Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Annals of Nuclear Energy 29 (2002) 805819

www.elsevier.com/locate/anucene

Improved methodology for generation of axial


ux shapes in digital core protection systems
Gyu-Cheon Leea,*, Won-Pil Baekb, Soon Heung Changc
a
Korea Power Engineering Company, Inc.,150 Dukjin, Yuseong, Daejeon, South Korea
b
Korea Atomic Energy Research Institute, 150 Dukjin, Yuseong, Daejeon, South Korea
c
Korea Advanced Institute of Science and Technology, 373-1 Gusung, Yuseong, Daejeon, South Korea

Received 17 April 2001; received in revised form 22 June 2001; accepted 22 June 2001

Abstract
An improved method of axial ux shape (AFS) generation for digital core protection sys-
tems of pressurized water reactors is presented in this paper using an articial neural network
(ANN) techniquea feedforward network trained by backpropagation. It generates 20-node
axial power shapes based on the information from three ex-core detectors. In developing the
method, a total of 7173 axial ux shapes are generated from ROCS code simulation for training
and testing of the ANN. The ANN trained 200 data predicts the remaining data with the average
root mean square error of about 3%. The developed method is also tested with the real plant data
measured during normal operation of Yonggwang Unit 4. The RMS errors in the range of
0.9 2.1% are about twice as accurate as the cubic spline approximation method currently used
in the plant. The developed method would contribute to solve the drawback of the current
method as it shows reasonable accuracy over wide range of core conditions. # 2002 Elsevier
Science Ltd. All rights reserved.

1. Introduction

A nuclear power plant should be operated by maintaining a sucient margin from


acceptable fuel design limits in terms of departure from nucleate boiling ratio and
local power density; this is usually achieved by plant monitoring systems and pro-
tection systems.

* Corresponding author. Tel.: +82-42-868-8750; fax: +82-42-861-1485.


E-mail address: gclee@ns.kopec.co.kr (G.-C. Lee).

0306-4549/02/$ - see front matter # 2002 Elsevier Science Ltd. All rights reserved.
PII: S0306-4549(01)00076-7
806 G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819

The Korean Standard Nuclear Power Plants (KSNPs) utilize a digital core pro-
tection system called the core protection calculator (CPC) system and a monitoring
system called the core operating limit supervisory system (COLSS). Both systems
have the capability of on-line real-time calculation for fuel limiting parameters (i.e.
nucleate boiling ratio and local power density) using measurable plant parameters,
resulting in the enhancement of operational exibility. They use several input para-
meters, such as core inlet temperature, system pressure, core ow rate, heat ux, radial
and axial ux shapes, where the axial ux shape (AFS) is obtained by applying tting
methods to available power detector signals. KSNPs utilize two types of power detec-
tors: in-core detectors for the COLSS and ex-core detectors for the CPC. The major
advantage of ex-core detectors is the fast response required for protection purposes;
however, they are located only at three axial positions (bottom, middle and top) and
at each radial quadrant of the core. One of the topics in the CPC is, therefore, how
to precisely construct the AFS with the information of only three ex-core signals.
The ex-core detector signals are aected by many parameters such as core power,
control rods position, xenon condition, and core burnup, etc. Since the relationship
between those parameters and detector signals are so complex, it is impractical to
analytically model the nonlinear relationship among them. The current CPC utilizes
the least square tting and cubic spline curve tting methods to construct the 20-
node AFS. The actual plant data measured during start-up tests (2080% of the
rated power) are used to determine the correlation coecients.
The AFS generated from the CPC is relatively accurate at beginning of cycle, but the
error tends to increase as core burnup increases due to the change of core conditions
from start-up test conditions. In addition, the measured data during startup from 20 to
80% of the rated power does not include all kinds of AFS that may occur during
operational transients or accident conditions. A larger uncertainty in the AFS predic-
tion will force to put a larger penalty into the CPC, resulting in the reduction of the
thermal margin. This leads to a need for more ecient methods in generating the AFS.
The articial neural networks (ANNs) are ecient and reliable algorithms capable of
performing functional input/output mappings (Ricotti and Zeo, 1999). Their parallel,
multi-parametric character, and their computing speed, allows them to assert themselves
as a powerful computational tool in many elds of research and application (Moon and
Chang, 1994; Kim and Chang, 1997; Garis et al., 1998), especially when the underlying
physical or mathematical model is complicated. Multilayer feedforward networks
generally have hidden layers between input and output layers. Each layer consists of
computational units. In the networks, activation functions for the hidden units play a
role to introduce nonlinearity into the network. Without nonlinearity, hidden units
would not make nets more powerful than just plain perceptrons, which do not have any
hidden units so that a composition of linear functions is again a linear function. It has
been proven that multilayer feedforward networks are capable of approximating any
functional continuous mapping from one nite dimensional space to another, provided
the number of hidden units is suciently large (Funahashi, 1989; Hornik et al., 1990).
This paper presents a new methodology for generation of the AFS by taking
advantage of the ANN algorithm, which can produce a reasonable accuracy at any
core conditions.
G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819 807

