Vous êtes sur la page 1sur 8

Applied Soft Computing 12 (2012) 3648–3655

Contents lists available at SciVerse ScienceDirect

Applied Soft Computing


journal homepage: www.elsevier.com/locate/asoc

Soft computing based multi-objective optimization of steam cycle power plant


using NSGA-II and ANN
Farzaneh Hajabdollahi, Zahra Hajabdollahi, Hassan Hajabdollahi ∗
Department of Mechanical Engineering, Behbahan Branch, Islamic Azad University, Behbahan, Iran

a r t i c l e i n f o a b s t r a c t

Article history: In this paper a steam turbine power plant is thermo-economically modeled and optimized. For this pur-
Received 30 May 2011 pose, the data for actual running power plant are used for modeling, verifying the results and optimization.
Received in revised form 18 May 2012 Turbine inlet temperature, boiler pressure, turbines extraction pressures, turbines and pumps isentropic
Accepted 5 June 2012
efficiency, reheat pressure as well as condenser pressure are selected as fifteen design variables. Then,
Available online 3 July 2012
the fast and elitist Non-dominated Sorting Genetic Algorithm (NSGA-II) is applied to maximize the ther-
mal efficiency and minimize the total cost rate (sum of investment cost, fuel cost, and maintenance cost)
Keywords:
simultaneously. The results of the optimal design are a set of multiple optimum solutions, called ‘Pareto
Steam turbine cycle
Thermal efficiency
optimal solutions’. The optimization results in some points show 3.76% increase in efficiency and 3.84%
Total cost rate decrease in total cost rate simultaneously, when it compared with the actual data of the running power
NSGA-II plant. Finally as a short cut to choose the system optimal design parameters a correlation between two
Artificial Neural Network objectives and fifteen decision variables with acceptable precision are presented using Artificial Neural
Network (ANN).
© 2012 Elsevier B.V. All rights reserved.

1. Introduction rate of exergy destruction as two objective functions. Sahin and


Ali [16] performed an optimal performance analysis of a combined
The optimization of power generation systems is one of the most Carnot cycle in a cascade form, including internal irreversibility for
important subjects in the energy engineering field. Owing to the steady-state operation. Koch et al. used the Genetic Algorithm to
high prices of energy and the decreasing fossil fuel resources, makes find the optimum design configuration and process variables in
the optimum design of energy consumption management methods a combined cycle power plant [17]. Bracco and Siri optimized a
important. Recently, thermo-economic and multi-objective analy- combined cycle power plant with a single pressure HRSG in exergy
sis and optimization of thermal systems have become a key solution point of view [18]. Valdes et al. applied Genetic Algorithm to find
in providing a better system in optimal energy consumption and the optimum thermodynamic parameters of a heat recovery steam
optimal system configuration [1–5]. In addition, in the last decade generator (HRSG) in a combined cycle power plant [19]. Aljundi
the application research of soft computing such as fuzzy set, Genetic performed the energy and exergy analysis of steam power plant to
Algorithm and Artificial Neural Network has become one of the find the points with energy and exergy losses [20]. Ogaji and Singh
most important topics in industrial applications and decision mak- performed the system identification for a two-shaft aeroderivative
ing [6–10]. In particular, in the field of industrial power plant gas turbine to specify the fault by means of Artificial Neural Net-
application for energy production including the gas turbine and work [6]. Bertini et al. introduced a novel fitness approximation
combined cycle power plant, it has been mainly applied for system technique coupled with Fuzzy Logic and Genetic Algorithm to opti-
identification [6,7]. Rosen and Dincer [11] used thermo-economic mize the conflicting objectives in a combined cycle power plant
analysis of power plants and applied it to a coal fired electricity [7].
generating power plant. Ameri et al. applied exergy and Genetic In this paper after thermo-economic modeling of steam cycle
Algorithm tools for thermal modeling and optimal design of steam power plant (SCPP), this equipment is optimized by maximiz-
cycle power plant [12]. Ahmadi and Dincer applied multi objective ing the thermal efficiency as well as minimizing the total cost
optimization for a gas turbine and combined cycle power plant rate, simultaneously. Turbine inlet temperature, boiler pressure,
[13–15]. They considered the total cost and exergy efficiency or turbines extraction pressures, turbines and pumps isentropic
efficiency, reheat pressure as well as condenser pressure are taken
as fifteen design parameters and fast and elitist Non-dominated
Sorting Genetic Algorithm (NSGA-II) is applied to provide a set of
∗ Corresponding author. Tel.: +98 913 2924318; fax: +98 913 2924318.
Pareto multiple optimum solutions. After the thermal modeling
E-mail address: Hassan.Hajabdollahi@gmail.com (H. Hajabdollahi).

