Article

Design of Gas Cyclone Using Hybrid Particle Swarm

Optimization Algorithm
Xueli Shen and Daniel C. Ihenacho *

Department of Software Engineering, Liaoning Technical University, Huludao 12500, China;

[email protected]
* Correspondence: [email protected], Tel.: +86‐139‐4187‐4807

Abstract: The method of searching for an optimal solution inspired by nature is referred to as
particle swarm optimization. Differential evolution is a simple but effective EA for global
optimization since it has demonstrated strong convergence qualities and is relatively
straightforward to comprehend. The primary concerns of design engineers are that the traditional
technique used in the design process of a gas cyclone utilizes complex mathematical formulas and
a sensitivity approach to obtain relevant optimal design parameters. The motivation of this research
effort is based on the desire to simplify complex mathematical models and the sensitivity approach
for gas cyclone design with the use of an objective function, which is of the minimization type. The
process makes use of the initial population generated by the DE algorithm, and the stopping
criterion of DE is set as the fitness value. When the fitness value is not less than the current global
best, the DE population is taken over by PSO. For each iteration, the new velocity and position are
updated in every generation until the optimal solution is achieved. When using PSO independently,
the adoption of a hybridised particle swarm optimization method for the design of an optimum gas
Citation: Shen, X.; Ihenacho, D.C. cyclone produced better results, with an overall efficiency of 0.70, and with a low cost at the rate of
Design of Gas Cyclone Using Hybrid 230 cost/second.
Particle Swarm Optimization
Algorithm. Appl. Sci. 2021, 11, 9772. Keywords: particle swarm optimization (PSO); differential evolution (DE); gas cyclone; hybridised
https://doi.org/10.3390/app11209772
particle swarm optimization; evolutionary algorithm

Academic Editors: Peng‐Yeng Yin,

Ray‐I Chang, Youcef Gheraibia,
Ming‐Chin Chuang, Hua‐Yi Lin
and Jen‐Chun Lee
1. Introduction
Particle swarm optimization (PSO) and differential evolution (DE) are two stochastic,
Received: 4 September 2021 population‐based optimization EAs in evolutionary algorithms.
Accepted: 12 October 2021 PSO was developed by Kennedy and Eberhart and was originally intended to
Published: 19 October 2021 simulate social behaviour. Every solution in PSO is a “bird” in the search space [1].
However, when the technique is implemented, it is referred to as a particle. All of the
Publisher’s Note: MDPI stays
particles have fitness values that the fitness function evaluates in order to optimize them,
neutral with regard to jurisdictional
as well as velocities that control their flight. The particles navigate through the problem
claims in published maps and
space by following the current best particles. Although the PSO algorithm has attracted a
institutional affiliations.
lot of attention in the last decade, it unfortunately has a premature convergence issue,
which is common in complicated optimization problems. To improve PSO’s search
performance, certain strategies for adjusting parameters such as inertia weights and
Copyright: © 2021 by the authors.
acceleration coefficients have been developed.
Licensee MDPI, Basel, Switzerland.
Storn and Price [2] presented DE as a simple but effective EA for global optimization.
This article is an open access article The DE method has progressively gained popularity and has been utilized in a variety of
distributed under the terms and practical applications, owing to its shown strong convergence qualities and ease of
conditions of the Creative Commons understanding [3]. DE has been successfully applied in a variety of engineering
Attribution (CC BY) license disciplines [4–8]. The selected trial vector generation technique and related parameter
(https://creativecommons.org/license values have a significant impact on the performance of the traditional DE algorithm.
s/by/4.0/).

Appl. Sci. 2021, 11, 9772. https://doi.org/10.3390/app11209772 www.mdpi.com/journal/applsci

Appl. Sci. 2021, 11, 9772 2 of 26

Premature convergence can also be caused by poor methodology and parameter selection.
DE scholars have proposed a number of empirical guidelines and proposals for selecting
trial vector generation techniques and their associated control parameter settings during
the last decade [9–12].
Despite the fact that PSO has been successfully applied to a wide range of challenges,
including test and real‐world scenarios, it has faults that might cause the algorithm
performance to suffer. The major problem is a lack of diversity, which leads to a poor
solution or a slow convergence rate [13]. The hybridization of algorithms, in which two
algorithms are combined to generate a new algorithm, is one of the groups of modified
algorithms used to improve the performance. DE is applied to each particle for a certain
number of iterations in order to choose the best particle, which is then added to the
population [14,15]. The barebones DE [16] is a proposed hybrid version of PSO and DE.
The evolving candidate solution is created using DE or PSO and is based on a specified
probability distribution [17]. In a hybrid metaheuristic [13], the strengths of both
techniques are retained.
Cyclone separators are a low‐cost and low‐maintenance way of separating
particulates from air streams. A cyclone is made up of two parts: an upper cylindrical
element called the barrel and a lower conical part called the cone, as seen in Figure 1. The
air stream enters the barrel tangentially and goes downhill into the cone, generating an
outer vortex. The particles are separated from the air stream by a centrifugal force caused
by the increased air velocity in the outer vortex. When the air reaches the bottom of the
cone, an inner vortex forms, reversing the direction of the air and exiting out the top as
clean air, while particulates fall into the dust collection chamber attached to the bottom of
the cyclone.

Figure 1. Schematic flow diagram of a cyclone [18].