2. Training of the ANN for axial ux shapes

Among many kinds of networks, a feedforward network trained by backpropagation


proposed by Rumellhart (1986) is the most widely used supervised training method
which is a steepest descent method of computing the interconnection weights that
minimize the total squared error between the actual network output and the true value.
Because the algorithm is introduced in many papers (e.g. Sarle, 1997), its mathematical
formulations are not repeated in this paper.
The number of input and output nodes is determined by the characteristics of the
problem. Three input nodes are selected since the input data are three ex-core
detector signals in this problem, as illustrated in Fig. 1. The number of output nodes
is determined as 20 since CPC produces the same number of axial power relative to
the average, which would give an equal basis for comparison. The number of hidden
layers and their nodes are determined by trial-and-error method to get an optimum
value. The activation function used in this study is a tangent hyperbolic function,
which is a kind of dierentiable sigmoid function.
Reliable performance of the ANN can be achieved by reecting a wide range of
AFSs and the corresponding detector signals that may appear during normal oper-
ating conditions and accident conditions. Since it is dicult to get such a wide range
data through testing at a real plant, nuclear design codes are used to obtain the
required data. The ROCS code (Loretz et al., 1989) is a two-group diusion theory
code which solves the two- or three-dimensional diusion equations in Cartesian

Fig. 1. Structure of a multi-layered feed forward neural network for axial ux shape generation.
808 G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819

geometry by applying the higher order coarse-mesh nodal expansion method. It mod-
els control rods insertion and feedback eects of moderator density and temperature,
fuel temperature, and xenon and samarium poisons. Based on the core physics, it pro-
vides the AFS and the corresponding ex-core detector signals. Detailed mathematical
models, solving algorithms, and the calculational and experimental verication of
ROCS code are well described by Ober et al. (1977). Using the ROCS code, total 3606
AFS sets for YGN 4 (an operating KSNP) cycle 4 are generated by varying core power,
control rods position, xenon condition and core burnup. A number of training sets are
extracted from the 3606 sets, by applying systematic sampling methods instead of
random sampling to include typical shape patterns. Four sampling methods are
considered for choosing the training sets from the total sets in accordance with:
. the order of top-to-bottom (T/B) detector signal ratio (sampling at even data
interval)
. the order of the bottom detector signal (sampling at even data interval)
. classication of AFS patterns (sampling the same number for 15 patterns)
. the order of axial shape index (ASI) (sampling at even data interval)
To verify the propriety of the ANN, various training cases with varying number of
sample sets, hidden layers, and nodes for the hidden layers are tested as shown in the
next section.

3. Test results of the ANN

The trained ANNs are tested to the data sets which were not used in the training
out of total 3606 sets. The RMS error between the target AFS and the ANN output
for each test set is represented as follows:
v
!
u REF 2
u 1 NX node
Pj ANN
 Pj
RMS error t 1
Nnode j1 PREF
j

where

Nnode =number of axial nodes,


PREF
j =target power from ROCS output at j th node,
PANN
j =generated power from the ANN at j th node.