1568-4946/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.asoc.2012.06.006
F. Hajabdollahi et al. / Applied Soft Computing 12 (2012) 3648–3655 3649

Turbine:
Nomenclature  
ẆT,a = ṁi hi − ṁe he (5)
a annualized factor (–)
Cin investment cost ($)
Ctotal total cost rate ($/s) ẆT,a hi − he,a
T = = (6)
Cfuel fuel cost ($/year) ẆT,s hi − he,s
h enthalpy (kJ/kg K)
cfuel price of fuel ($/kg) Condenser:
ṁ mass flow rate (kg/s)
 
Q̇cond = ṁi hi − ṁe he (7)
Q̇ rate of heat transfer (kW)
Ẇ power (kW)
Pump:
LHV fuel lower heating value (kJ/kg)
P pressure (kPa) Ẇp,s i (Pe − Pi )
p = = (8)
TTD terminal temperature difference (◦ C) Ẇp,a he − hi
y depreciation time (year)
r interest rate (–) Feed water heater:
 
Greek abbreviation ṁi hi = ṁe he (9)
ε total cycle thermal efficiency (–)
 hours of operation per year Terminal temperature difference or TTD is used for the closed
 efficiency (–) feed water heaters and defined as the temperature difference
ϕ maintenance factor (–) between saturated steam from extraction and feed water leaving
the heater [21].
Subscripts In addition the total thermal efficiency of the cycle is computed
i inlet from:
o outlet Ẇnet
b boiler ε= (10)
ṁfuel .LHV
s isentropic
a actual where LHV and ṁfuel are fuel lower heating value and fuel mass flow
cond condenser rate respectively and Ẇnet is output net power estimated as follow:
p pump  
T turbine Ẇnet = ẆT,a − Ẇp,a (11)

3. Genetic Algorithm for multi-objective optimization


and optimization of the SCPP, the Artificial Neural Network is used
to find a closed form equation for the optimum design parameters 3.1. Definition of multi-objective optimization
(decision variables) versus the total cost rate and efficiency. This
closed form equation enables designers to find the optimum design A multi-objective optimization problem requires the simulta-
parameters for the optimum total cost rate and efficiency. neous satisfaction of a number of different and often conflicting
objectives. It is required to mention that no combination of
decision variables can optimize all objectives, simultaneously.
2. Thermal modeling
Multi-objective optimization problems generally show a possibly
uncountable set of solutions, whose evaluated vectors represent
In order to do the thermal modeling, mass and energy balances
the best possible trade-offs in the objective function space. Pareto
on the system are required to determine the flow rates and energy
optimality is the key concept to establish a hierarchy among the
transfer rates at the control surface. Appling the first law of ther-
solutions of a multi-objective optimization problem, in order to
modynamic in the steady state, one can find the formula for mass
determine whether a solution is really one of the best possible
and energy balance as follow [21]:
trades-off [22]. A multi-objective problem consists of optimizing
Mass balance equation:
(i.e., minimizing or maximizing) several objectives simultaneously,
  with a number of inequality or equality constraints. The problem
ṁi = ṁe (1)
can be formally written as follows:
Find x = (xi ) ∀i = 1, 2, . . . , Nparam such as
Energy balance equation:
fi (x) is a minimum (respectively maximum) ∀i = 1, 2, . . . , Nobj
 
Q̇ − Ẇ = ṁe he − ṁi hi (2) Subject to:

gj (x) = 0 ∀j = 1, 2, . . . , M, (12)
where subscripts i and e refer to streams entering and leaving the
control volume, respectively. hk (x) ≤ 0 ∀k = 1, 2, . . . , K, (13)
The energy balance equations for the various parts of the steam
where x is a vector containing the Nparam design parameters,
turbine cycle as shown in Fig. 1 are as follow:
(fi )i=1, ..., Nobj the objective functions and Nobj the number of
Boiler:
objectives. The objective function (fi )i=1, ..., Nobj returns a vector
Q̇b containing the set of Nobj values associated with the elementary
b = (3)
ṁfuel .LHV objectives to be optimized simultaneously. The first multi-objective
GA, called vector evaluated GA (or VEGA), was proposed by Schaffer
Q̇b = ṁ27 (h32 − h27 ) + ṁ30 (h31 − h30 ) (4) [23]. An algorithm based on non-dominated sorting was proposed
3650 F. Hajabdollahi et al. / Applied Soft Computing 12 (2012) 3648–3655

