See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/253122833

Development of fast calculation method for ammonia refrigeration cycle and

parameter adjustment with genetic algorithm

Article · January 2007

CITATIONS READS

0 1,783

3 authors, including:

Eisuke Togashi Shin-ichi Tanabe

Kogakuin University Waseda University
29 PUBLICATIONS 53 CITATIONS 348 PUBLICATIONS 6,357 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Eisuke Togashi on 29 December 2013.

The user has requested enhancement of the downloaded file.

 Proceedings: Building Simulation 2007

DEVELOPMENT OF FAST CALCULATION METHOD FOR

AMMONIA REFRIGERATION CYCLE AND
PARAMETER ADJUSTMENT WITH GENETIC ALGORITHM

Eisuke Togashi, Shin-ichi Tanabe and Tomohiro Ataku

Waseda University, Dept. of Architecture, 3-4-1 Okubo Shinjuku-ku, Tokyo 169-8555

E-mail : [email protected] http://www.tanabe.arch.waseda.ac.jp

network to increase the calculation speed of a

ABSTRACT refrigeration cycle. The accuracy of the
A parameter adjustment method for an ammonia heat approximation and the calculation speed are
pump chiller using a genetic algorithm (GA) was evaluated. After that, some parameters are estimated
developed. The parameters were automatically from a performance data at rated point with using the
adjusted by the data of performance at rated point. genetic algorithm. Output values from the simulation
Upon using the proposed adjustment method for model are compared with the rated value.
parameters, output values of the simulation model
agreed quite well with the performance data at rated PHYSICAL MODEL OF
point. The deviation of output was less than 1.2 [%] REFRIGERATION CYCLE
from the rated value. To speed up the adjustment
The refrigeration cycle consists of some smaller
process, a new approximation method with neural
components. In this study, the model of a compressor,
network was also proposed. This method decreases
an evaporator, a condenser, and an expansion valve
the calculation time required in obtaining refrigerant
are all combined to calculate the refrigeration cycle.
thermodynamic properties. The time required to
A brief calculation flow chart with each component
calculate saturated and other refrigerant states were
is described below.
decreased by 14 times and 33 times respectively,
while the average relative error was less than 0.04 Compressor
[%] and 0.5 [%] when compared to the exact solution
from REFPROP. The time required to calculate the A compression work of the compressor can be
refrigerant cycle decreased 20 times, while relative generally expressed as a function of inlet and outlet
error was within 3 [%]. pressures as shown in equation (1) (Fujiwara 1985).
A flow rate of the refrigerant is a function of a
KEYWORDS density, and it is calculated by solving an equation of
state (EOS).
Ammonia heat pump, Parameter adjustment,
( pol −1) / pol
REFPROP, Approximation, Refrigeration cycle pol − 1 ⎧⎪⎛ Pout ⎞ ⎫⎪
W = η vVol Pin ⎨⎜⎜ ⎟⎟ − 1⎬ (1)
pol P
INTRODUCTION ⎩⎪⎝ in ⎠ ⎪⎭

Although a detailed physical model of the Gout = ηv ⋅ Vol ⋅ ρin (2)

refrigeration cycle is suitable for evaluating a ρ in = f EOS ( Pin , H in ) (3)
performance at a low load operation or during a
starting up process, the HVAC simulation program Evaporator
uses more of the performance data to simulate the In the evaporator, by drawing heat from an ambient
heat source equipment. air, the two-phase refrigerant could change its state to
There might be two reasons for this. The first reason super heated. Equation (4) shows the heat balance in
is because a detailed physical model requires greater two-phase region, where Δθ is a logarithmic mean
computation time than the simple model which temperature difference (LMTD) expressed in
utilizes the performance data file. It is mainly due to equation (6). Similarly, the heat balance in the super
an iterative computation on calculating the heated region can be expressed in equation (5).
thermodynamic properties. The second reason is Temperatures necessary to calculate the LMTD could
because a detailed physical model needs much more be calculated by solving the EOS.
physical parameters to simulate; it is usually difficult Figure 1 shows a calculation flow chart of the
to look for all the parameters. evaporator. The output refrigeration state is either
In this study, thermodynamic properties of a two-phase or super heated, and an iterative
refrigerant are approximately evaluated by neural