The simulation of an optimum gas cyclone with low cost is the primary focus of this
research with the use of PSO, DE, and hybrid DEPSO algorithms using an objective
function which is of a minimization type. The efficient global optimization is a major
advantage of the DE algorithm. Furthermore, the diversity of the entire population is
easily maintained throughout the process, preventing individuals from falling into a local
optimum. PSO, on the other hand, has the advantage of a quick convergence speed. The
best individual particle across the entire iteration is saved to obtain the lowest fitness
values. Combining the benefits of DE and PSO, the use of a hybrid DEPSO strategy is
proposed for this research with the goal of fast convergence and efficient global
Appl. Sci. 2021, 11, 9772 3 of 26

optimization. The proposed DEPSO method first reduces the search space using the DE
algorithm, and then the obtained populations are used as the initial population by the PSO
to achieve a fast convergence rate to a final global optimum. Moreover, DEPSO utilizes
the crossover operator feature of DE to improve the distribution of information between
candidates, and based on the fitness function, the hybrid algorithm can determine the
global minimum cost value much better than the use of PSO alone due to its drawbacks,
such as high computational complexity, slow convergence, sensitivity to parameters, and
so forth.
The overview of this paper is as follows:
‐ Sections 2 and 3 describe the PSO and DE algorithms, respectively.
‐ Section 4 provides insight about the DESPO algorithm.
‐ Section 5 gives a brief description about the gas cyclone device’s operation.
‐ Section 6 consists of the experimental results and performance evaluation of each
algorithm.
‐ Finally, Section 7 concludes the research work.

2. PSO
The theory behind particle swarm intelligence (PSO) is primarily nature‐inspired,
emulating the behaviour of animal societies by following the member of the group that is
closest to the food source, such as flocks of birds and schools of fish. For example, a flock
of birds during food search will follow a member of the flock that is closest to the position
of the food source (the potential solution). This is achieved because of the simultaneous
interactions made with each member of the flock in search of the best position. This
continues until the food source is discovered (the best solution).
The PSO algorithm follows this process in search of the best solution to a given
problem. The algorithm consists of a swarm of particles in which each particle is a
potential solution, which usually leads to the best solution. This best solution is known as
the global best and is called 𝑔 𝑡 or 𝑔 .
After obtaining the two best values, the particle updates its velocity and positions
with Equations (1) and (2):
vi(t 1)  wvi(t) r1c1(Pi(t)  xi(t))r2c2(g(t)  xi(t)) (1)

xi(t 1)= xi(t) vi(t 1) (2)

where:
w is inertia;
𝑣 𝑡 is the particle velocity;
𝑥 𝑡 is the current particle (solution);
𝑃 𝑡 and 𝑔 𝑡 are defined as personal best and global best, respectively;
𝑟 𝑟 are random numbers between (0,1);
𝑐 𝑐 are learning factors;
i is the i‐th particle.
The pseudo‐code (Algorithm 1) and the flowchart (see Figure 2) of the procedure are
shown below.

Algorithm 1 PSO
Input: Objective function: fi; lower bound: lb; upper bound: ub; population size: Np;
velocity: v; dimension size: D; termination criterion: T; inertia weight: w; learning rates: c1
and c2.
Output: Return gbest as the best estimation of the global optimum
Initialize the controlling parameters
Population = Initialize Population (Np,D,ub,lb,v)
Evaluate the objective function, fi
Assign Pbest as Population and fpbest as fi
Identify the solution with the best fitness and assign that solution as gbest and fitness as fgbest
Appl. Sci. 2021, 11, 9772 4 of 26

for t = 0 to T
for i = 0 to Np
Calculate the velocity, vi of i‐th particle
Calculate the new position, Xi of i‐th particle
Bound Xi
Evaluate objective function fi of i‐th particle
Update population using Xi and fi
if (fi < fpbest,i) then
Pbest,i = Xi
fbest,i = fi
end if
if (fPbest,i < fgbest) then
gbest = Pbest,i
fgbest = fpbest,i
end if
Update Inertia, w
end for
end for

Figure 2. Flowchart of PSO algorithm.

Appl. Sci. 2021, 11, 9772 5 of 26

3. DE
Similar to other evolutionary computation methods, DE starts with an initial
population that is initialized randomly. After determining the population, a new candidate
individual is generated by applying mutation and crossover operators. This candidate then
becomes the input of the selection operator, and through a hard selection mechanism
between the candidate and the current member of the population, if the candidate is better
than the current member, it will enter the next generation; otherwise, the current member
remains in the population. Algorithm 2 is the pseudo‐code of the standard DE. In the
algorithm, NP denotes the population size, D is the dimension size, F is the scaling factor, Cr
is the crossover rate, randint (1, D) is the random integer [1, D], and rand is the random real
number in [0, 1]. The main operator in DE is the mutation operator. These mutations take
advantage of diverse mechanisms and strategies for generating the donor vector. In order to
distinguish between DE variants, the notation DE/X/Y/Z is proposed by Storn and Price,
where DE denotes differential evolution, X denotes the base vector, Y denotes number of
difference vectors, and finally Z determines the type of crossover.
The pseudo‐code and flowchart (see Figure 3) of the procedure is as follows.

Algorithm 2 Pseudo‐code for classic differential evolution

Inputs: Fitness function: f; lb: lower bound; ub: upper bound: Np: population size; T: termination
criteria; F: scaling factor/mutation rate; Pc/Cr: crossover probability; D: dimension size.
Output: Best_Vector
Population = Initialize Population (Np,D,ub,lb)
while (T ≠ True) do
Best_Vector = Evaluate_Population (Population)
Vx = Select_Random_Vector (Population)
Index = Find_Index_of_Vector(Vx) //Specify row number of a vector
Select_Random_Vector (Population, v1,v2,v3)//where v1 ≠ v2 ≠ v3 ≠ vx
Vy = v1 + F (v2 – v3)// donor vector
for (i = 0;i + +;I < D – 1)//Loop for starting Crossover
operation
if (randj [0,1] < Cr) then
u[i] = vx[i]
else u[i] = vy[i]
end for //end crossover operation
if (f(u) ≤ f(vx)) then
Update Population (u, Index, Population)
end//While loop
return Best_Vector
Appl. Sci. 2021, 11, 9772 6 of 26

Figure 3. Flowchart of classic differential evolution.