Fig. 1. Schematic diagram of steam turbine cycle studied in this paper.

by Srinivas and Deb [24] and called non-dominated sorting genetic- 3.4. Controlled elitism sorting
algorithm (NSGA). This was later modified by Deb et al. [25] which
eliminated higher computational complexity, lack of elitism and To preserve diversity, the influence of elitism is controlled
the need for specifying the sharing parameter. This algorithm is by choosing the number of individuals from each subpopulation,
called NSGA-II which is coupled with the objective functions devel- according to the geometric distribution [27]:
oped in this study for optimization.
1 − c q−1
Sq = S c , (15)
1 − cw
3.2. Non-dominated sorting and Pareto front To form a parent search population, Pt+1 (t denote the genera-
tion), of size S, where 0 < c < 1 and w is the total number of ranked
As defined by Deb and Goel [26], an individual X(a) is said to non-dominated.
constrain-dominate an individual X(b) , if any of the following con-
ditions are true: 3.5. Crowding distance

The crowding distance metric proposed by Deb and Goel [26]


(1) X(a) and X(b) are feasible, with
is utilized, where the crowding distance of an individual is the
(a) X(a) is no worse than X(b) in all objective, and
perimeter of the rectangle with its nearest neighbors at diagonally
(b) X(a) is strictly better than X(b) in at least one objective.
opposite corners. So, if individual X(a) and individual X(b) have same
(2) X(a) is feasible while individual X(b) is not.
rank, each one has a larger crowding distance is better.
(3) X(a) and X(b) are both infeasible, but X(a) has a smaller constraint
violation.
3.6. Crossover and mutation

Here, the constraint violation (X) of an individual X is defined Uniform crossover and random uniform mutation are employed
to be equal to the sum of the violated constraint function values to obtain the offspring population, Qt+1 .The integer-based uniform
[27] crossover operator takes two distinct parent individuals and inter-
changes each corresponding binary bits with a probability, 0 < pc ≤

B 1. Following crossover, the mutation operator changes each of the
(X) = (gj (X))gj (X), (14) binary bits with a mutation probability, 0 < pm < 0.5.
j=1
3.7. Historical archive
where  is the Heaviside step function. A set of non-dominated
individuals is used to form a Pareto-optimal fronts. The NSGA-II algorithm has been modified to include an archive
of the historically non-dominated individuals, Ht . Archive is used
to update the data at each iteration.
3.3. Tournament selection
4. Artificial Neural Network
Each individual competes in exactly two tournaments with ran-
domly selected individuals, a procedure which imitates survival of The feed-forward neural networks are the most popular archi-
the fittest in nature. tectures due to their structural flexibility, good representational
F. Hajabdollahi et al. / Applied Soft Computing 12 (2012) 3648–3655 3651

Table 1
Purchase equipment cost functions [12].

Components Investment cost


  3    
0.05 Ta − 866 K−1
CST = 0.7
a51 .PST 1+ × 1 + 5 exp ,
Steam turbine 1 − ST 10.42 K −1

−0.7
a51 = 3880.5 $k W
a
CBoiler = a1 (ṁBoiler ) 2 p T  SH/RSH
   a4  
Pe − P̄e 1 − ¯ 1 Te − T̄e
P = exp ,  = 1 + , T = 1 + a5 exp
a3 1 − 1 a6
Boiler Te − TiSH ṁRSH TeRSH − TiRSH
SH/RSH = 1 + + .
Te ṁBoiler TeRSH
−1
T̄e = 593 ◦ C, P̄e = 28 bar, ¯ 1 = 0.9, a1 = 208582 $kg s−1

a2 = 0.8, a3 = 150 bar, a4 = 7, a5 = 5, a6 = 10.42 C
Q̇Cond
CCond = a61 . + a62 .ṁCW + 70.5.Q̇Cond × (−0.6936.Ln(T̄CW − Tb ) + 2.1898)
Condenser k.
Tin
−1
a61 = 280.74 $ m−2 , a62 = 746 $kg s−1 , k = 2200 Wm−2 K−1
a2 −0.7
Deaerator CDA = a1 (mwater ) , a1 = 145315 $k W , a2 = 0.7
  −1
Pump CPump = a71 .PPump
0.71
1+ 0.2
1−Pump
, a71 = 705.48 $kg s−1

