Sie sind auf Seite 1von 8

Parametric optimization design for supercritical CO

2
power cycle using genetic
algorithm and articial neural network
Jiangfeng Wang
a
, Zhixin Sun
a
, Yiping Dai
a,
*
, Shaolin Ma
b
a
Institute of Turbomachinery, Xian Jiaotong University, No.28 Xianning West Road, Xian 710049, PR China
b
Dongfang Steam Turbine Works, Deyang 618201, PR China
a r t i c l e i n f o
Article history:
Received 16 April 2009
Received in revised form 16 June 2009
Accepted 28 July 2009
Available online 20 August 2009
Keywords:
Articial neural network
Genetic algorithm
Optimization
Power cycle
Supercritical CO
2
Waste heat recovery
a b s t r a c t
Supercritical CO
2
power cycle shows a high potential to recover low-grade waste heat due to its better
temperature glide matching between heat source and working uid in the heat recovery vapor generator
(HRVG). Parametric analysis and exergy analysis are conducted to examine the effects of thermodynamic
parameters on the cycle performance and exergy destruction in each component. The thermodynamic
parameters of the supercritical CO
2
power cycle is optimized with exergy efciency as an objective func-
tion by means of genetic algorithm (GA) under the given waste heat condition. An articial neural net-
work (ANN) with the multi-layer feed-forward network type and back-propagation training is used to
achieve parametric optimization design rapidly. It is shown that the key thermodynamic parameters,
such as turbine inlet pressure, turbine inlet temperature and environment temperature have signicant
effects on the performance of the supercritical CO
2
power cycle and exergy destruction in each compo-
nent. It is also shown that the optimum thermodynamic parameters of supercritical CO
2
power cycle
can be predicted with good accuracy using articial neural network under variable waste heat conditions.
2009 Elsevier Ltd. All rights reserved.
1. Introduction
In recent years, there have been a large number of waste heats
being released into the environment, such as exhaust gases from
turbines and engines and waste heat from industrial plants, which
lead to serious environmental pollution. Therefore, more and more
attention has been paid to low-grade waste heat recovery for its
potential in reducing fossil fuel consumption and alleviating envi-
ronmental problems.
Supercritical CO
2
power cycle shows a high potential to recover
low-grade waste heat. Because there is a better temperature glide
matching between heat source and working uid in the heat recov-
ery vapor generator, and CO
2
can easily reach its supercritical state
(the critical pressure and temperature of CO
2
are 7.38 MPa and
31.1 C, respectively). Supercritical CO
2
power cycle also shows
no pinch limitation in the heat recovery vapor generator. In addi-
tion, CO
2
is inexpensive, non-toxic, non-explosive, and abundant
in nature, and the knowledge of its thermodynamic properties is
sufcient. Therefore, some researchers has explored the supercrit-
ical CO
2
power cycle.
Chen et al. [1] examined the performance of the CO
2
trans-
critical power cycle utilizing energy from low-grade waste heat
in comparison to an Organic Rankine cycle (ORC) using R123 as
working uid. They found that when utilizing the low grade heat
source with equal mean thermodynamic heat rejection tempera-
ture, the carbon dioxide trans-critical power cycle had a slightly
higher power output than the ORC. Zhang et al. investigated a ther-
modynamic cycle powered by solar energy for both power and
heat generation using supercritical carbon dioxide as a working
uid on theoretical aspect [24], and they examined the effects
of various design conditions and climate conditions on the perfor-
mances of this CO
2
-based Rankine cycle. They also set up an exper-
imental system to validate the feasibility of this supercritical
carbon dioxide cycle [58]. Cayer et al. [9] expressed a detailed
analysis of a carbon dioxide trans-critical power cycle using an
industrial low-grade stream of process gases as its heat source.
They examined the effect of high pressure on the thermal ef-
ciency, exergy efciency, total UA values (UA is the product of
the overall heat transfer coefcient by the area) and total heat ex-
change surface with xed temperature and mass ow rate of the
heat source, xed maximum and minimum temperatures in the
cycle and a xed sink temperature.