4. Hybridised PSO
Hybridised PSO is a method that combines PSO with various classical and
evolutionary optimization algorithms to make use of the strengths of both approaches
while compensating for their flaws. DEPSO, a hybridised PSO which combines DE and
PSO, is the suggested hybridised PSO for this research.
DEPSO follows the same steps as the standard DE method until the trial vector is
created. If the trial vector meets the criteria, it is added to the population; otherwise, the
algorithm moves on to the PSO phase and creates a new candidate solution. Iteratively,
the technique is repeated until the optimal value is found. The incorporation of the PSO
phase generates a disturbance in the population, which aids in population diversification
and the output of an optimum solution. The following is an illustration of the algorithm’s
procedure [19]:
1. Population initialization: The individual 𝑥, with the population number NP, is
randomly generated to form an initial population in a 𝐷‐dimensional space. All the
individuals should be generated within the bounds of the solution space. The initial
individuals are generated randomly in the range of the search space. Additionally,
the associated velocities of all particles in the population are generated randomly in
the 𝐷‐dimension space. Therefore, the initial individuals and the initial velocity can
be expressed as follows [19]:
𝑋 0 𝑥 , 0 ,𝑥 , 0 ,…,𝑥 , 0
𝑥, 0 𝑖𝑖𝑥 𝑟𝑎𝑛𝑑 , 0,1 𝑥 𝑥 (3)
𝑉 0 𝑣 , 0 ,𝑣 , ,…,𝑣 , 0 ,
Appl. Sci. 2021, 11, 9772 7 of 26

where 1 ≤ i ≤ Np ≤ j ≤ D, [xmin, xmax] is the range of the search space, and rand(0,1) is a random
number chosen between 0 and 1.
2. Iteration Loop of DE: The individual mutation operation is denoted as time 𝑡. By
randomly choosing three individuals from the previous population, the mutant
individual (𝑡 + 1) can be generated as [19]:

𝑉𝑖 (𝑡 + 1) = 𝑋𝑟1(𝑡) + 𝐹(𝑋𝑟2 (𝑡)−𝑋𝑟3 (𝑡)) (4)

where 𝐹 is a differential weight between 0 and 1.
The crossover operation aims to construct a new population 𝑢𝑖,(𝑡 + 1), which is chosen
from the current individuals and mutant individuals in order to increase the diversity of
the generated individuals [19]:
𝑈𝑖 (𝑡 + 1) = {𝑢𝑖,1 (𝑡 + 1) , 𝑢𝑖,2 (𝑡 + 1) , . . . , 𝑢𝑖,𝐷 (𝑡 + 1)}
𝑣 𝑡 1 , 𝑖𝑖𝑖𝑓 𝑖𝑖𝑟𝑎𝑛𝑑 0,1 𝐶 𝑖𝑖𝑜𝑟 𝑖𝑖𝑗 𝑗 𝑖 (5)
𝑢, 𝑡 1
𝑥 𝑡 , 𝑖𝑖𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒, 𝑖𝑖
where rand(0, 1) is a random number chosen from 0 to 1, 𝑗rand is an integer chosen from 1
to 𝐷 randomly, and Cr is a crossover parameter that is randomly chosen from 0 to 1.
In the selection operation, the crossover vector 𝑈(𝑡+1) is compared to the target vector
𝑋𝑖(𝑡) by evaluating the fitness function value based on a greedy criterion, and the vector
with a smaller fitness value is selected as the next generation vector:
𝑈 𝑡 1 , 𝑖𝑖𝑖𝑓 𝑖𝑖𝑓 𝑋 𝑡 𝑓 𝑈 𝑡 1 𝑖𝑖
𝑋 𝑡 1 𝑖𝑖 (6)
𝑋 𝑡 , 𝑖𝑖𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒, 𝑖𝑖
The global best part is updated with the minimum fitness value (𝑔best) and the
personal‐best part (𝑝best).
3. Iteration Loop of PSO: The velocity and position equations of individuals remains
unchanged and follows the earlier stated Equations of (1) and (2) respectively in page
3 as shown below:

vi(t 1)  wvi(t) r1c1(Pi(t)  xi(t))r2c2(g(t)  xi(t))

xi(t 1)= xi(t) vi(t 1)
The pseudo code (Algorithm 3) and flowchart (see Figure 4) of DEPSO are shown
below.
Algorithm 3 DEPSO
Inputs: Fitness function: f; lb: lower bound; ub: upper bound; Np: population size; T:
termination criteria; F: scaling factor/mutation rate; Pc/Cr: crossover probability; D:
dimension size; w: inertia weight; learning rates: c1 and c2.
Output: BestVector
Population = Initialize Population (Np,D,ub,lb)
while (T ≠ True) do
Best_Vector = Evaluate_Population (Population)
Vx = Select_Random_Vector (Population)
Index = Find_Index_of_Vector(Vx) //Specify row number of a vector
Select_Random_Vector (Population, v1,v2,v3)//where v1≠ v2≠v3≠vx
Vy = v1 + F (v2 − v3)//donor vector
for (i = 0; i + +; I < D − 1)//Loop for starting Crossover
operation
if (randj[0,1] < Cr) then
u[i] = vx[i]
else u[i] = vy[i]
end for //end crossover operation
Appl. Sci. 2021, 11, 9772 8 of 26

if (f(u) ≤ f(vx)) then

Update Population (u, Index, Population)
else
for i = 0 to Np
Calculate the velocity, vi of i‐th particle
Calculate the new position, Xi of i‐th particle
Bound Xi
Evaluate objective function fi of i‐th particle
Update population using Xi and fi
if (fi < fpbest,i) then
Pbest,i = Xi
fbest,i = fi
end if
if (fPbest,i < fgbest) then
gbest = Pbest,i
fgbest = fpbest,i
Update Population (u, Index, Population)
end if
Update Inertia, w
end for
end//While loop
return BestVector
Appl. Sci. 2021, 11, 9772 9 of 26