capabilities and availability of a large number of training algorithms 6. Discussion and results
[28]. Here the individual element inputs are I1 , I2 , . . . , IR multiplied
by weights w11 , w12 , . . . , w1R and the weighted values are fed to 6.1. Model verification
the summing junction. The neuron has a bias b, which is summed
with the weighted inputs to form the net input n. This sum, n, is the To have a good verification results from the developed code,
argument of the transfer function F: the results in this study are compared with the Montazer Ghaem
steam cycle power plant that is schematically shown in Fig. 1. This
a = F(n) = F(w11 I1 + w12 I2 + · · · + w1R IR + b) (16) power plant is located near the Tehran city, the capital of Iran. For
this purpose, 3% pressure drop are considered in each side of the
This network consists of neurons arranged in layers in which feed water heaters, 1% pressure drop from turbine outlet to the
every neuron is connected to all neurons of the next layer. inlet of heater, 5% pressure drop in boiler, 0.7 for pumps isentropic
efficiency, TTD = +3 degree for LP feed water heaters and TTD = −3
5. Objective functions, design parameters and constraints degree for HP feed water heaters. The results of our simulation and
the actual power plant data with the above assumption along with
In this study, the total cycle efficiency and total cost rate are the same input values listed in Table 5, are shown in Table 2. Results
considered as two objective functions. The efficiency is defined in show that the difference percentage points of two mentioned mod-
Eq. (10) and the total cost rate is computed from: eling output results are acceptable.

Ctotal = (a.ϕ.Cin + Cfuel ) × (17) 6.2. Optimization

where ϕ = 1.06 is considered for the maintenance factor, Cin is the The steam cycle optimum design parameters are obtained for
purchase cost of kth component in US dollar listed in Table 1 and a actual running power plant described in the previous section. The
is the annual cost coefficient defined as: power station should deliver 160 MW output net power. The Effi-
r ciency and total cost rate are considered as two objective functions.
a= (18)
1 − (1 + r)−y Design parameters (decision variables) and the range of their vari-
ations are listed in Table 3. The number of iterations for finding
where r and y are the interest rate and depreciation time respec- the global extremum in the whole searching domain is about
tively. 2.9 × 1034 . System is optimized for depreciation time y = 20 years,
Cfuel and in Eq. (17) are fuel cost in a year and a coefficient for interest rate r = 0.1 and 0.1 $/kg as the fuel cost. The Genetic
converting the total annual cost into the cost per time unit defined Algorithm Optimization is performed for 700 generations, using a
as below:

Cfuel = cfuel × ṁf ×  × 3600 (19) Table 2


The comparison of modeling output and the corresponding results from actual run-
ning power plant.
= ( × 3600)−1 (20)
Output variables Actual Present paper Difference (%)
where N is the annual number of the operating hours of the unit Qb (MW) 376.93 372.16 1.265
and cfuel is the fuel unit cost. Qc (MW) 222.21 218.03 1.88
To maximize the efficiency value and to minimize the total cost WHPT (MW) 44.167 46.427 5.12
WIPT (MW) 65.405 64.131 6.53
rate, fifteen design parameters including the turbine inlet temper-
WLPT (MW) 49.158 48.370 1.95
ature, boiler pressure, turbines extraction pressures, turbines and WNet (MW) 155.74 156.04 0.192
pumps isentropic efficiency, reheat pressure as well as condenser WP1 (kW) 87.182 98.5302 13.01
pressure are selected. The constrains are introduced to insure that WP2 (kW) 25.64 22.1437 13.63
the temperature and steam quality in point 7 (T7 and x7 ) are WP3 (kW) 2877.81 2769.2 3.77
ε 0.3718 0.3773 1.48
selected higher than the 40 ◦ C and 0.95 respectively.
3652 F. Hajabdollahi et al. / Applied Soft Computing 12 (2012) 3648–3655

Table 3
The design parameters and their range of variation.

Variables From To

T32 (C) 500 560


P1 (MPa) 1 2
P3 (MPa) 0.4 1
P5 (MPa) 0.3 0.4
P7 (MPa) 0.005 0.05
P8 (MPa) 0.05 0.1
P10 (MPa) 0.1 0.3
P28 (MPa) 2 4
P32 (MPa) 8 18
P1 0.5 0.8
P2 0.5 0.8
P3 0.5 0.8
HPT 0.7 0.9
IPT 0.7 0.9
LPT 0.7 0.9

Fig. 3. The distribution of Pareto-optimal points solutions using NSGA-II.