Up to the present, none of the published studies on the super-
critical CO
2
power cycle focused on the parameter optimization
to convert low-grade waste heat to useful work as much as possi-
ble under the given waste heat condition. So, the aim of this study
is to conduct the parameter optimization for the supercritical CO
2
power cycle to recover as much waste heat as possible from indus-
trial production.
0306-2619/$ - see front matter 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.apenergy.2009.07.017
* Corresponding author. Tel./fax: +86 029 82668704.
E-mail address: freego810211@gmail.com (Y. Dai).
Applied Energy 87 (2010) 13171324
Contents lists available at ScienceDirect
Applied Energy
j our nal homepage: www. el sevi er . com/ l ocat e/ apener gy
In the present paper, based on the examination of the effects of
the major thermodynamic parameters on the system performance
and the exergy destruction in each component for the supercritical
CO
2
power cycle, a parameter optimization is conducted to maxi-
mize the exergy efciency by means of genetic algorithm under
the given waste heat condition. Because the parametric optimiza-
tion using genetic algorithm takes a long time to get the optimum
values of thermodynamic parameters, which is not competent for
conducting the online optimization operation, an articial neural
network with the multi-layer feed-forward network type and
back-propagation training is also used to achieve parametric opti-
mization design rapidly.
2. Thermodynamic analysis of supercritical CO
2
power cycle
The schematic diagram and Ts diagram of the supercritical CO
2
power system are shown in Figs. 1 and 2, respectively. The cycle is
composed of following processes.
Process 23: a constant-pressure heat absorption process in the
heat recovery vapor generator;
Process 34: a non-isentropic expansion process in the turbine;
Process 41: a constant-pressure heat rejection process in the
condenser;
Process 12: a non-isentropic compression process in the
pump.
For the cycle performance simulation, the following assump-
tions are also made:
(1) The system reaches a steady state, and pressure drop in
pipes and heat losses to the environment in the condenser,
heat recovery vapor generator, turbine and pump are
neglected.
(2) The condenser outlet state is saturated liquid, and its tem-
perature is assumed to be approximately 6 C higher than
the environment temperature.
(3) The isentropic efciency of the turbine and pump are 75%
and 70%, respectively.
In the HRVG, when the turbine inlet parameters such as pres-
sure and temperature are given, the turbine inlet state is obtained.
The pump outlet temperature could be obtained by pump model.
By giving the temperature difference between pump outlet and
waste heat exhaust from the HRVG, the temperature of waste heat
exhaust could be obtained. The heat transfer process can be de-
scribed by the energy balance.
m
g
h
in;g
h
out;g
mh
3
h
2
1
Nomenclature
d expected output
E exergy, kW
Er error
h enthalpy, kJ kg
1
I exergy destruction, kW
Is sum of the weighted inputs
m mass ow rate, kg s
1
o actual output
p pressure, MPa
Q heat addition, kW
s entropy, kJ kg
1
K
1
t temperature, C
T temperature, K
w weight value
W power, kW
Greek symbols
g efciency, learning rate
l momentum factor
Subscripts
CND condenser
exg second law of thermodynamics
g waste heat
HRVG heat recovery vapor generator
in input
loss exergy loss
out output
PUMP pump
thm rst law of thermodynamics
TBN turbine
0 environment state
1,2,3,4 state point
Fig. 1. Schematic diagram of the supercritical CO
2
power system.
Fig. 2. Ts diagram of the supercritical CO
2
power system.
1318 J. Wang et al. / Applied Energy 87 (2010) 13171324
The absorption heat from the waste heat to the working uids is
Q
in
mh
3
h
2
2
In the turbine, the isentropic efciency of the turbine can be ex-
pressed as
g
TBN

h
3
h
4
h
3
h
4s
3
The generating power can be given as
W
TBN
mh
3
h
4
4
In the condenser, the rejecting heat to the environment is expressed
as
Q
out
mh
4
h
1
5
In the pump, the isentropic efciency of the pump can be expressed
as
g
PUMP

h
2s
h
1
h
2
h
1
6
The work input by the pump is
W
PUMP
mh
2
h
1
7
The thermal efciency of the supercritical CO
2
power cycle is de-
ned on the basis of the rst law of thermodynamics as the ratio
of net power output to the heat addition.
g
thm

W
TBN
W
PUMP
Q
in
8
From the viewpoint of the rst law of thermodynamics and energy
conservation used to determine the overall thermal efciency, work
and heat are equivalent. On the other hand, based on the second law
of thermodynamics, exergy quanties the difference between work
and heat in terms of irreversibility. Therefore, the exergy efciency
can evaluate the cycle performance from the energy quality and it is
chosen to be the criterion for the cycle performance evaluation to
recover low-grade waste heat.