Figure 4. Flowchart of DEPSO.

Appl. Sci. 2021, 11, 9772 10 of 26

5. Gas Cyclone Design

Cyclones are devices used for sizing, classification, and screening of particulate
materials in mixtures with fluid (gases or liquid). Cyclones come in many sizes and shapes
and have no moving parts. The mode of operation involves the process of subjecting the
flowing fluid to swirl around the cylindrical part of the device. They impact the cyclone
walls, fall down the cyclone wall (by gravity), and are collected in a hopper. The most
important parameter of a cyclone is its collection efficiency and the pressure drop across the
unit. Cyclone efficiency is increased through:
(a) Reduction in cyclone diameter, gas outlet diameter, and cone angle;
(b) Increasing the cyclone body length.
Capacity is, however, improved by increasing the cyclone diameter, inlet diameter,
and body length. Increasing the pressure drop give rise to:
 Increase in separation efficiency;
 Higher capacity;
 Decrease in the underflow to throughput ratio;
 Cleaner overflow.
The design parameters include:
‐ Feed rate Q m3/s;
‐ Cyclone diameter Dc, m;
‐ Set cut size dpc, μm;
‐ Cyclone efficiency;
‐ Cyclone pressure drop, DP;
‐ Cyclone geometry (D (cyclone diameter), De(vortex finder diameter), h(cylindrical
height), H(overall height), a(inlet height), b(inlet width), B(cone outlet diameter),
S(vortex finder height));
‐ Temperature, °C;
‐ Pressure, N/m2;
‐ Viscosity, Ns/m2;
‐ Density fluid, kg/m3;
‐ Mass median diameter MMD, (μm);
‐ Geometric standard deviation GSD, μm;
‐ Number of cyclones;
‐ Fluid density, Kg/m3;
‐ Particle density, Kg/m3.
Given the cyclone geometry (as shown in Figure 5) and the operating conditions,
there are five design parameters that can be specified for a design.
These parameters are:
1. Feed rate, Q;
2. Cyclone diameter, D;
3. Set cut size, dpc;
4. Cyclone efficiency;
5. Cyclone pressure drop, DP.
Appl. Sci. 2021, 11, 9772 11 of 26

De

S
b

a h

D H

Figure 5. The gas cyclone geometry.

Cyclone geometry:
‐ a is the inlet height
‐ b is the inlet width
‐ D is the cyclone diameter
‐ De is the vortex finder diameter
‐ h is the cylindrical height
‐ B is the cone diameter
‐ H is the overall height
‐ S is the vortex finder height.

5.1. Particle Cut‐Off Size DPC

According to Wang et al. [20–22], cyclone performance depends on the geometry and
operating parameters of the cyclone, as well as the particle size distribution of the
entrained particulate matter. Several mathematical models have been developed to
predict cyclone performance. Lapple [23] developed a semi‐empirical relationship to
predict the cut point of cyclones designed according to the classical cyclone design
method, where cyclone cut point is defined as the particle diameter corresponding to a
50% collection efficiency. Wang et al. showed that Lapple’s approach did not discuss the
effects of particle size distribution on cyclone performance. The Lapple model was based
on the terminal velocity of particles in a cyclone [23]. From the theoretical analysis,
Equation (7) was derived to determine the smallest particle that will be collected by a
cyclone if it enters at the inside edge of the inlet duct:
µ
dpc = (7)
ᴨ

where:
Appl. Sci. 2021, 11, 9772 12 of 26

dpc (cut‐point) = diameter of the smallest particle that will be collected by the cyclone
if it enters on the inside edge of the inlet duct (μm);
μ = gas viscosity (kg/m/s);
b = width of inlet duct (m);
Ne = number of turns of the air stream in the cyclone;
Vi = gas inlet velocity (m/s);
𝜌𝑝 = particle density (kg/m3);
ρg = gas density (kg/m3).

5.2. Fractional Efficiency Calculation

The most important parameters in cyclone operation are pressure drop and collection
efficiency. The pressure drop is given by the difference between the static pressure at the
cyclone entry and the exit tube. The fractional efficiency for the j‐th particle size, according
to Lapple, is given as [23]:

ɳj =
(8)

where:
dpc = diameter of the smallest particle that will be collected by the cyclone with 50%
efficiency,
dpj = diameter of the j‐th particle.
The overall collection efficiency of the cyclone is a weighted average of the collection
efficiencies for the various size ranges, namely:
ɳ
ɳ= ∑ (9)

where:
ɳ = overall collection efficiency;
ɳj = fractional efficiency for j‐th particle size;
m = total mass of particle;
mj = mass of particle in the j‐th particle size range.