- 596 -
 Proceedings: Building Simulation 2007

calculation is needed to determine a suitable enthalpy Gin = Cv ρin Pin − Pout (9)
of the output refrigerant.
ρ in = f EOS ( Pin , H in ) (10)
In this model, a cooling air state is assumed to move
along the saturation line on the psychometric chart Condenser
when cooling air is saturated.
The model of the condenser is almost as same as that
of the evaporator. In this study, though an output
Qsh = ARsh K sh Δθ sh (4) state of the refrigerant could be sub cooled, a heat
(
= Gref H ref ,out − H ref ,vs ) transfer area of sub cooled and two-phase are treated
= Gair (H air ,in − H air , sh,out )
as one unit, because a heat exchange at sub cool state
is very low.
Qtp = ARtp Ktp Δθtp (5) Expansion valve
(
= Gref H ref ,vs − H ref ,in ) Equation (9) shows pressure drop at an expansion
= Gair (H air ,in − H air ,tp ,out ) valve at a given flow rate. It is a function of an inlet
refrigerant density which can be calculated by
Δθ =
(T1in − T2in ) − (T1out − T2out ) (6) solving the EOS.
ln ((T1in − T2in ) / (T1out − T2out ))
Refrigeration cycle
T = f EOS (P, H ) (7)
Finally, all four components (compressor, evaporator,
Tvs = f EOS (Pvs ) (8) condenser, expansion valve) are connected together
and state variables are solved simultaneously.
Figure 2 shows a calculation flow chart of the
Input: Gref,in, Hin, Pin, Tair,in, Gair,in refrigeration cycle. According to this chart, the inlet
enthalpy and the inlet pressure of each components,
Calculate saturated state (Tref,vs, Href,vs) at Pin
compression work, and flow coefficient of expansion
Calculate Qvs from Href,vs, Href,in, Gref
Calculate Hair,tp,out, Tair,tp,out from Qvs, Hair,in, Gair are treated as state variables. The object of this
Calculate Δθvs from Tair,tp,in, Tair,vs,out, Tref,vs, Tref,in calculation is to estimate an electricity consumption
Calculate Rvs from A, Ktp, Δθtp
Yes: Output refrigerant No: Output refrigerant
state is super heated state is two phase Input: Tair,in, Gair,in, Twat,in, Gwat,in
Rvs < 1.0
Assumption P0, P1, H0, H1, H2, W, Cv
Tair,tp,out = Tair,vs,out Assumption Href,tp,out

Assumption: Href,sh,out Calculate Tref,tp,out Pcmp,in = Pevp,in = P0, Pexp,in = Pcnd,in = P1

from Pin, Href,tp,out Hevp,in = H0, Hcmp,in = H1, Hcnd,in = H2
Calculate Tref,sh,out
from Pin, Href,sh,out Calculate Qtp from Calculate Compressor output
Href,in, Href,tp,out, Gref (Gevp,in = Gexp,in = Gcnd,in) = Gcmp,out
Calculate Qsh from
Href,vs, Href,vs,out, Gref Calculate Tair,tp,out from
Qtp, Hair,in, Gair Calculate Condenser output
Calculate Tair,sh,out from Calculate Evaporator output
Qsh, Hair,in, Gair Calculate Expansion output
Calculate Δθtp from Tair,in,
Tair,tp,out, Tref,in, Tref,tp,out
Calculate Enthalpy at (Tref,vs,P0 + 5) : Hset1
Calculate Δθsh from Tair,in,
Calculate Enthalpy at (Tref,ls,P1 – 1) : Hset2
Tair,sh,out, Tref,vs, Tref,sh,out Calculate Qtp’ from
A, Ktp, Δθtp
Calculate Qsh’ from P0 – Pexp,out = ε0, P1 – Pcmp,out = ε1 Adjust
No H0 – Hexp,out = ε2, H1 – Hevp,out = ε3 assumption
A, Ksh, Δθsh Qtp - Qtp’<ε (Use derivative
H2 – Hcmp,out = ε4 information)
No H1 – Hset1 = ε5, H0 – Hset2 = ε6
Qsh - Qsh’<ε
No
ε0~ε6 < ε
Tair,out = Tair,vs,out·Rvs+ Tair,sh,out·(1- Rvs) Tair,out = Tref,tp,out
Hout = Href,sh,out Hout = Href,tp,out Yes
Pout = Pin Pout = Pin Calculate El
Gref,out = Gref,in Gref,out = Gref,in
Output: Tair,out, El, Twat,out
Output: Gref,out, Hout, Pout, Tair,out
Figure 2 Calculation flow chart of the
Figure 1 Calculation flow chart of the evaporator refrigeration cycle