Optimum values of two objectives for five typical points from


A to E (Pareto-optimal fronts) along with the actual data from the
running power plant are listed in Table 4.
The optimization results in design point E show 3.76% increase
in efficiency and 3.84% decrease in total cost rate when it compared
with the actual data of the running power plant which is a notice-
able improvement. It is clear that the design point E is dominant
over the actual data of the running power plant in both efficiency
and total cost rate.
To provide a useful tool for the optimal design of the steam tur-
bine power plant, the following equation for efficiency versus the
total cost rate is curved for the Pareto curve (Fig. 3).

4.007ε2 − 42.67ε + 18.12


Ctotal ($/s) = (21)
ε2 − 37.78ε + 16.12
Eq. (21) is valid in the range of 0.3915 < ε < 0.4177 for effi-
ciency. The interesting point in Eq. (21) is that considering a
Fig. 2. Results of all evaluations during 700 generations using NSGA-II. A clear numerical value for the efficiency in mentioned range, provides the
approximation of the Pareto front is visible on the lower part of the figure. minimum total cost rate for that optimal point.
Optimum values of fifteen design parameters for five typical
search population size of M = 150 individuals, crossover probability points from A to E (Pareto-optimal fronts) are listed in Table 5.
of pc = 0.9, gene mutation probability of pm = 0.035 and controlled It results show that the all turbines isotropic efficiency should be
elitism value c = 0.55. located at their maximum range to makes the system thermoeco-
The results of optimum efficiency and total cost rate for all nomically optimized.
points evaluated over 700 generations are depicted in Fig. 2. The To have a better insight in some important optimum design
Pareto-optimal curve (best rank) is clearly visible in the lower part parameters such as boiler exhaust steam pressure, HP turbine inlet
of the figure which is separately shown in Fig. 3. The Pareto opti- temperature (TIT) and reheat pressure, the distribution of efficiency
mum results clearly reveal the conflict between two objectives, the versus the mentioned parameters for optimum points in Pareto
efficiency and the total cost rate. Any change that increases the effi- curve are shown in Figs. 4–6 respectively. It can be concluded
ciency, leads to an increase in the total cost rate and vice versa. This that the efficiency increases by increase of boiler pressure, TIT and
shows the need for multi-objective optimization technique in opti- reheat pressure in the optimum selected range with the fixed net
mal design of steam cycle power plant. It is shown in Fig. 3, which power.
the maximum efficiency exists at design point A (0.4177), while
the total cost rate is the biggest at this point. On the other hand the 6.3. Selection of final optimum design
minimum total cost rate occurs at design point E (1.376 $/s), with
a smallest efficiency value (0.3915) at that point. Design point A is Each point on the Pareto frontier has the potential of final
the optimal situation at which, efficiency is a single objective func- optimum design. However, selection of a single optimum point
tion, while design point E is the optimum condition at which the from existing points on the Pareto front needs a process of
total cost rate is a single objective function. decision-making. In fact, this process is mostly carried out based

Table 4
The optimum values of efficiency and total cost rate for the design points A–E in Pareto- optimal fronts.

A B C D E Actual data

Efficiency 0.4177 0.4137 0.4086 0.4027 0.3915 0.3773


Total cost rate ($/s) 1.954 1.749 1.597 1.489 1.376 1.431
F. Hajabdollahi et al. / Applied Soft Computing 12 (2012) 3648–3655 3653

Table 5
The optimum values of design parameters for optimum selected points A–E in Pareto optimum front.