5.3. Pressure Drop Calculation

The energy consumed in a cyclone is most frequently expressed as the pressure drop
across the cyclone. This pressure drop is the difference between the gas static pressure
measured at the inlet and outlet of the cyclone. Many models have been developed to
determine this pressure drop. Some of the commonly used equations to calculate the
pressure drop are:
(a) The Koch and Licht Pressure Drop Equation.
Koch and Licht (1977) expressed the cyclone pressure drop as
∆P = 0.003𝜌gVi2 NII (10)
where:
𝜌g = gas density (lbm/ft3);
Vi = inlet velocity (ft/s);
NII = number of velocity heads (inches of water) and is expressed as
.
NII = K ( ) (11)

K = 16 for no inlet vane 7.5 with neutral inlet vane;

a, b = inlet height and width, respectively.
(b) The Ogawa Equation.
Another pressure drop equation due to Ogawa (1984) takes the form of
Appl. Sci. 2021, 11, 9772 13 of 26

1
∆P ρ V Ν∏ (12)
2
The prediction of the performance of the cyclone separators is a challenging problem
for the designers owing to the complexity of the internal aerodynamic process and dust
particles. Hence, modern numerical simulations are needed to solve this problem. Fluid
flows have long been mathematically described by a set of nonlinear, partial differential
equations, namely the Navier–Stokes equations.
A cyclone is required to remove carbon dust particles from affluent air coming from
a thermal power station at a rate of 5.0 m3/s at 65 °C. The dust particles are assumed to
have a normal size distribution with an MMD of 20μm and a GSD of 1.5μm. For energy
cost consideration, the pressure drop in the cyclone is required to be not more than 1000
N/m2. The density of the dust particle is 2250 kg/m3, and the cyclone is operating at
atmospheric pressure. To attain overall efficiency of 70% and above, the recommended
required geometry/size and cut size of the cyclone are: Q = 5 m3/s, temp = 65 °C, MMD =
20μm, GSD = 15μm, pressure 1000 N/m2, and ρ dust particle = 2250 kg/m2.
This study develops PSO, DE, and DEPSO models for the design optimization of a
5.0 m3/s gas cyclone instrument.

6. Results and Discussion

The cost of the operation of a gas cyclone is the primary problem set for this research,
instead of using two conflicting objective functions, namely efficiency and pressure drop.
In general, the total cost per unit will be a function of a fixed cost cyclone and the energy
cost of operating the cyclone:
Ct = Fixed cost + energy cost.
Assuming the cost of the cyclone depends on its diameter, the fixed cost can be
expressed as
Cfixed = fNeDc2/YH (13)
where:
f is an investment factor to allow for installation;
Dc is the cyclone diameter;
Ne is number of turns of the air stream in the cyclone;
H is the time worked per year;
Y is the number of years.
The energy cost is given by
Cenergy = Q∆PCe (14)
where:
Q is the feed rate (m3/s);
∆P is the pressure drop (Pa);
Ce is cost per unit energy.
Hence,
Ct = fNeβ1/YH + ρf εQ3 β2/2ao2 bo2N2 (15)
where:
β1 = [dpc2(ρs − ρf)ᴨNtNeQ/9aobo2 μN]2/3;
β2 = [dpc2(ρs − ρf)ᴨNtQ/9aobo2 μN]−4/3.
The objective function is of a minimization type, and the EC methods search for an
optimized cyclone geometry with low cost per unit. Using the table of parameters as
shown in Tables 1–9, the cost performance and overall efficiency of each of the algorithms
are shown in the figures below.
The following can be observed from the plots and tables below:
Appl. Sci. 2021, 11, 9772 14 of 26

1. It can be seen from Table 10 that DEPSO at 1000 iterations with a particle size of 40
has a cost‐effective value of 230 naira/seconds, with an overall efficiency of 0.70.
2. It can also be seen from Table 10 and Figures 6–19 that the PSO algorithm fails to
provide any reasonable results when applied to complex optimization problems, due
to its premature convergence, high computational complexity, slow convergence,
sensitivity to parameters, and so forth. DEPSO solves this issue by utilizing the
crossover operator of DE to improve the distribution of information between
candidate solutions.
3. In terms of the best efficiency at low cost for DE algorithm, at a particle size of 60 at
400 iterations, its cost and overall efficiency are 229.365 naira/seconds and 0.749,
respectively, as shown in Figures 10 and 11.
4. In terms of the best efficiency at a low cost for the DEPSO algorithm, at a particle size
of 70 at 1000 iterations, its cost and overall efficiency are 249.192 naira/seconds and
0.786, respectively, as shown in Figures 12 and 13.
5. As seen from Table 10, the DE algorithm’s lowest cost is 218.938 naira/seconds with
an overall efficiency of 0.716.

Table 1. Cyclone input data.

Parameters Value
Feed rate, Q 5.0
Particle size 1 0.00002
Particle size 2 0.0000015
Particle Density 2250
Viscosity 0.00002039
Delta pressure 1000
Particle mass 1 0.000056
Particle mass 2 0.000042
Number of particle range 1
Assumed natural length 2.078

Table 2. Investment/cost data.

Parameters Value
Investment factor 3.9
Number of years 10
Period of operation/year, (seconds) 0.0000015
Cost/square m of cyclone 20.5
Number of cyclone 1

Table 3. Parameters list with 40‐particle size.

Parameters Value
Particle size 40
Dimensions 7
Probability of crossover range, Cr 0.5
Scaling factor 1.0
w 0.729
c1 2
c2 2
Appl. Sci. 2021, 11, 9772 15 of 26

Figure 6. Cost performance of DE, DEPSO, and PSO algorithms at 40 particles.

Figure 7. Overall efficiency of DE, DEPSO, and PSO algorithms at 40 particles.