- 597 -
 Proceedings: Building Simulation 2007

and a heating rate at a given ambient temperature or

Bias Linear
under other conditions. State variables and input Function
Bias θk θj
parameters should be changed in accordance with the Sigmoid
intended use. For example, when the flow rate of a x i Function
cooling air is need provided that a heating rate is
given, the heating rate will be the input parameter y j

while the flow rate of the cooling air becomes state

variable. Output
x τ wkj Layer
The result of the simulation will be described later. wik
Input Layer Hidden Layer
FAST CALCULATION OF Figure 3 Architecture of three-layer neural network
REFRIGERANT PROPERTIES
NIST REFPROP database (Anon.) is mostly used to ⎛ σ ⎛ τ ⎞ ⎞
calculate the thermodynamic properties of ⎜
⎝ i =1
∑
y j (x ) = f sig ⎜ wkj ⋅ f sig ⎜ wik xi + θ k ⎟ + θ j ⎟ (11)
⎜ ⎟
⎠
⎟ ∑
refrigerants. Because of unavoidable iterations in ⎝ k =1 ⎠
calculations, the use of REFPROP leads to high 2
computational time delay. f sig (x ) = −1 (12)
1 + e−x
To accelerate the computation, Cleland (1986) and
Charters (1987) has developed some formulae for Sub cooled region Two phase region Super heated region
pure refrigerants approximation. Corberan developed
approximation method by linear interpolation.
Guoliang (2005, 2006) proposed implicit curve-
fitting method which has applicable to mixed Rated refrigeration cycle
Pressure [MPa]

refrigerants. Although they achieved significant

results for certain range on the mollier diagram, none
of them mentioned approximation of density at two- Approximation range
phase state. There was difficulty in approximating
density at a two-phase state because of the
nonlinearity around the saturated liquid line. As
Enthalpy [kJ/kg]
shown in equation (10), the density state is
indispensable to evaluate the pressure drop at the Figure 4 Mollier diagram with approximation range
expansion valve. and rated refrigeration cycle

In this section, an approximation method which uses Sat. liquid line (REFPROP) Sat. liquid line (Approximation)
Sat. vapor line (Approximation) Sat. vapor line (REFPROP)
a neural network is proposed to rapidly calculate 2000
refrigerant thermodynamic properties. A neural 1700 Sub cooled Super heated
network is a non-linear statistical data modeling tool. region region
Pressure [kPa]

1400 Two phase

Cybenko (1989) showed that every continuous region
1100
function can be closely approximated by a multi-
800
layer perceptron with just one hidden layer (which
500
means three-layer neural network). Figure 3 shows
architecture of a three-layer neural networks. An 200
0 500 1000 1500 2000
output value will be calculated by equation (11) and Enthalpy [kJ/kg]
(12). In this study, Levenberg-Marquardt method
(Bogdan 1999) was used to train the neural network. Figure 5 Approximation results of saturated states

Since the curve of the thermodynamic property

Saturated refrigerant properties
inflects around the saturated lines on the mollier
diagram, area is divided into three regions by the For saturated refrigerants, enthalpy, temperature and
saturated lines. Different neural networks should be density are expressed as a function of pressure.
established for different regions. A range with a Value of pressure is directly used as input value of a
pressure from 200 [kPa] to 2000 [kPa] and enthalpy neural network.
from 400 [kJ] to 2000 [kJ] is selected for
approximation since it contains the entire range of Figure 5 shows the approximation result of saturated
refrigeration cycle at the rated point. Figure 4 shows lines of ammonia. In order to check the accuracy of
a mollier diagram of ammonia with the the approximation, 200 points are randomly
approximation range and the rated refrigeration cycle. distributed in the ranges selected above. Table 1
shows the maximum and average value of a relative
error, where the relative error is calculated by

- 598 -
 Proceedings: Building Simulation 2007

equation (13). Table 2 shows a calculation speed of X App − X REF

REFPROP and the neural network. It is shown that Errrel = ×100 (13)
the average relative error of the output from the X REF
neural network is less than 0.01 [%] while the
H − H sl
calculation speed is up to 14 times faster than that of Hr = (14)
REFPROP. H sv − H sl