Variables A B C D E Actual data

T32 (C) 559.9635 557.2884 550.1593 548.1850 545.7967 537.7778


P1 (MPa) 1.0479 1.0108 1.0055 1.0653 1.0201 1.4989
P3 (MPa) 0.5514 0.5502 0.5201 0.5188 0.5049 0.8191
P5 (MPa) 0.3530 0.3510 0.3413 0.3434 0.3301 0.4216
P7 (MPa) 0.0078 0.0078 0.0078 0.0078 0.0078 0.00847
P8 (MPa) 0.0521 0.0504 0.0509 0.0504 0.0507 0.0778
P10 (MPa) 0.3000 0.2996 0.2999 0.3000 0.2999 0.2764
P28 (MPa) 2.2177 2.1892 2.1129 2.1352 2.0886 3.1336
P32 (MPa) 17.7773 14.8543 12.7300 10.5678 8.0183 12.5110
P1 0.6475 0.6834 0.7975 0.7346 0.7125 0.7
P2 0.6481 0.7870 0.5755 0.6235 0.6213 0.7
P3 0.7982 0.7979 0.7967 0.7965 0.7795 0.7
HPT 0.8996 0.8994 0.8994 0.8997 0.8992 0.82
IPT 0.8994 0.8994 0.8997 0.8956 0.8738 0.86
LPT 0.9000 0.9000 0.9000 0.9000 0.9000 0.8

In the LINMAP method, each objective is nondimensionalized


using the following relation:
Fij
Fijn =
(22)
m
2
(F )2
i=1 ij

where i is the index for each point on the Pareto frontier, j is the
index for each objective in the objectives space and m denotes
the number of points in the Pareto front. After nondimensionaliza-
tion of two objectives, the distance of each solution on the Pareto
frontier from the ideal point obtained. The closest point of Pareto
frontier to the equilibrium point (the design point D) might be
considered as a desirable final solution with the 0.4027 thermal effi-
ciency and 1.489 $/s total cost rate along with its optimum design
parameters listed in Table 5.

Fig. 4. Distribution of thermal efficiency versus boiler exhaust steam pressure for 6.4. Correlations between the objective functions and design
optimum points in Pareto optimum front in Fig. 3. variables

The main goal of this section is to estimate a correlation between


on engineering experiences and importance of each objective for
two optimum objective functions and fifteen decision variables for
decision makers. The process of final decision-making in Fig. 3
a specified case study. Actually the designer can find the minimum
is usually performed with the aid of a hypothetical point named
of total cost rate for any choosing efficiency in the range of 0.3915 <
as equilibrium point, that both objectives have their optimal val-
ε < 0.4177 using Eq. (21). Using these two optimum values (total
ues independent of the other objectives [29]. It is clear that it is
coat rate and efficiency) and this correlation, designers can find
impossible to have both objectives at their optimum point, simul-
the corresponding values of fifteen optimum design parameters.
taneously. The equilibrium point is not a solution located on the
For this purpose, the Artificial Neural Network by means of a hid-
Pareto frontier. In this paper, LINMAP method was used to find the
den layer with two neurons for two input parameters (efficiency
final optimum solution in Pareto front [29].

Fig. 5. Distribution of thermal efficiency versus turbine inlet temperature (TIT) for Fig. 6. Distribution of thermal efficiency versus reheat pressure for optimum points
optimum points in Pareto optimum front in Fig. 3. in Pareto optimum front in Fig. 3.
3654 F. Hajabdollahi et al. / Applied Soft Computing 12 (2012) 3648–3655

Table 6
The values of constants in Eq. (22) for the estimated optimum design parameters in the presented case study.

b1 b2 b3 b4 b5 w1 w2 w3 w4 × 103

T32 (C) −2.045 157.77 93.286 −1634.558 882.270 −65.027 −337.287 505.5 822.28
P1 (MPa) −2.045 157.77 93.286 −1634.558 0.808 −65.027 0.2226 505.5 −10.991
P3 (MPa) −2.045 157.77 93.286 −1634.558 .359 −65.027 0.154 505.5 −1.4298
P5 (MPa) −2.045 157.77 93.286 −1634.558 0.299 −65.027 0.0010777 505.5 −0.023943
P7 (MPa) −2.045 157.77 93.286 −1634.558 0.00775 −65.027 0.00006784 505.5 −0.0012728
P8 (MPa) −2.045 157.77 93.286 −1634.558 0.0426 −65.027 0.00875 505.5 −0.14148
P10 (MPa) −2.045 157.77 93.286 −1634.558 0.300 −65.027 0.00065 505.5 0.1233
P28 (MPa) −2.045 157.77 93.286 −1634.558 6.597 −65.027 −5.562 505.5 9.9576
P32 (MPa) −2.045 157.77 93.286 −1634.558 214.015 −65.027 −204.982 505.5 −1111.2
P1 −2.045 157.77 93.286 −1634.558 1.356 −65.027 −0.6535 505.5 32.139
P2 −2.045 157.77 93.286 −1634.558 0.342 −65.027 0.3133 505.5 −17.67
P3 −2.045 157.77 93.286 −1634.558 0.895 −65.027 −0.1023 505.5 0.4927
HPT −2.045 157.77 93.286 −1634.558 0.936 −65.027 − 0.0388 505.5 0.72803
IPT −2.045 157.77 93.286 −1634.558 0.910 −65.027 0.01123 505.5 0.15629
LPT −2.045 157.77 93.286 −1634.558 0.904 −65.027 −0.00492 505.5 0.11332