Table 4. Parameters list with 50‐particle size.

Parameters Value
Particle size 50
Dimensions 7
Probability of crossover range, Cr 0.5
Scaling factor 1.0
w 0.729
c1 2
c2 2
Appl. Sci. 2021, 11, 9772 16 of 26

Figure 8. Cost performance of DE, DEPSO, and PSO algorithms at 50 particles.

Figure 9. Overall efficiency of DE, DEPSO, and PSO algorithms at 50 particles.

Table 5. Parameters list with 60‐particle size.

Parameters Value
Particle size 60
Dimensions 7
Probability of crossover range, Cr 0.5
Scaling factor 1.0
W 0.729
c1 2
c2 2
Appl. Sci. 2021, 11, 9772 17 of 26

Figure 10. Cost performance of DE, DEPSO, and PSO algorithms at 60 particles.

Figure 11. Overall efficiency of DE, DEPSO, and PSO algorithms at 60 particles.

Table 6. Parameters list with 70‐particle size.

Parameters Value
Particle size 70
Dimensions 7
Probability of crossover range, Cr 0.5
Scaling factor 1.0
w 0.729
Appl. Sci. 2021, 11, 9772 18 of 26

c1 2
c2 2

Figure 12. Cost performance of DE, DEPSO, and PSO algorithms at 70 particles.

Figure 13. Overall efficiency of DE, DEPSO, and PSO algorithms at 70 particles.

Table 7. Parameters list with 80‐particle size.

Parameters Value
Particle size 80
Dimensions 7
Probability of crossover range, Cr 0.5
Scaling factor 1.0
w 0.729
Appl. Sci. 2021, 11, 9772 19 of 26

c1 2
c2 2

Figure 14. Cost performance of DE, DEPSO, and PSO algorithms at 80 particles.

Figure 15. Overall efficiency of DE, DEPSO, and PSO algorithms at 80 particles.
Appl. Sci. 2021, 11, 9772 20 of 26

Table 8. Parameters list with 90‐particle size.

Parameters Value
Particle size 90
Dimensions 7
Probability of crossover range, Cr 0.5
Scaling factor 1.0
w 0.729
c1 2
c2 2

Figure 16. Cost performance of DE, DEPSO, and PSO algorithms at 90 particles.
Appl. Sci. 2021, 11, 9772 21 of 26

Figure 17. Overall efficiency of DE, DEPSO, and PSO algorithms at 90 particles.

Table 9. Parameters list with 100‐particle size.

Parameters Value
Particle size 100
Dimensions 7
Probability of crossover range, Cr 0.5
Scaling factor 1.0
w 0.729
c1 2
c2 2
Appl. Sci. 2021, 11, 9772 22 of 26

Figure 18. Cost performance of DE, DEPSO, and PSO algorithms at 100 particles.

Figure 19. Overall efficiency of DE, DEPSO, and PSO algorithms at 100 particles.
Appl. Sci. 2021, 11, 9772 23 of 26

Table 10. Lowest cost value.

PSO DE DEPSO
Overall Overall Overall
Cost Cost Cost
Efficiency Efficiency Efficiency
297.715 0.793 218.938 0.716 230 0.70
Iteration
100–1000 1000 1000
Number
Particle Size 50–100 70 40

The optimized cyclone geometry, using PSO, DE, and hybrid DEPSO algorithms, is
shown in Tables 11–14 below.

Table 11. PSO parameters used for design parameter optimization.

Parameters Value
Maximum iteration 1000
Particle size 50
Dimensions 7
Probability of crossover range, Cr 0.5
Scaling factor 1.0
w 0.729
c1 2
c2 2

Table 12. DE parameters used for design parameter optimization.

Parameters Value
Maximum iteration 1000
Particle size 70
Dimensions 7
Probability of crossover range, Cr 0.5
Scaling factor 1.0
w 0.729
c1 2
c2 2

Table 13. DEPSO Parameters used for design parameter optimization.

Parameters Value
Maximum iteration 1000
Particle size 40
Dimensions 7
Probability of crossover range, Cr 0.5
Scaling factor 1.0
w 0.729
c1 2
c2 2
Appl. Sci. 2021, 11, 9772 24 of 26

Table 14. Design parameters.

Design Parameters are in Analytical Method

PSO DE Hybrid DEPSO
the Ratio of Dc (Stairmand Model)
Overall cyclone height 4.0 3.925 3.371 3.884
Inlet height, ao 0.5 0.491 0.494 0.489
Gas outlet diameter, Deo 0.5 0.168 0.411 0.309
Cylindrical height of
1.5 0.977 0.817 1.006
cyclone, ho
Inlet width, bo 0.2 0.213 0.242 0.209
Gas outlet length, So 0.5 0.674 1.177 1.171
Dust outlet diameter, Bo 0.375 0.485 0.447 0.179
Pressure drop (Pa) 1000 469.034 364.668 493.24
Inlet velocity m/s 19.86 11.066 9.706 11.36
Cut size dpc (μm) 8.67855 × 10 −6 1.021 × 10 −5 1.26 × 10 −5 9.96 × 10−6
Cost/second (Naira/s) ‐ 297.715 218.938 230
Overall efficiency 0.890 0.793 0.716 0.70
Particle size ‐ 50 70 40

The following can be observed in Table 14 above:

1. Due to the problem of being a minimization type, each algorithm’s efficiency is lower
compared to the Stairmand’s model. The pressure drop of DE is the lowest, followed
by PSO and DEPSO. Moreover, each algorithm’s pressure drop is considerably lower
than the Stairmand’s model. An increase in pressure drop will lead to an increase in
separation efficiency, higher capacity, and cleaner overflow.
2. Another factor that leads to lower efficiency is the inlet velocity of the cyclone; as
seen in Table 14 above, all algorithms have a lower inlet velocity when compared to
the Stairmand’s model. An increase in inlet velocity leads to an increase in separation
efficiency, thereby leading to an increase in the overall efficiency, because of higher
resultant centrifugal force.