Sub cooled, two-phase and super heated Figure 6 and 7 show the approximation result of sub
refrigerant properties cooled, two-phase and super heated region. Drastic
increase of density around the saturated liquid line is
For sub cooled, two-phase, and super heated clearly represented by the neural network. In order to
refrigerants, temperature and density are expressed as check the accuracy of the approximation, 40000
a function of enthalpy and pressure. Pressure and points are randomly distributed in the ranges selected
enthalpy are directly used as input value to calculate above. Table 3 shows the maximum and average
temperature and density in the super heated region. value of relative error. Table 4 shows the calculation
Since density increases exponentially around the speed of REFPROP and neural network. It is shown
saturated liquid line, the relative position of enthalpy that the average relative error of output from neural
between saturated lines should be used for an input network is less than 0.5 [%] while the calculation
variable to calculate density at the sub cooled and speed is up to 33 times faster than that of REFPROP.
two-phase region. The relative position of an
enthalpy is calculated by equation (14). Densities at
saturated states are also used for input variables. Table 2 Calculation speed of saturated state

REFPROP vers.7 Approximation

Table 1 Max and average value of relative error 453.13 [msec] 31.25 [msec]

Sat. liquid line Sat. vapor line

9.88 · 10-3 9.63 · 10-5 Table 3 Maximum and average value of relative error
error [%] error [%]

Enthalpy
relative relative
Ave.

Temp. 4.16 · 10-3 6.61 · 10-3

Density 1.29 · 10-4 9.63 · 10-3 Sub cool Two-phase Super heat
Enthalpy 6.02 · 10-2 1.05 · 10-3 Ave. relative Temp. 2.14 · 10-2 2.85 · 10-2 2.42 · 10-3
4.32 · 10-3 4.48 · 10-1 3.09 · 10-3
Max

Temp. 1.24 · 10-2 1.67 · 10-2 error [%] Density

Density 2.63 · 10-4 5.62 · 10-2 Max relative Temp. 2.74 · 10-1 1.29 · 10-1 1.82 · 10-2
error [%] Density 3.91 · 10-1 2.66 · 10-0 5.62 · 10-2

REFPROP vers.7 Approximation Relative error [%]

Figure 6 Approximation result of Temperature

Relative error [%]

REFPROP vers.7 Approximation

Figure 7 Approximation result of Density

- 599 -
 Proceedings: Building Simulation 2007

Table 4 Calculation speed of parents are offspring are

sub cooled, two-phase, super heated region selected randomly generated by
UNDX
REFPROP vers.7 Approximation
135234.38 [msec] 4046.75 [msec]

PARAMETER ADJUSTMENT WITH

GENETIC ALGORITHM
Although theoretical models are versatile, they offspring 1 : selected by fitness
normally need number of physical parameters to offspring 2 : selected randomly
execute a simulation successfully. In this section, a
Figure 9 Alternation of generations by MGG
physical parameters adjustment method using genetic
algorithm is proposed.
Genetic algorithm In this study, a value of overall heat transfer
coefficients, heat transfer areas, polytropic index, and
A genetic algorithm is a random search technique to the volumetric efficiency are the parameters that
find optimal solutions. For continuous quantity, a need to be optimized. In the beginning, these
real-coded genetic algorithm is used. Physical parameters are randomly set within certain bounds.
parameters such as over all heat transfer coefficient, Table 5 shows the minimum and the maximum
surface area, and so on, are treated as DNA. values of the parameters. Fitness of DNAs is
UNDX (unimodal normal distribution crossover) is calculated by equation (15). The value of the fitness
adopted for cross over method (Ono I. 1997). Figure is between 0 and 1 as a function of a total relative
8 shows generation of child DNA by UNDX. Two error. The total relative error is defined by the
normal distributions (parallel and vertical to relative error of the heat load and the electricity
deviation vector between parent1 and parent2) are consumption.
made to generate children randomly.
Table 5 Minimum and Maximum value of parameters

parent 2 Min. Max.

Name of parameter
parent 3 value value
d2 Polytropic index [-] 0.7 0.85
Volumetric coefficient [-] 1.17 1.57
Heat transfer area [m2] 1000 1500
Evaporator Condenser

d1
Overall heat transfer coef [kW/Km2] 40 60
2
parent 1 σ1 Overall heat transfer coef [kW/Km ] 40 60
2
Heat transfer area [m ] 800 1300
σ2
σ1 = αd1 Overall heat transfer coef [kW/Km2] 40 60
σ2 = βd2 / n Overall heat transfer coef [kW/Km ] 2
40 60