To evaluate the ANN performance, the average and the maximum RMS errors of
the total test sets are introduced as follows:
v
!
u
1 NXdata u X
N node
P ANN
 P REF
Average RMS error t 1 i;j i;j
2
Ndata i1 Nnode j1 PREF
i;j
G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819 809

v
0 ! 1
u REF 2
X
NX
Ndata u ANN
t 1 node
P i;j  P i;j
Maximum RMS error Max@ REF
A 3
i1
N node j1
P i;j

where
Ndata =number of total test sets,
PREF
i;j =target power from ROCS output at jth node of ith test set,
PANN
i;j =generated power from the ANN at jth node of ith test set.

Table 1 summarizes the test results of various sampling methods with the combi-
nation of hidden layers and its node numbers. The total test sets of ANN trained
with 100, 150 and 200 samplings are 3506, 3456 and 3406, respectively. The test
results show that the performance of the ANNs does not signicantly vary with the
sampling methods. Regardless of various sampling methods, the number of hidden
layers and their nodes, all the training cases show that the average RMS error is
about 3%. In addition, about 85% of the test sets are within 5% RMS error and at
least 97% of the test sets are within 10% RMS error for all the trained cases. Fig. 2
shows cumulative number of test sets according to RMS errors for the training case

Table 1
Test results of various training cases

Training cases Results

Sampling Sample No. of No. of Average Max 5% 10% 15% 20%


method sets for hidden node RMS RMS RMS RMS RMS RMS
training laver error error error error error error
(%) (%) (%) (%)

T/B detector order 100 1 15 0.0312 0.2104 86.04 99.20 99.63 99.94
25 0.0326 0.1659 85.31 99.06 99.88 100.00
2 15, 15 0.0328 0.2235 85.63 98.34 99.52 99.89
25, 25 0.0333 0.1364 83.79 98.89 100.00 100.00
150 1 15 0.0340 0.2467 84.01 97.73 99.34 99.97
25 0.0308 0.1857 87.64 99.31 99.94 100.00
2 15, 15 0.0327 0.1339 84.85 99.47 100.00 100.00
25, 25 0.0311 0.1918 87.79 99.39 99.82 100.00
200 1 15 0.0296 0.1019 88.31 99.91 100.00 100.00
25 0.0299 0.1041 87.42 99.93 100.00 100.00
2 15, 15 0.0304 0.1280 84.41 99.41 100.00 100.00
25, 25 0.0296 0.1017 88.15 99.94 100.00 100.00
ASI order 200 1 15 0.0317 0.1138 86.58 99.53 100.00 100.00
2 15,15 0.0311 0.1185 86.73 99.35 100.00 100.00
Bottom detector 200 1 15 0.0302 0.0932 88.15 100.00 100.00 100.00
2 15, 15 0.0307 0.1508 87.76 99.49 99.97 100.00
Shape pattern 200 1 15 0.0309 0.1233 87.76 99.29 100.00 100.00
2 25, 25 0.0318 0.1291 88.64 99.71 100.00 100.00
810 G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819

Fig. 2. Test results for YGN 4 cycle 4.

based on the order of bottom detector signals with 200 sample-training sets and one
hidden layer that reveals the smallest maximum RMS error among the training
cases. The average and the maximum RMS errors are 3.02 and 9.32%, respectively.
Moreover, 95% of the test sets are within about 6% RMS errors.
To the training case mentioned in Fig. 2, another test has been performed to conrm
whether the ANN can show equal performances for another cycle. A total 3567 test sets
for YGN 4 cycle 3 were generated by using ROCS code. As shown in Fig. 3, the average
and the maximum RMS errors for YGN 4 cycle 3 are 2.85 and 12.6%, respectively.
Similar to the YGN 4 cycle 4 results, the average RMS error is about 3% and about
99% of the test sets are below 8% RMS error.

Fig. 3. Test results for YGN 4 cycle 3.


G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819 811

From the above evaluation, it has been clearly conrmed that the AFS generation
method using neural network is feasible, and furthermore, it has a good capability of
AFS generation that has nearly no dependency on the core burnup.