Consider p
0
and T
0
to be the environment pressure and temper-
ature as the specied dead reference state. The following assump-
tions are made to calculate the exergy of each state point:
(a) It is assumed that only physical exergies are used for ue gas
and steam ows.
(b) Chemical exergies of the substances are neglected.
(c) Kinetic and potential exergies of materials are ignored.
The exergy of the state point can be considered as
E
i
mh
i
h
0
T
0
s
i
s
0
9
The exergy balance for an open thermodynamic system can be ex-
pressed as [10]
X
E
in

X
E
out
I 10
The exergy destruction in the HRVG can be given as
I
HRVG
E
in
E
2
E
out
E
3
11
The exergy destruction in the turbine can be given as
I
TBN
E
3
W
TBN
E
4
12
The exergy destruction in the condenser can be given as
I
CND
E
4
E
5
E
1
E
6
13
The exergy destruction in the pump can be given as
I
PUMP
W
PUMP
E
1
E
2
14
The exergy efciency of the supercritical CO
2
power cycle can be
given as
g
exg

E
in

P
I E
loss
E
in
15
where E
in
is the exergy of waste heat uid which is input to the cy-
cle and E
loss
is the exergy which is carried by the exhaust and cool-
ing water.
3. Methodology for parametric optimization
In order to recover as much waste heat as possible, it is neces-
sary to optimize the performance of the supercritical CO
2
power
cycle. The exergy efciency is selected as objective function for
parameter optimization.
3.1. Genetic algorithm
The GA, which is presented rstly by professor Holland [11], is a
stochastic global search method that simulates the natural biolog-
ical evolution. Based on the Darwinian survival-of-ttest principle,
the genetic algorithm operates on a population of potential solu-
tions to produce better and better approximations to the optimal
solution. The GA differs from more traditional optimization tech-
niques because it involves a search from a population of solutions
and not from a single point and it can prevent the convergence to
suboptimal solutions.
The GA encodes a potential solution to a specic domain prob-
lem on a simple chromosome-like data structure (which consti-
tutes an individual), where genes are parameters of the problem
to be solved. In the present study, the oat-point coding is used
in parameter optimization for the supercritical CO
2
power cycle.
Each chromosome vector is coded as a vector of oating point
numbers of the same length as the dimension of the search space.
Chromosome is dened as a real number vector, X = (x
1
, x
2
, . . ., x
n
),
x
i
2 R, i = 1,2, . . ., n, where i is ith parameter for the supercritical
CO
2
power cycle, n is the number of optimizing parameters.
The GA uses tness function to evaluate adaptability of individ-
ual without external information in the evolution search. The
adaptability is expressed by the tness value. A bigger tness value
means a better adaptability subjected to constraints and a better
viability of the individual. Fitness function which is not con-
strained by denition domain, continuity and differentiability, re-
quires that the objective function is dened as a form of non-
negative maximum. In this optimization, the exergy efciency is
selected as the tness function.
The GA operators include selection operator, crossover operator
and mutation operator. Selection operator is responsible for select-
ing the parents to create the next generation of solutions. The par-
ent is chosen with a probability based on its tness. The higher the
tness is, the higher the probability of selection is. The rank-based
model is selected for this optimization in the supercritical CO
2
power cycle.
Crossover operator is the basic operator for producing new
chromosomes. It produces new individuals that have some parts
of both parents genetic material. The simple arithmetic crossover
is applied to this optimization problem due to very simple opera-
tion, which is presented as follows:
c
1
af
1
1 af
2
c
2
af
2
1 af
1

16
where a is a random number between 0 and 1, f
1
and f
2
are parents
individuals which are selected to crossover each other, c
1
and c
2
are
children individuals which are produced by crossover.