Figure 8 Generation of offspring by UNDX

1
Fit = (15)
Figure 9 shows alternation of generations by a MGG Err + 1
(minimum generation gap) method, which has an
n n
advantage in avoiding a local solution. First, three
parents are selected randomly and a certain number
Err = ∑
j =1
Δload rel , j + ∑ Δelec
k =1
rel , k (16)
of offspring are generated by UNDX. Then, the
fitness of the offspring is evaluated and the most
Result of parameter adjustment for rated value
adopted one is selected as the first child. The second
child is randomly selected from the rest. Finally, 5 points of rated value (altered to outdoor dry bulb
parent 1 and parent 2 are replaced by the selected temperature) are used to evaluate the fitness.
two offspring.
200 individuals (DNAs) are initially generated and
6000 generations are passed over to evolve DNAs.
Figure 10 shows the value alternation of heat transfer
area of the evaporator and the condenser. It is seems
that parameters are converged into a certain value

- 600 -
 Proceedings: Building Simulation 2007

with the passage of generation. Figure 11 shows the Initial 1000th generation 2000th generation 6000th generation
1450
alternation of the fitness with the passage of

Heat transfer area of evaporator [m2]

generation. Other parameters are also estimated.
1250
Table 6 shows final value of the each parameter
gained from the fittest DNA.
1050
Figure 12 and Table 7 shows the calculation result of
the refrigeration cycle with estimated parameters 850
against an outdoor dry bulb temperature. The values Solution
at rated points are also plotted. Simulation outputs 650

are in good consistent with the rated value. The value 1000 1200 1400 1600 1800 2000 2200
Heat transfer area of condenser [m2]
of relative error is less than 1.2 [%].
Figure 10 Alternation of the heat transfer area
Both REFPROP and the approximation method Maximum value Average value
described above are used to calculate the 0.25
thermodynamic properties. It seems that there is 0.20
slight difference at lower temperature but the

Fitness [-]
0.15
deviation ratio is not bigger than 3 [%]. Table 8
0.10
shows the calculation speed of the refrigeration cycle
0.05
with the two methods. The rated refrigeration cycles
are calculated 50 times. The calculation which uses 0.00
0 1000 2000 3000 4000 5000 6000
the approximation method is about 20 times faster Generation [-]
than that of using REFPROP.
Figure 11 Alternation of the fitness

Table 6 Final value of the each parameter

REFPROP Approximation Rated performance

Name of parameter Value

Electricity consumption [kW]

54

Polytropic index [-] 1.00507 53 53.4 Electricity consumption [kW]

52
Volumetric coefficient [-] 0.736 52.3
51
Heat transfer area [m2] 1580.95
Evaporator Condenser

51.2
50
50.2
Overall heat transfer coef [kW/K·m2] 71.59 49 49.0
2 48
Overall heat transfer coef [kW/K·m ] 55.16 230

Heat transfer area [m2] 1214.98 220 Thermal load [kW]

Thermal load [kW]

2 210 219.4
Overall heat transfer coef [kW/K·m ] 46.94
200 209.6
2
Overall heat transfer coef [kW/K·m ] 100.51 190 199.9
190.7
180
181.5
170
Table 7 Calculation result of the refrigeration cycle 4.2

with estimated parameters 4.1 COP [-]

4.11
4.0
COP [-]

4.01
Outdoor dry bulb temperature [C] 3.9
3.90
3.8
7 6 5 4 3 3.80
3.7
Electricity Rated 53.4 52.3 51.2 50.2 49.0 3.70
3.6
consumption Ref 53.1 52.3 51.2 50.0 49.4 7 6 5 4 3
[kW] App 53.1 52.3 51.2 50.0 48.7 Outdoor Dry bulb Air Temp. [C]
Rated 219.4 209.6 199.9 190.7 181.5 Figure 12 Calculation result of the refrigeration
Thermal load
Ref 216.7 209.6 200.0 190.8 186.9 cycle with estimated parameters
[kW]
App 216.9 209.7 200.0 190.8 181.5
Rated 4.11 4.01 3.90 3.80 3.70
COP
Ref 4.08 4.01 3.90 3.82 3.79
[-]
App 4.08 4.01 3.91 3.81 3.73
CONCLUSION
In this paper, a parameter adjustment method for an
ammonia heat pump chiller using a genetic algorithm
Table 8 Calculation speed of the refrigeration cycle
(GA) was developed. To speed up adjustment
REFPROP vers.7 Approximation process, a new approximation method with neural
99421.88 [msec] 4968.75 [msec]
network is also proposed.
1) The calculation speed of saturated states and
other refrigerant states increased 14 times and 33