4. Comparison with the current method in CPC

There is a recommendation that the axial shape related coecients should be re-
evaluated if the CPC RMS error is greater than 8% during the operation. This indi-
cates that the current CPC is designed conservatively enough to allow such value of
RMS error. Then, the results of the ANN indicate that about 6% RMS error is accep-
table with 95% of probability even though we consider the extreme cases such as
severely top- or bottom-skewed AFSs, which may occur at abnormal operations or
accident conditions. Therefore, it is considered that the performance of the ANN is
fairly excellent compared with the method used in the current CPC.
To compare the performance of the ANN with the current CPC, 15 AFS patterns are
arbitrarily chosen from the YGN 4 cycle 4 test sets. Table 2 shows that the core condi-
tions of the chosen sets. They include various core conditions in combination with the
control rod position, the core power, and the core burnup. The axial shape index (ASI)
is dened as the ratio of the lower-half core power minus upper-half core power to the
total power, which gives crude information of the axial ux shape. For example, if the
ASI is 0.863, the axial shape is badly bottom-skewed as shown in Fig. 4. With the ASIs
shown in Table 2, it is considered that the sets are properly chosen from the extremely
top-skewed to the bottom-skewed shapes without bias. With the core conditions, the

Table 2
Test cases used for comparison of the AFS from ANN and CPC

Case No. Burnup ASI Power (%) Control rods (% withdrawal)

Group 4 Group 5 PSR 1 PSR 2

1 EOC 0.863 100 27 0 70 70


2 BOC 0.385 100 100 0 35 35
3 BOC 0.370 100 73 0 100 100
4 EOC 0.335 80 50 0 100 100
5 MOC 0.166 60 100 73 70 70
6 BOC 0.187 100 100 50 50 50
7 EOC 0.089 60 100 50 10 10
8 MOC 0.033 80 100 0 100 100
9 BOC 0.077 20 73 0 27 27
10 EOC 0.156 40 100 27 27 27
11 EOC 0.487 100 100 100 100 100
12 BOC 0.048 40 100 27 73 73
13 EOC 0.649 20 100 0 100 100
14 EOC 0.609 100 100 100 100 100
15 EOC 0.550 100 100 27 35 35
812 G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819

Table 3
Eect of the RMS error on the MDNBR

Case ASI CPC ANN Target CPC ANN


RMS error RMS error MDNBR MDNBR MDNBR

1 0.863 0.5115 0.0286 1.2897 1.6159 1.2883


2 0.385 0.3351 0.0122 2.4579 2.1374 2.4960
3 0.370 0.3154 0.0202 2.5380 2.2063 2.24990
4 0.335 0.3028 0.0447 0.5067 2.1985 2.3801
5 0.166 0.1338 0.0285 3.1531 2.8072 3.1667
6 0.187 0.1359 0.047 2.9982 2.7886 3.1124
7 0.089 0.1137 0.0131 3.1657 3.0894 3.1723
8 0.033 0.0709 0.0079 2.0977 3.2023 2.9222
9 0.077 0.0499 0.0325 2.8457 2.9173 2.9182
10 0.156 0.0682 0.0211 2.3072 2.4859 2.3219
11 0.487 0.1381 0.0167 1.9836 1.8461 1.9885
12 0.048 0.0676 0.0248 2.8871 3.1480 2.8566
13 0.649 0.3931 0.0093 1.4842 1.4292 1.4922
14 0.609 0.4380 0.0286 1.8007 1.6649 1.8056
15 0.550 0.2364 0.0196 1.9128 1.7746 1.9191

corresponding AFSs of the CPC are obtained by simulation.1 Fig. 4 shows the com-
parison of the axial ux shapes generated from the ANN and those from the CPC for
some cases. Solid lines represent the target (true) values come out from the ROCS code,
and dot lines and dash-dot-dot lines represent the AFS resulted from the CPC and the
ANN, respectively. As shown in Fig. 4, the performance of the ANN is evidently better
than that of the current method used in CPC for every axial ux shapes pattern tested,
especially when the AFSs are badly skewed or saddle types.
To see the eect of AFSs uncertainty on the minimum departure from nucleate boil-
ing ratio (MDNBR), a test run was performed for the AFSs generated from the CPC
and the ANN with the 15 cases in Table 2. The remaining parameters required to cal-
culate the DNBR are assumed as constant values of core inlet temperature (569 K),
primary system pressure (15.51 MPa), core mass ow rate (1.53104 kg/s), and radial
peaking factor (1.29). It should be noted that the calculated MDNBR values in Table 3
are not for calculating the real situations but just for evaluating the AFSs eect on
MDNBR. The target MDNBR in Table 3 means the MDNBR obtained with the AFS
from ROCS code, and similarly, the CPC MDNBR and the ANN MDNBR do the
MDNBR obtained with the AFS generated from the CPC and from the ANN,
respectively. The larger RMS error cases obviously result in big MDNBR dierences
from the target. It is noted that the CPC MDNBR dierences from the target for the
cases 9 (0.077ASI) and 10 (0.156ASI) are relatively small compared with the
others. This is because the shape annealing matrix and boundary point power cor-
relation coecients used in the CPC for tting AFS are obtained from the data