Fig. 9. Comparison of optimization results and predicted results by ANN for total
Fig. 7. The schematic diagram of neural network with two input neurons and fifteen
cost rate.
output neurons.

the linear transfer function (n) was applied at the output layout.
and total cost rate) and an output layer with fifteen neurons for
A close form equation for decision variables versus objective func-
fifteen output parameters (decision variables) were applied using
tions (efficiency and total cost) was derived by training the network
feed forward algorithm (Fig. 7).
as follows:
Tangent-sigmoid transfer function (−1 + 2/(1 + e−2n )) was used
at the input layer with n as the transfer function input. Furthermore b1
Decision variables =
1 + e(w1 ε+w2 Ctotal +b2 )
b3
+ + b5 (22’)
1 + e(w3 ε+w4 Ctotal +b4 )
where b is the bias value and w is weighting function shown in
Table 6, for the each decision variables. The corresponding errors
for the estimated efficiency and the total cost rate (two objective
functions) are shown in Figs. 8 and 9. Results show that with apply-
ing the above proposed correlations (Eq. (22)), the estimated total
cost rate and efficiency, are accurate within −0.8% to +0.3% and −2%
to 1% respectively, which are acceptable for engineering problems.

7. Conclusions

A steam cycle power plant was optimally designed using the


fast and elitist Non-dominated Sorting Genetic Algorithm (NSGA-
II) technique. The design parameters (decision variables) were
the turbine inlet temperature, boiler pressure, turbines extraction
Fig. 8. Comparison of optimization results and predicted results by ANN for effi- pressures, turbines and pumps isentropic efficiency, reheat pres-
ciency. sure as well as condenser pressure. In this presented optimization
F. Hajabdollahi et al. / Applied Soft Computing 12 (2012) 3648–3655 3655