7. Conclusions
This research utilized a hybrid particle swarm optimization approach to design a gas
cyclone with design parameters at a low cost. When compared to PSO individually, as
shown in Table 14, the adoption of a hybridised particle swarm optimization technique
produced a lower cost of 230 naira/second, while the use of DE had a cost value of 218.938
naira/second.
In this research, it can be concluded that:
 The PSO algorithm, when applied to complex optimization problems, can be
improved via the combination with other algorithms, such as the DE algorithm,
thereby improving the distribution of information between candidates, due to the
crossover feature present within it.
 When both the cognitive component c1 and the social component c2 have the same
value, both exploration and exploitation are evenly balanced. Furthermore, the
inertia weight, w, when less than one, i.e., w < 1, has an effect on the distribution of
information across the particles. In addition, when inertia weight is less than one, i.e.,
w < 1, the velocity of the particle decreases until it reaches zero, thereby causing
exploitation to gradually increase and exploration to gradually decrease. In most
cases, changing the value of inertia weight, w, during optimization will balance out
exploration and exploitation. Some studies, such as [24–26], have developed an
inertia weight application.
 When compared to PSO, DEPSO had a better cost value but lower efficiency.
Appl. Sci. 2021, 11, 9772 25 of 26

 As shown in Table 14, DE and DEPSO algorithms produced suitable results for
design purposes with a low cost, while PSO showed the highest cost value.
 Due to the problem set being of a minimization type, the efficiency suffers
considerably, but if the problem was of a maximization type, of which maximum
overall efficiency is required, the cost and overall efficiency would be considerably
high, provided that the cost of operation is not the primary focus.
 Having made use of the DE/rand/1 strategy with PSO to optimize the cost of
operation of the gas cyclone, various other strategies of DE, such as DE/best/1,
DE/rand/2, DE/best/2, and DE/target‐to‐best/1, can also be used.
 PSO is a concept which can be used to solve numerous optimization problems and
can be improved through combination with other algorithms. Algorithms such as
GA, ACO, TLBO, and so forth can, in fact, be explored to form a hybridised PSO
algorithm. Some research on such combinations includes [27–33].

Author Contributions: Conceptualization, X.S. and D.C.I.; methodology, D.C.I.; writing—original

draft preparation, D.C.I.; writing‐review and editing, X.S. and D.C.I.; supervision, X.S. All authors
have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Conflicts of Interest: The author declares no conflict of interest.

Abbreviations
ACO Ant Colony Optimization
DE Differential Evolution
DEPSO Differential Evolution PSO
EC Evolutionary Computation
EA Evolutionary Algorithm
GA Genetic Algorithm
PSO Particle Swarm Optimization
TLBO Teaching‐Learning Based Optimization

References
1. Kennedy, J.; Eberhart, R. Particle swarm Optimization. In Proceedings of the IEEE international Conference on Neural
Networks, Perth, WA, Australia, 27 November–1 December 1995; Volume 4, pp. 1942–1948.
2. Storn, R.; Price, K. Differential Evolution&ndash; A Simple and Efficient Heuristic for global Optimization over Continuous
Spaces. J. Glob. Optim. 1997, 11, 341–359.
3. Liu, J.; Lampinen, J. On setting the control parameter of the differential evolution method. In Proceedings of the 8th International
Conference on Soft Computing (MENDEL’02), Brno, Czech Republic, 8–10 June 2002; pp. 11–18.
4. Tang, Y.; Gao, H.; Zou, W.; Kurths, J. Identifying controlling nodes in neuronal networks in different scales. PLoS ONE 2012, 7,
e41375.
5. Mallipeddi, R.; Suganthan, P.; Pan, Q.‐K.; Tasgetiren, M.F. Differential evolution algorithm with ensemble of parameters and
mutation strategies. Appl. Soft Comput. 2011, 11, 1679–1696.
6. Das, S.; Konar, A. Automatic image pixel clustering with an improved differential evolution. Appl. Soft Comput. 2009, 9, 226–
236.
7. Storn, R. Differential evolution design of an IIR‐filter. In Proceedings of the 1996 IEEE International Conference on Evolutionary
Computation (ICEC’96), Nagoya, Japan, 20–22 May 1996; pp. 268–273.
8. Varadarajan, M.; Swarup, K. Differential evolution approach for optimal reactive power dispatch. Appl. Soft Comput. 2008, 8,
1549–1561.
9. Mallipeddi, R.; Suganthan, P.N. Differential evolution algorithm with ensemble of parameters and mutation and crossover
strategies. In Proceedings of the Swarm Evolutionary and Memetic Computing Conference, Chennai, India, 16–18 December
2010; pp. 71–78.
10. Brest, J.; Greiner, S.; Bôskovi’c, B.; Mernik, M.; Zumer, V. Self‐adapting control parameters in differential evolution: A
comparative study on numerical benchmark problems. IEEE Trans. Evol. Comput. 2006, 10, 646–657.
11. Yang, Z.; Tang, K.; Yao, X. Self‐adaptive differential evolution with neighborhood search. In Proceedings of the IEEE Congress
on Evolutionary Computation (CEC’08), Hong Kong, China, 1–6 June 2008; pp. 1110–1116.
Appl. Sci. 2021, 11, 9772 26 of 26

