design of axial fan using inverse
design method, J. Mech. Sci. Technol. 22 (10) (2008) 1883–1888.
sideration of other criteria in the context of a multi-objective
[22] K.-S. Lee, K.-Y. Kim, A. Samad, Design optimization of low-speed axial flow fan
optimization approach by the implementation of the same meta- blade with three-dimensional RANS analysis, J. Mech. Sci. Technol. 22 (10)
heuristics. The computational time may be reduced substantially by (2008) 1864–1869.
the integration of substituted models, where the variables will be [23] P. Song, J. Sun, Blade shape optimization for transonic axial flow fan, J. Mech.
Sci. Technol. 29 (3) (2015) 931–938.
directly connected to the objective function and to the considered [24] R. Robert, N. Ricardo, B. Farid, Pompes rotodynamiques –
constraints. aérohydrodynamique des profils et aubages de pompes hélices, Techniques
de l’Ingénieur, Réf. BM4304 V1.
[25] R. Robert, N. Ricardo, B. Farid, Pompes rotodynamiques – dimensionnement
References et analyse des performances des pompes hélices, Techniques de l’Ingénieur,
Réf. BM4305.
[1] A.J. Stepanoff, Centrifugal and Axial Flow Pumps: Theory, Design, and [26] K. Deb, Multi-objective Optimization Using Evolutionary Algorithms, vol. 16,
Application, Krieger Publishing Company, 1957. John Wiley & Sons, 2001.
[2] J.P. Veres, Centrifugal and Axial Pump Design and Off-Design Performance [27] A. Ponsich, C. Azzaro-Pantel, S. Domenech, L. Pibouleau, Constraint handling
Prediction, National Aeronautics and Space Administration, 1994. strategies in genetic algorithms application to optimal batch plant design,
[3] Z. Dugao, Z. Jiang, Optimization design of an axial-flow fan used for mining Chem. Eng. Process.: Process Intensif. 47 (3) (2008) 420–434.
local-ventilation, Comput. Ind. Eng. 31 (3) (1996) 691–696. [28] K. Deb, An efficient constraint handling method for genetic algorithms,
[4] D. Desai, J. Zhou, Optimization design of an axial-flow fan used in local Comput. Methods Appl. Mech. Eng. 186 (2) (2000) 311–338.
ventilation in mining industry, in: ASME 2011 International Mechanical [29] A.E. Smith, D.W. Coit, T. Baeck, D. Fogel, Z. Michalewicz, Penalty functions
Engineering Congress and Exposition, American Society of Mechanical Handbook of Evolutionary Computation, vol. 97 (1, 1995, pp. C5.
Engineers, 2011, pp. 879–883. [30] J.H. Holland, Adaptation in Natural and Artificial Systems: An Introductory
[5] K. OH, HW, K. KIM, Conceptual design optimization of mixed-flow pump Analysis with Applications to Biology, Control, and Artificial Intelligence, vol.
impellers using mean streamline analysis, Proc. Inst. Mech. Eng. Part A: J. viii, Univ. Michigan Press, Oxford, England, 1975.
Power Energy 215 (A1) (2001) 133–138. [31] D. Goldberg, in: N.N. Schraudolph (Ed.), Genetic Algorithms in Search,
[6] D. Sorensen, J. Sorensen, Toward improved rotor-only axial fans – Part I: A Optimization, and Machine Learning, vol. 3, Addison-Wesley, Reading, MA,
numerically efficient aerodynamic model for arbitrary vortex flow, J. Fluids 1989, p. 1.
Eng. 122 (2) (2000) 318–323. [32] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution Programs,
[7] D. Sorensen, M. Thompson, J. Sorensen, Toward improved rotor-only axial Springer Science & Business Media, 1996.
fans – Part II: Design optimization for maximum efficiency, J. Fluids Eng. 122 [33] Y. Kaya, M. Uyar, et al., A Novel Crossover Operator for Genetic Algorithms:
(2) (2000) 324–329. Ring Crossover, 2011 arXiv:1105.0355.
[8] E. Benini, A. Toffolo, A parametric method for optimal design of [34] L. Song, NGPM – A NSGA-II Program in Matlab, College of Astronautics, North
two-dimensional cascades, Proc. Inst. Mech. Eng. Part A: J. Power Energy 215 Western Polytechnical University, China, 2011.