- 601 -
 Proceedings: Building Simulation 2007

times respectively, while the average relative Parameter adjustment with genetic algorithm
error was less than 0.04 [%] and 0.5 [%] when Fit Fitness of DNA [-]
compared to exact solution. Err Total relative error [-]
2) The calculation speed of refrigerant cycle Δload Relative error of thermal load [-]
increased 20 times faster, while maximum Δelec Relative error of electricity consumption [-]
relative error was within 3 [%]. ACKNOWLEDGMENT
3) Parameter adjustment method by genetic
This study was partially funded by the Waseda
algorithm was applied with the help of the
University Grant for Special Research Projects. This
proposed approximation method to speed up the
study is partially supported by Waseda University
iterative calculation process. Moreover, outputs
project research institutes, Comprehensive research
values of the simulation are consistent with
organization “Institute of Architectural Environment
values at rated point as demonstrated using
Design & Information System”. This research was
adjusted parameters values.
partially supported by the Ministry of Education,
Science, Sports and Culture, Grant-in-Aid for
NOMENCLATURE
Scientific Research (A), No.19206063, 2007. The
Physical model of refrigeration cycle authors express gratitude to Professor S. Kawai and
Mr. S. Yamaguchi for valuable advices about
T Temperature [K]
physical model of heat pump chiller.
P Pressure [kPa]
ρ Density [kg/m3] REFERENCES
H Enthalpy [kJ/kg]
G Flow rate [kg/sec] Anonymity http://www.nist.gov/data/nist23.htm
W Compression work [kW] M. Fujiwara, K. Tani, T. Matsunaga and M.
El Electricity consumption [kW] Watanabe, 1985 “Analysis of screw compressor
Q Amount of heat exchanger duty [kW] performance (2nd report, Simulation program)”,
A Heat transfer area [m2] The Japan Society of Mechanical Engineers,
R Rate of area [-] Vol.51 No.466, pp.1816-pp.1824
K Overall heat transfer coefficient [kW/m2-K]
pol Polytropic index [-] Guo Liang Ding, Zhigang Wu, Kaijian Wang and
Masaharu Fukaya, 2005. “An implicit curve-
ηv Volumetric efficiency [-]
fitting method for fast calculation of thermal
Vol Volume [m3]
properties of pure and mixed refrigerants”,
Cv Flow coefficient [m2] International Journal of Refrigeration, pp.921-
Δθ Logarithmic mean temp. difference [K] pp.932
fEOS(·) Equation of state (EOS)
Guo Liang Ding, Zhigang Wu, Kaijian Wang and
Subscripts:
Masaharu Fukaya, 2006. “Extension of the
sh super heat cmp compressor applicable range of the implicit curve-fitting
tp two-phase evp evaporator method for refrigerant thermodynamic properties
sv saturated vapor cnd condenser to critical pressure”, International Journal of
sl saturated liquid exp expansion Refrigeration, pp.1-pp.15
in inlet air outdoor air
out outlet ref refrigerant Corberan J.M., Gonzalvez J., Fuentes D,
wat water set setpoint “Calculation of refrigerant properties by linear
interpolation of bidimensional meshes”,
First calculation of refrigerant properties http://www.imst-
fsig(·) Sigmoid function art.com/ficherosdesc/interpolation.pdf
τ Number of neurons in input layer Cybenko G., 1989, “Approximation by superposition
σ Number of neurons in hidden layer of a sigmoidal function”, Math. Control Signals
j Number of neurons in output layer Systems, 2, 303-314
θ Threshold value
Bogdan M., Wilamowski, Yixin Chen, 1999,
w Connection weight between neurons
“Efficient Algorithm for Training Neural
x Input value
Networks with one Hidden Layer”, Proc. of the
y Output value International Joint Conference on Neural
Errrel Relative error Networks (IJCNN'99), Washington, D.C., USA,
Hr Relative value of enthalpy July 10-16, pp 1725-1728.
Subscripts:
Ono I. and Kobayashi S., 1997, “A Real-coded
APP value calculated from approximation formula Genetic Algorithm for Function Optimization
REF value calculated from REFPROP Using Unimodal Normal Distribution
Crossover”, Proceedings of the seventh

- 602 -
 Proceedings: Building Simulation 2007

International Conference on Genetic Algoritms,

pp.246-253

- 603 -

View publication stats