J. Wang et al. / Applied Energy 87 (2010) 13171324 1319
Mutation is needed because even if selection and crossover
search new solutions together, they tend to cause rapid conver-
gence and there is a danger of losing potentially useful genetic
material. The role of mutation in GA is to restore lost or unexplored
genetic material into the population to prevent the premature con-
vergence of the GA to suboptimal solutions. Random mutation is
adopted to optimize the parameters for the supercritical CO
2
power cycle. It is achieved by selecting individuals from the range
of the parameter according to mutation probability.
3.2. Articial neural network
ANN is a computational system that simulates the microstruc-
ture of a biological neurons system. It mimics the learning process
of a natural brain, organizes itself to gain knowledge from given
examples and applies the knowledge to solve new problems. It
operates like a black box, only cares about the inputs and out-
puts, and requires no detailed information about the system.
ANN is able to handle noisy and incomplete data, and deal with
non-linear problems. Once it is trained it can perform prediction
at high speed for highly complex and ill-dened problems, just
as humans usually decide on an intuitional basis. Because of its
advantages, ANN has been widely used in diverse applications such
as classication, forecasting, control systems, optimization and
decision making [12].
The ANN consists of many units that represent neurons. An arti-
cial neuron architecture is shown in Fig. 3. Each unit is a basic
unit of the information process. Units are interconnected via links
that contain weight values. Weight values help the neural-network
to express knowledge. Each input is multiplied by a connection
weight and summed, then passes through a transfer function to
generate a result, and nally the output is obtained. In ANN, the
data used as inputs is transmitted through the network, layer by
layer, and a set of outputs is obtained.
The multi-layer feed forward (MLFF) neural network is the most
widely used network. A MLFF neural network typically employs
three or more layers for the architecture: an input layer, an output
layer, and at least one hidden layer, shown as in Fig. 4. In MLFF net-
works, neurons are arranged in layers with connectivity between
the neurons of different layers. The layer that receives inputs is
called the input layer, and the layer that gives the output is called
the output layer. Other layers, as they do not receive any direct in-
put or contribute to output directly, are called hidden layers. How-
ever, it is not necessary to have more than one hidden layer, since
it has been proved that one hidden layer is enough to approximate
any continuous function as long as it has a sufcient number of
neurons [13].
The transfer function for neurons in a MLFF network can be lin-
ear or non-linear. A sigmoid function is a widely used non-linear
activation function whose output lies between 0 and 1 and is de-
ned as:
f x
1
1 e
x
17
Once a neural network is congured, a training process must be
done to perform a particular function. Training is normally accom-
plished through an adaptive procedure or algorithm that adjusts
weights of the connections between the neurons, either from the
information outside the network or by the neurons themselves in
response to the input. The ANN reads the input and output values
in the training data set and changes the values of the weights to
minimize the differences between the actual outputs and expected
outputs using a learning algorithm (e.g., back-propagation algo-
rithm). Initially, due to the random weights assigned randomly to
the connections, the differences between the actual outputs and ex-
pected outputs are usually large, but through many training cycles,
the differences are reduced. If a satisfactory level of accuracy is
reached, the training stops, and the network uses the weights to
make decisions in test data.
There are different kinds of learning algorithms. One popular
algorithm is the back-propagation algorithm. Back-propagation
training is a gradient descent algorithm. It tries to improve the per-
formance of the neural network by reducing the total error by
changing the weights along its gradient.
The learning of ANN is a procedure of modifying the weights.
Fig. 5 shows the schematic diagram of error back propagation.
The error for each neuron is the difference squared between the ex-
pected output and the actual output, dened as:
Er
k1
m

1
2
o
k1
m
d
m

2
18
where o
k1
m
is the actual output for each neuron and d
m
is the ex-
pected output. Each weight w(N + 1) is updated from its previous
state value w(N) according to
w
1
w
2
w
n
y
x
i
w
i
x
2
x
n
f
weights
inputs
output
transfer
function

x
1
Fig. 3. An articial neuron model.
Fig. 4. Schematic diagram of a multi-layer feed forward neural network.
1 , k k
ji
w
k k
mj
w
, 1 +
1 + k
m
E
k
j
E
Fig. 5. Schematic diagram of error back propagation.