12. Zhang, J.; Sanderson, A.C. JADE: Adaptive Differential Evolution with Optional External Archive. In IEEE Transactions on
Evolutionary Computation; IEEE: Manhattan, NY, USA, 2009; Volume 13, pp. 945–958.
13. Pant, M.; Thangaraj, R.; Abraham, A. DE‐PSO: A new hybrid meta‐heuristic for solving global optimization problems. New
Math. Nat. Comput. 2011, 7, 363–381.
14. Kannan, S.; Slochanal, S.R.; Subbaraj, P.; Padhy, N.P. Application of particle swarm optimization technique and its variants to
generation expansion planning problem. Electr. Power Syst. Res. 2004, 70, 203–210.
15. Talbi, H.; Batouche, M. Hybrid particle swarm with differential evolution for multimodal image registration. In Proceedings of
the IEEE International Conference on Industrial Technology (ICIT ‘04), Hammamet, Tunisia, 1 December 2004; pp. 1567–1572.
16. Omran, M.G.H.; Engelbrecht, A.P.; Salman, A. Differential evolution based particle swarm optimization. In Proceedings of the
IEEE Swarm Intelligence Symposium (SIS ‘07), Honolulu, HI, USA, 1–5 April 2007; pp. 112–119.
17. Hao, Z.F.; Guo, G.H.; Huang, H. A particle swarm optimization algorithm with differential evolution. In Proceedings of the 6th
International Conference on Machine Learning and Cybernetics (ICMLC ‘07), Hong Kong, China, 19–22 August 2007; pp. 1031–
1035.
18. Wang, L. Theoretical Study of Cyclone Design. Ph.D. Thesis, Texas A&M University, College Station, TX, USA, 2004. Available
online: https://core.ac.uk/download/pdf/147123938.pdf (accessed on 2 October 2021).
19. Mao, B.; Xie, Z.; Wang, Y.; Handroos, H.; Wu, H. A Hybrid Strategy of Differential Evolution and Modified Particle Swarm
Optimization for Numerical Solution of a Parallel Manipulator. Math. Probl. Eng. 2018, 2018, 1–9.
20. Wang, L.; Parnell, C.B.; Shaw, B.W. 1D2D, 1D3D, 2D2D cyclone fractional efficiency curves for fine dust. In Proceedings of the
2000 Beltwide Cotton Conferences, San Antonio, TX, USA, 4−8 January 2000.
21. Wang, L.; Parnell, C.B.; Shaw, B.W. A New Theoretical Approach for Predicting Number of Turns and Cyclone Pressure Drop; St.
Joseph, Mich.: ASAE: Washington, DC, USA, 2001.
22. Wang, L.; Parnell, C.B.; Oemler, J.A.; Shaw, B.W.; Lacey, R.E. Analysis of Cyclone Collection Efficiency; St. Joseph, Mich.: ASAE:
Washington, DC, USA, 2003.
23. Lapple, C.E. Processes use many collector types. Chem. Eng. 1951, 58, 144‐151.
24. Müller, P.K.S.; Airaghi, S.; Marchetto, J. Optimization algorithms based on a model of bacterial chemotaxis. In From Animals to
Animats 6: Proceedings of the Sixth International Conference on Simulation of Adaptive Behavior; 1st Ed;A Bradford Book: London,
UK, 2000; pp. 375–384. Available online: https://direct.mit.edu/books/book/4493/From‐Animals‐to‐Animats‐6Proceedings‐of‐
the‐Sixth (accessed on 2 October 2021).
25. Passino, K.M. Biomimicry of bacterial foraging for distributed optimization and control. IEEE Control. Syst. Mag. 2002, 22, 52–
67.
26. Karaboga, D.; Basturk, B. Artificial Bee Colony (ABC) optimization algorithm for solving constrained optimization problems.
In Foundations of Fuzzy Logic and Soft Computing; Melin, P., Castillo, O., Aguilar, L.T., Pedrycz, W., Eds.; Springer: New York,
NY, USA, 2007; pp. 789–798.
27. Mesloub, S.; Mansour, A. Hybrid PSO and GA for global maximization. Int. J. Open Probl. Comput. Sci. Math. 2009, 2, 597–608.
28. Premalatha, K.; Natarajan, A.M. Hybrid PSO and GA models for document clustering. Int. J. Adv. Soft Comput. Its Appl. 2010, 2,
302–320.
29. Wang, H.; Li, Y. Hybrid teaching‐learning‐based PSO for trajectory optimisation. Electron. Lett. 2017, 53, 777–779.
30. Kao, Y.; Chen, M.‐H.; Huang, Y.‐T. A Hybrid Algorithm Based on ACO and PSO for Capacitated Vehicle Routing Problems.
Math. Probl. Eng. 2012, 2012, 1–17.
31. Dai, Q. A hybrid PSO/ACO algorithm for land cover classification. In Proceedings of the 2nd International Conference on
Information Science and Engineering, Hangzhou, China, 3–5 December 2010; pp. 3428–3430.
32. Menghour, K.; Souici‐Meslati, L. Hybrid ACO‐PSO Based Approaches for Feature Selection. Int. J. Intell. Eng. Syst. 2016, 9, 65–
79. https://doi.org/10.22266/ijies2016.0930.07.
33. Azad‐Farsani, E.; Zare, M.; Azizipanah‐Abarghooee, R.; Askarian‐Abyaneh, H. A new hybrid CPSO‐TLBO optimization
algorithm for distribution network reconfiguration. J. Intell. Fuzzy Syst. 2014, 26, 2175–2184.