problem, the efficiency and total cost rate were considered as two [10] D. Peidro, P. Vasant, Transportation planning with modified S-curve member-
objective functions (the efficiency was maximized and total cost ship functions using an interactive fuzzy multi-objective approach, Applied
Soft Computing 11 (2011) 2656–2663.
rate was minimized). A set of Pareto optimal front points were [11] M. Rosen, I. Dincer, Exergoeconomic analysis of power plants operating on
extracted. The results revealed the level of conflict between the various fuels, Applied Thermal Engineering 23 (2003) 643–658.
two objectives. The optimization results in design point E show [12] M. Ameri, P. Ahmadi, A. Hamidi, Energy and exergy and exergoeconomic anal-
ysis of a steam power plant: a case study, International Journal of Energy
3.76% increase in efficiency and 3.84% decrease in total cost rate Research 33 (2009) 499–512.
simultaneously, compared with the actual data of the running [13] P. Ahmadi, I. Dincer, Thermodynamic analysis and thermoeconomic optimiza-
power plant. It was also observed that the efficiency increases tion of a dual pressure combined cycle power plant with a supplementary firing
unit, Energy Conversion and Management 52 (2011) 2296–2308.
by the increase of boiler pressure, TIT and reheat pressure in the
[14] P. Ahmadi, I. Dincer, Exergoenvironmental analysis and optimization of a
optimum selected range with the fixed net power. Furthermore cogeneration plant system using Multimodal Genetic Algorithm (MGA), Energy
a correlation between the optimal values of two objective func- 35 (2010) 5161–5172.
[15] P. Ahmadi, I. Dincer, Thermodynamic and exergoenvironmental analyses, and
tions was proposed. Finally a correlation between two optimum
multi-objective optimization of a gas turbine power plant, Applied Thermal
objective functions and fifteen decision variables for the presented Engineering 31 (2011) 2529–2540.
case study were proposed with acceptable accuracy using Artificial [16] B. Sahin, K. Ali, Thermo-dynamic analysis of a combined Carnot cycle with
Neural Network. This correlation enables designers to find the internal irreversibility, Energy 20 (1995) 1285–1289.
[17] C. Koch, F. Cziesla, G. Tsatsaronis, Optimization of combined cycle power plants
corresponding fifteen optimum design parameters for minimum using evolutionary algorithms, Chemical Engineering and Processing 46 (2007)
total cost rate and maximum efficiency. 1151–1159.
[18] S. Bracco, S. Sir, Exergetic optimization of single level combined gasesteam
power plants considering different objective functions, Energy 35 (2010)
References 5365–5373.
[19] M. Valdes, M.D. Duran, A. Rovira, Thermoeconomic optimization of combined
[1] H. Hajabdollahi, P. Ahmadi, I. Dincer, An exergy-based multi objective optimiza- cycle gas turbine power plants using genetic algorithms, Applied Thermal Engi-
tion of a heat recovery steam generator (HRSG) in a combined cycle power plant neering 23 (2003) 2169–2182.
(CCPP) using evolutionary algorithm, International Journal of Green Energy 8 [20] I.H. Aljundi, Energy exergy analysis of a steam power plant in Jordan, Applied
(2011) 44–64. Thermal Engineering 29 (2009) 324–328.
[2] S. Sanaye, H. Hajabdollahi, Multi-objective optimization of rotary regenerator [21] M.M. El-Wakil, Powerplant Technology, McGraw-Hill, 2002.
using genetic algorithm, International Journal of Thermal Sciences 48 (2009) [22] C.M. Fonseca, P.J. Fleming, Multiobjective optimization, in: T. Back, D.B. Fogel, Z.
1967–1977. Michalewicz (Eds.), Handbook of Evolutionary Computation, Oxford University
[3] S. Sanaye, H. Hajabdollahi, Thermal-economic multi-objective optimization of Press, 1997.
plate fin heat exchanger using genetic algorithm, Applied Energy 87 (2010) [23] J.D. Schaffer, Multiple objective optimization with vector evaluated genetic
1893–1902. algorithms, in: Proceedings of the International Conference on Genetic Algo-
[4] S. Sanaye, H. Hajabdollahi, Multi-objective optimization of shell and tube heat rithm and their Applications, 1985.
exchangers, Applied Thermal Engineering 30 (2010) 1937–1945. [24] N. Srinivas, K. Deb, Multi-objective optimization using non-dominated sort-
[5] P. Ahmadi, H. Hajabdollahi, I. Dincer, Cost and entropy generation minimization ing in genetic algorithms, Journal of Evolutionary Computation 2 (3) (1994)
of a cross flow Plate-Fin Heat Exchanger (PFHE) using multi-objective genetic 221–248.
algorithm, Journal of Heat Transfer – Transactions of the ASME 133 (2011) [25] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multi-objective
021801. genetic algorithm: NSGA-II, IEEE Transactions on Evolutionary Computation 6
[6] S.O.T. Ogaji, R. Singh, Advanced engine diagnostics using artificial neural net- (2) (2002) 182–197.
works, Applied Soft Computing 3 (2003) 259–271. [26] K. Deb, T. Goel, Controlled elitist non-dominated sorting genetic algo-
[7] I. Bertini, M.D. Felice, A. Pannicelli, S. Pizzuti, Soft computing based optimization rithms for better convergence, in: Proceedings of the First International
of combined cycled power plant start-up operation with fitness approximation Conference on Evolutionary Multi-criterion Optimization, Zurich, 2001,
methods, Applied Soft Computing 11 (2011) 4110–4116. pp. 385–399.
[8] T. Ganesan, P. Vasant, I. Elamvazuthi, Optimization of nonlinear geological [27] K. Deb, Multi-objective Optimization Using Evolutionary Algorithms, John
structure mapping using hybrid neuro-genetic techniques, Mathematical and Wiley and Sons Ltd., Chichester, 2001.
Computer Modelling 54 (2011) 2913–2922. [28] S. Haykin, Neural Networks: A Comprehensive Foundation, 2nd ed., Prentice-
[9] M. Díaz-Madroñero, D. Peidro, P. Vasant, Vendor selection problem by using Hall, New York, 1999.
an interactive fuzzy multi-objective approach with modified S-curve mem- [29] H. Sayyaadi, R. Mehrabipour, Efficiency enhancement of a gas turbine cycle
bership functions, Computers and Mathematics with Applications 60 (2010) using an optimized tubular recuperative heat exchanger, Energy 38 (2012)
1038–1048. 362–375.

Vous aimerez peut-être aussi