1320 J. Wang et al. / Applied Energy 87 (2010) 13171324
w
k;k1
ji
N 1 w
k;k1
ji
N Dw
k;k1
ji
N 19
where Dw(N) is the incremental change in the weight, which can be
written as
Dw
k;k1
ji
N gd
k
j
o
k1
i
20
where g is the learning rate, o is the corresponding output, and d
k
j
is
dened as
@Er
k
j
@o
k
j
f
0
Is
k
j
, Is is the sum of the weighted inputs.
The weight change Dw(N) in the output and hidden layer neu-
rons can be calculated using Eqs. (18) and (19), respectively.
Dw
k1;k
mj
N gd
k1
m
o
k
j
go
k1
m
d
m
f
0
Is
k1
m
o
k
j
21
Dw
k;k1
ji
N gd
k
j
o
k1
i
g
X
m
d
k1
m
w
k1;k
mj
!
f
0
Is
k
j
o
k1
i
22
Since back-propagation employs a form of gradient descent, it is
very easy for the training process to get trapped in a local minimum.
However, it can be avoided by adding a momentum term to the
weight change as follows:
DwN 1 gdo lDwN 23
where l is the momentum factor. Thus, the new value of weight be-
comes equal to the previous value of the weight plus the weight
change, which includes the momentum term. This training method
is known as back-propagation with momentum.
A higher learning rate has the advantage of faster learning, but
it may cause the weights to bounce around error minimum, thus
failing to learn properly. In contrast, a small learning rate is more
safe, but it drives the learning slowly. A momentum term is widely
applied to accelerate the learning and reduce an oscillation. Appro-
priate values of these parameters aid the network learning.
The conguration of the multilayer ANN has to be determined
by experience since there are no denitive rules to select the num-
ber of hidden layers and the number of neurons in each hidden
layer. In the present study, the input of the system is selected as
heat source temperature, and the outputs are selected as turbine
inlet pressure, turbine inlet temperature and exergy efciency.
So, the input layer in ANN has one neuron for the heat source tem-
perature, the hidden layer has six neurons, and the output layer
has three neurons for the turbine inlet pressure, the turbine inlet
temperature and the exergy efciency. Neurons in input layer have
no transfer function. Logistic sigmoid (logsig) transfer function has
been used in hidden layer and pureline transfer function in output
layer.
4. Results and discussion
The supercritical CO
2
power cycle can be heated by the waste
heat from industrial production, solar energy, geothermal heat or
any other heat sources. In the present study, the waste heat, which
is composed of 70% steam and 30% air, is used as heat source to
simulate the supercritical CO
2
power cycle. The initial parameters
used to simulate the supercritical CO
2
power cycle are shown in
Table 1.
The simulation of the supercritical CO
2
power cycle is carried
out using a simulation program written in Fortran by authors.
The thermodynamics properties of CO
2
were calculated by REF-
PROP 6.01 [14] developed by the National Institute of Standards
and Technology of the United States.
The parametric analysis is performed to evaluate the effects of
each key parameter on the supercritical CO
2
power cycle, such as
turbine inlet pressure, turbine inlet temperature and so on.
Fig. 6 shows the effect of turbine inlet pressure on the exergy
efciency for different heat source temperature while the other
parameters are kept constant as those in Table 1. It can be seen that
the exergy efciency increases at rst to maximum and then de-
creases as the turbine inlet pressure increases. It is known that
the enthalpy drop across the turbine increases as the pressure ratio
increases, thus the turbine power output increases. By subtracting
pump input from the turbine power output, the net power output
increases. This is why the exergy efciency increases at rst. But
the enthalpy gains from an increased pressure ratio do not make
up for the decrease in vapor ow rate generated by HRVG, thus
the net power output decreases afterwards, resulting in a decrease
in exergy efciency. In addition, it is found that the exergy
Table 1
Initial parameters for simulating the supercritical CO
2
power cycle.