(4) (2001) 465–473. [35] J. Kenndy, R. Eberhart, Particle swarm optimization, Proceedings of IEEE
[9] A. Oyama, M.-S. Liou, S. Obayashi, Transonic axial-flow blade optimization: International Conference on Neural Networks, vol. 4 (1995) 1942–1948.
evolutionary algorithms/three-dimensional Navier–Stokes solver, J. Propuls. [36] Y. Shi, R. Eberhart, A modified particle swarm optimizer, Evolutionary
Power 20 (4) (2004) 612–619. Computation Proceedings, 1998. IEEE World Congress on Computational
[10] C.-h. Cho, S.-y. Cho, C. Kim, Development of an axial-type fan with an Intelligence. The 1998 IEEE International Conference on IEEE (1998) 69–73.
optimization method, Front. Energy Power Eng. China 3 (4) (2009) 414–422. [37] F. Van Den Bergh, An Analysis of Particle Swarm Optimizers, University of
[11] W. Tiow, K.F.C. Yiu, M. Zangeneh, Application of simulated annealing to Pretoria, 2006 (Ph.D. thesis).
inverse design of transonic turbomachinery cascades, Proc. Inst. Mech. Eng. [38] J. Wei, L. Guangbin, L. Dong, Elite particle swarm optimization with mutation,
Part A: J. Power Energy 216 (1) (2002) 59–73. System Simulation and Scientific Computing, 2008. ICSC 2008. Asia Simulation
[12] T.J. Carrigan, B.H. Dennis, Z.X. Han, B.P. Wang, Aerodynamic shape Conference – 7th International Conference on IEEE (2008) 800–803.
optimization of a vertical-axis wind turbine using differential evolution, ISRN [39] X.-S. Yang, S. Deb, Cuckoo search via lévy flights, Nature & Biologically
Renew. Energy 2012 (2012) 1–16. Inspired Computing, 2009. NaBIC 2009. World Congress on IEEE (2009)
[13] B.H. Dennis, G.S. Dulikravich, Z.-X. Han, Optimization of turbomachinery 210–214.
airfoils with a genetic/sequential-quadratic-programming algorithm, J. [40] X.-S. Yang, Cuckoo search and firefly algorithm: overview and analysis, in:
Propuls. Power 17 (5) (2001) 1123–1128. Cuckoo Search and Firefly Algorithm, Springer, 2014, pp. 1–26.
[14] S. Shahpar, A comparative study of optimisation methods for aerodynamic [41] X.-S. Yang, S. Deb, Engineering optimisation by cuckoo search, Int. J. Math.
design of turbomachinery blades, ASME Paper (2000-GT) (2000) 523. Model. Numer. Optim. 1 (4) (2010) 330–343.
[15] A. Safari, K. Hajikolaei, H. Lemu, G. Wang, A high-dimensional model [42] R.V. Rao, V.J. Savsani, D. Vakharia, Teaching-learning-based optimization: a
representation guided PSO methodology with application on compressor novel method for constrained mechanical design optimization problems,
airfoil shape optimization, in: ASME Turbo Expo 2016: Turbomachinery Comput. Aided Des. 43 (3) (2011) 303–315.
Technical Conference and Exposition, American Society of Mechanical
Engineers, 2016, p. V02CT45A013.
[43] R. Rao, V. Savsani, J. Balic, Teaching-learning-based optimization algorithm for [47] T.-Y. Chen, Calculation of the move limits for the sequential linear
unconstrained and constrained real-parameter optimization problems, Eng. programming method, Int. J. Numer. Methods Eng. 36 (15) (1993) 2661–2679.
Optim. 44 (12) (2012) 1447–1462. [48] L. Lamberti, C. Pappalettere, Move limits definition in structural optimization
[44] S.C. Satapathy, A. Naik, K. Parvathi, A teaching learning based optimization with sequential linear programming. Part I: Optimization algorithm, Comput.
based on orthogonal design for solving global optimization problems, Struct. 81 (4) (2003) 197–213.
SpringerPlus 2 (1) (2013) 130. [49] G. Dantzig, M. Thapa, Linear Programming 1: Introduction (Springer Series in
[45] R. Fletcher, Quadratic Programming, Practical Methods of Optimization, Operations Research and Financial Engineering), 1997.
second edition, 2000, pp. 229–258. [50] Manjunatha, J.R. Nataraj, Design and analysis of impeller blade for axial flow
[46] M.S. Bazaraa, H. Sherali, C. Shetty, Nonlinear Programming: Theory and pumps, Int. J. Eng. Res. Manag. Stud. (2015) 1–31.
Algorithms, John Wiley & Sons, New York, 1993.

View publication stats