Environment temperature (C) 15.00
Environment pressure (MPa) 0.10135
Heat source mass ow rate (kg/s) 210
Turbine isentropic efciency (%) 75.00
Pump isentropic efciency (%) 70.00
Cooling water initial temperature (C) 15.00
Cooling water mass ow rate (kg/s) 2000
Fig. 6. The effect of turbine inlet pressure on the exergy efciency for different heat
source temperature.
Fig. 7. The effect of turbine inlet temperature on the exergy efciency for different
heat source temperature.
J. Wang et al. / Applied Energy 87 (2010) 13171324 1321
efciency increases with an increasing in heat source temperature.
It is obvious that higher heat source temperature conforms to high-
er exergy efciency.
Fig. 7 shows the effect of turbine inlet temperature on the exer-
gy efciency for different heat source temperature while the other
parameters are kept constant as those in Table 1. It can be seen that
the exergy efciency increases as the turbine inlet temperature in-
creases. In addition, it is obvious that the exergy efciency in-
creases with an increase in heat source temperature.
Fig. 8 shows the effect of environment temperature on the exer-
gy efciency for different heat source temperature while the other
parameters are kept constant as those in Table 1. It is obvious that
the exergy efciency decreases with an increase in environment
temperature. The reason for this is that an increase in environment
temperature results in an increase in condensing pressure, which
reduces the turbine power.
Exergy analysis has been performed to evaluate the exergy
destructions in the system. The exergy destruction of each compo-
nent is calculated based on the assumptions in Table 1. Figs. 911
show the effects of thermodynamic parameters on the exergy
destruction in each component. It can be seen that the biggest
exergy destruction basically occurs in the HRVG followed by the
turbine, the condenser or the pump. The major exergy destruction
in the HRVG or the condenser is due to heat transfer over a nite
temperature difference, and the exergy destruction in the turbine
or pump is due to the friction losses of the ow through the turbine
or the pump, the non-ideal adiabatic expansion or compression in
the turbine or the pump, and the corresponding irreversibilities.
The turbine inlet pressure has great effect on the exergy
destruction in each component, shown in Fig. 9. As the turbine in-
let pressure increases, the exergy destruction in the HRVG de-
creases. With the increasing turbine inlet pressure, the exergy
destruction in the turbine or pump increases. This is because an in-
crease in the turbine inlet pressure results in an increase in pres-
sure difference through the turbine or pump, which leads to an
increase in entropy through the turbine or pump. In addition, when
the turbine inlet pressure increases, the turbine outlet temperature
decreases. This could result in a decrease in heat transfer temper-
ature difference for the condenser, thus, the exergy destruction in
the condenser decreases.
As shown in Fig. 10, it can be seen that as the turbine inlet tem-
perature increases, the exergy destruction in the HRVG increases,
and the exergy destruction in the condenser increases. This is be-
cause an increase in the turbine inlet temperature can result in a
decrease in the heat transfer temperature difference for the HRVG.
In addition, due to an increase in the turbine inlet temperature, the
turbine outlet temperature increases, thus, the heat transfer tem-
perature difference in the condenser increases, resulting in an in-
crease in the exergy destruction. It also can be seen that as the
turbine inlet temperature increases, the exergy destruction in the
pump and the turbine decreases.
As shown in Fig. 11, it can be seen that the exergy destruction in
the turbine and pump decrease as the environment temperature
Fig. 10. The exergy destruction in each component vs. turbine inlet temperature.
Fig. 11. The exergy destruction in each component vs. environment temperature.
Fig. 8. The effect of environment temperature on the exergy efciency for different
heat source temperature.
Fig. 9. The exergy destruction in each component vs. turbine inlet pressure.
1322 J. Wang et al. / Applied Energy 87 (2010) 13171324
increases. This is because that a condensing pressure increases
with the increasing environment temperature, thus, the pressure
differences through the turbine and the pump decrease, resulting
in an increase in entropy through the turbine or pump. In addition,
due to the increasing condensing pressure, the turbine outlet tem-
perature increases, and this results in an increase in the heat trans-
fer temperature difference for the condenser, thus, the exergy
destruction in the condenser increases.
From the discussions above, the turbine inlet pressure and tem-
perature have great impact on the exergy efciency of the super-
critical CO
2
power cycle under a given waste heat condition, so it
is necessary to conduct the parameter optimization to obtain the
maximum exergy efciency. It is obvious that the parameters cho-
sen for optimization are turbine inlet pressure and turbine inlet
temperature, and the objective function is the exergy efciency.
In addition, because the waste heat from industrial plant uctuates
greatly than environment condition, the condition of parametric
optimization could be assumed that the waste heat temperature
varies from60 C to 90 C and the environment temperature is kept
to be constant. Table 2 shows the optimization results of the super-
critical CO
2
power cycle using genetic algorithm under the differ-
ent waste heat temperature and the same environment.
Genetic algorithm is a good method to conduct optimization
problem, but the process is so long that it isnt suitable for online
optimization. In order to obtain the optimum values of thermody-
namic parameters for the supercritical CO
2
power cycle rapidly to
reach the optimum performance under the variable waste heat
conditions, an articial neural network is used to achieve paramet-
ric optimization design.
The parametric optimization results by the GA shown in Table 2
are selected as training samples for ANN to complete the training
process. After training, the ANN can achieve the parametric optimi-
zation with high speed. The number of training samples is four. A
back-propagation momentum learning method with a learning
rate of 0.2 and a momentum factor of 0.95 is adopted. The error
for back-propagation training is 1 10
5
and the training epoch
is set to 1000. When the ANN nishes the training process, several
examples gained by GA are selected as test data to predict the opti-
mum parameters for the supercritical CO
2
power cycle. Table 3
shows the predicted optimum parameters using ANN and the com-
parison results of the optimum parameters between ANN and GA.
It is indicated that the predicted parameters gained by ANN are in
substantial agreement with those gained by GA. The errors be-
tween ANNs data and GAs data are less than 0.1% with good
accuracy.
5. Conclusions
In the present study, the effects of parameters on the thermody-
namic performance and the exergy destruction in each component
are examined for the supercritical CO
2
power cycle in waste heat
recovery eld. Parameter optimization is conducted with exergy
efciency as objective function by means of genetic algorithm
under different waste heat conditions. An articial neural network
Table 3
The predicted optimum parameters using ANN and the comparison of optimum parameters between ANN and GA.
ANN GA Error (%) ANN GA Error (%) ANN GA Error (%)
Heat source initial temperature (C) 85 75 65
Turbine inlet pressure (MPa) 10.348 10.35 0.0193 9.5505 9.552 0.0157 8.7616 8.764 0.0274
Turbine inlet temperature (C) 74.928 74.97 0.0560 64.927 64.95 0.0354 54.913 54.94 0.0491
Exergy efciency (%) 29.555 29.58 0.0845 27.446 27.46 0.0510 24.687 24.7 0.0526
Table 2
The optimization results of the supercritical CO
2
power cycle under given waste heat conditions using GA.
Conditions Population size 50 50 50 50
Crossover probability 0.95 0.95 0.95 0.95
Mutation probability 0.05 0.05 0.05 0.05
Stop generation 200 200 200 200
Heat source initial temperature (C) 90 80 70 60
Optimization results Turbine inlet pressure (MPa) 10.751 9.948 9.155 8.376
Turbine inlet temperature (C) 79.93 69.94 59.91 49.96
Exergy efciency (%) 30.44 28.57 26.16 23.00
Additional results CO
2
mass ow rate (kg/s) 174.300 151.620 126.855 99.383
Heat source exhaust temperature (C) 39.71 38.40 37.06 35.71
Cooling water temperature (C) 18.89 18.24 17.58 16.92
Turbine outlet pressure (MPa) 5.857 5.857 5.857 5.857
Turbine outlet temperature (C) 35.09 31.28 27.75 24.61
Pump inlet temperature (C) 21.00 21.00 21.00 21.00
Pump outlet temperature (C) 29.71 28.40 27.06 25.71
Turbine work (kW) 3177.945 2279.117 1510.714 884.826
Pump work (kW) 1557.877 1136.782 769.522 462.196
Net power output (kW) 1620.068 1142.335 741.192 422.630
HRVG heat input (kW) 34129.961 28212.232 22323.017 16455.200
Condenser heat rejection (kW) 32509.892 27069.897 21581.825 16032.570
Thermal efciency (%) 4.75 4.05 3.32 2.57
Heat source exergy input (kW) 5322.184 3997.787 2833.513 1837.328
HRVG exergy destruction (kW) 1094.358 882.877 679.861 483.946
Turbine exergy destruction (kW) 995.094 721.677 483.445 285.874
Condenser exergy destruction (kW) 578.731 469.873 372.299 281.291
Pump exergy destruction (kW) 445.330 326.280 221.787 133.782
Exhaust exergy loss (kW) 371.374 303.902 238.902 176.741
Cooling water exergy loss (kW) 217.229 150.843 96.027 53.065
J. Wang et al. / Applied Energy 87 (2010) 13171324 1323
with the multi-layer feed-forward network type and back-propa-
gation training is used to achieve parameter optimization design
rapidly.
From the discussions above, it can be concluded that the waste
heat condition such as heat source temperature and environment
temperature, andthe major thermodynamic parameters suchas tur-
bineinlet pressure, turbineinlet temperaturehavesignicant effects
on the exergy efciency of the supercritical CO
2
power cycle.
It is also shown that the thermodynamic parameters of the
supercritical CO
2
power cycle can be predicted with good accuracy
by using articial neural network under variable waste heat condi-
tions to reach the optimum performance.
Acknowledgement
The study presented in this paper is nancially supported by
National High Technology Research and Development Program
(863Program) of China(Grant Nos. 2009AA05Z205, 2007AA05Z251).
References
[1] Chen Y, Lundqvist P, Johansson A, Platell P. A comparative study of the carbon
dioxide transcritical power cycle compared with an organic Rankine cycle with
R123 as working uid in waste heat recovery. Appl Therm Eng
2006;26:21427.
[2] Zhang XR, Yamaguchi H, Uneno D, Fujima K, Enomoto M, Sawada N. Analysis of
a novel solar energy-powered Rankine cycle for combined power and heat
generation using supercritical carbon dioxide. Renew Energy
2006;31:183954.
[3] Zhang XR, Yamaguchi H, Fujima K, Enomoto M, Sawada N. Theoretical analysis
of a thermodynamic cycle for power and heat production using supercritical
carbon dioxide. Energy 2007;32:5919.
[4] Zhang XR, Yamaguchi H, Uneno D. Thermodynamic analysis of the CO
2
-based
Rankine cycle powered by solar energy. Int J Energy Res 2007;31:141424.
[5] Zhang XR, Yamaguchi H, Fujima K, Enomoto M, Sawada N. A feasibility study of
CO
2
-based Rankine cycle powered by solar energy. JSME Int J B 2005;48:5407.
[6] Yamaguchi H, Zhang XR, Fujima K, Enomoto M, Sawada N. Solar energy powered
Rankine cycle using supercritical CO
2
. Appl Therm Eng 2006;26:234554.
[7] Zhang XR, Yamaguchi H, Fujima K, Enomoto M, Sawada N. Study of solar
energy powered transcritical cycle using supercritical carbon dioxide. Int J
Energy Res 2006;30:111729.
[8] Zhang XR, Yamaguchi H, Uneno D. Experimental study on the performance of
solar Rankine systemusing supercritical CO
2
. RenewEnergy 2007;32:261728.
[9] Cayer E, Galanis N, Desilets M, Nesreddine H, Roy P. Analysis of a carbon
dioxide transcritical power cycle using a low temperature source. Appl Energy
2009;86:105563.
[10] Lior N, Zhang N. Energy, exergy, and Second Law performance criteria. Energy
2007;32:28196.
[11] Holland JH. Adaptation in nature and articial systems: an introductory
analysis with applications to biology, control and articial
intelligence. Massachusetts: MIT Press; 1992.
[12] Kalogirou SA. Applications of articial neural-networks for energy systems.
Appl Energy 2000;67:1735.
[13] Cybenko G. Approximation by superpositions of a sigmoidal function. Math
Control Signal 1989;2:30314.
[14] NIST Standard Reference Database 23, NIST thermodynamic and transport
properties of refrigerants and refrigerant mixtures REFPROP, Version 6.01;
1998.
1324 J. Wang et al. / Applied Energy 87 (2010) 13171324

Das könnte Ihnen auch gefallen