1
The CPC simulation is actually performed by using CPC FORTRAN that was made to perform a lot of
simulations required in the design. It has basically the same algorithum with the real CPC installed in the
plant. The consistency of the CPC FORTRAN with the CPC has been veried according to the procedure.
G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819 813

Fig. 4. Comparison of AFSs generated from current CPC and ANN.


814 G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819

measured in the range between about 0.2 and 0.05 ASI, as mentioned in Section
1. The current CPC has a penalty to compensate these uncertainties, resulting in a
reduction of operation exibility. Better prediction of AFSs by the ANN would give
benecial eects on the MDNBR, thereby on the thermal margin.

5. Applicability of the ANN method

The acquired measured data during the startup and normal operation of YGN 4
cycle 4 are used for verifying the applicability of the AFS generation method devel-
oped in this study.
In accordance with the current power ascension program of the plant, the calibration
of ex-core detector signals was performed at about 20% power. And then, acquired
were the data needed for tting the shape annealing matrix and the boundary point
power correlation coecients, with increasing the core power to about 80%. Figs. 5 and
6 illustrate measured ex-core detector signal errors of a CPC channel according to the
reactor power and the burnup, respectively. The detector signal error is dened as
follows:

DM C
i  Di
Ei  100 4
DC
i

where
i =top, middle or bottom ex-core detector,
Ei =top, middle or bottom ex-core detector signal error,
DMi =measured ex-core detector signal,
DCi =CECOR synthesized ex-core detector signal.

Fig. 5. Variation of detector error to reactor power (Y4C4).


G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819 815

Fig. 6. Variation of detector error to burnup (Y4C4).

In the plant site, the results of CECOR code (Terney et al., 1975, 1976; Bier et
al., 1980) are used as the references for estimating the core power distributions. The
CECOR code is used not for an on-line monitoring but for an o-line evaluation of
the core power distribution with the measured data, as required or as necessary. Instead
of directly solving detailed, multi-group diusion theory calculations, the code synthe-
sizes signals from self-powered, xed in-core detectors into detailed three-dimensional
assembly and peak pin power distributions using libraries of pre-calculated coecients
generated by transport and standard diusion theory methods such as ROCS code.
Using the power in the assembly at each detector level, the detailed axial shape can be
synthesized by using a Fourier expansion. To avoid tediousness, detailed algorithms
are omitted in this paper. The axial tting technique is described in detail by Terney
et al. (1975). Since the results of AFS from ROCS and CECOR show excellent
agreements (Ober et al., 1977), it is believed that the results from CECOR code are
the same as those from ROCS code used in the ANN training.
Fig. 5 shows that the absolute values of top and bottom detector signal errors
increase almost linearly with core power while the middle detector signal error is
maintained very small. This is because the ex-core detectors are calibrated just once at
about 20% power and further calibrations are not performed until the end of cycle.
Fig. 6 shows that the absolute values of ex-core detector signal errors are about 5.5%
throughout the normal operations ( 100% power) and are almost independent on
core burnup. Since the ANN has been trained with the exact values, if somewhat dif-
ferent signal inputs should come into the CPC with the ANN, the constructed AFS
would go wide from the target. On that account, it is required that a compensating
logic should be inserted between the raw signals and the ANN to get proper AFSs.
For YGN 4 cycle 4, a simple compensating method has been applied, as an example,
based on the characteristic that the detector signal errors are linearly dependent on
the core power while almost independent on the burnup.
816 G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819

From linear ttings with the data for each detector and Eq. (4), the compensated
signal can be approximated as follows:

DM
DCo
i
i
5
1 Ei =100

where

DCo
i compensated signal;
Ei 0:534  0:05  P for bottom detector;
0:031  0:0056  P for middle detector; 6
0:032 0:0545  P for top detector;
and P percent of rated core power:

By applying the compensation to the developed method, AFSs are generated and
compared with the CECOR outputs, target values, at 12 burnup points during nor-
mal operation of the YGN 4 cycle 4.
Fig. 7 shows that the developed method gives RMS errors in the range of 0.92.1%
and almost independent of core burnup, while the current CPC gives 2.04.4%.
Hereby, we conclude that the performance of ANN is at least two times superior to the
current CPC during the normal operation. Since the variation of AFS is not severe
during normal operations, and the shape annealing matrix and the boundary point
power correlation coecients used in the current CPC are relatively well tuned in

Fig. 7. Comparison of RMS errors resulted from new method and COLSS/CPC for YGN 4 cycle 4.
G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819 817

this region, the dierences between them are not so big relative to the results shown
in previous section. In addition, except for beginning of cycle, the developed method
are superior to the Core Operating Limit Supervisory System (COLSS) (Pasquenza
et al., 1975) which utilizes measured signals from in-core detectors and the Fourier
series synthesis method. The larger RMS errors of the COLSS at end of cycle are
due to the fact that the accuracy of the Fourier series synthesis method depends on
the power shape.
By comparing with the measured data acquired in the site, the newly developed
method is applicable with a compensating logic. As a summary, Fig. 8 illustrates the
general procedure for application of the developed method to the digital core pro-
tection system. First, a set of target axial ux shapes and ex-core detector signals
must be prepared with a proper neural network. The set must include sucient axial
ux shapes that may appear during normal operations and accidents. In this study,
it was generated by ROCS code simulation with the variation of core power, control

Fig. 8. Application procedure of the developed method.


818 G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819

rod position, xenon condition and core burnup. Second, with a suitable sampling
method a number of samples are extracted from the total set for training. Third, the
number of hidden layers, and nodes for the hidden layers is to be found for optimization
of the neural network.
Then, the performance of the neural network is tested with the untrained sets whether
it is appropriate for use. If it should not show acceptable results, another type of neural
networks or sampling methods could be considered. Finally, to avoid unexpected AFS
construction results, an adjusting procedure should be considered to make sure that the
measuring detector signals are compatible with those were used in the training for the
same core conditions. After performing these procedures, the ANN method can be
easily applied to the CPC with simple algorithm changes.

6. Conclusions

A reasonably accurate and robust method for generation of axial ux shapes using
ex-core detector signals of pressurized water reactors has been based on a feedforward
neural network trained by backpropagation. The performance of the ANN method is
demonstrated by testing a number of untrained axial ux shapes that covers wide ran-
ges of core burnup, core power, xenon condition, and control rod position. The average
RMS errors are 3.02 and 2.85% for the YGN 4 Cycle 4 (total 3406 test sets) and YGN 4
Cycle 3 (total 3567 test sets), respectively. About 95% of the total test sets are within the
6% RMS error range. The developed method also predicts well the real plant data
measured during normal operation of Yonggwang Unit 4, showing RMS errors
about 2 times superior to the cubic spline approximation method currently used in
the plant.
Therefore, it is concluded that the proposed method is signicantly superior to the
current method of the CPC in several aspects, including the smaller average RMS
error and independence on core burnup. The method can be easily incorporated into
the on-line digital protection system algorithm by inserting the resultant weighting
matrix and a compensating logic for ex-core detector signals. This would result in
the reduction of uncertainty in DNBR calculation, thereby the increase of the
available thermal margin.
The ANN technique may also be applied to the estimation of DNBR as the future
work. In this case, the current complex calculation process of the CPC could be
replaced by a combination of two ANNs.

References

Bier, J.L., Terney, W.B., Zimnoch, F.S., 1980. NPSD-103-P. Combustion Engineering, Inc.
Funahashi, K., 1989. Neural Networks 2, 183192.
Garis, N.S., Pazsit, I., Sandberg, U., Andersson, T., 1998. Nuclear Technology 123, 278295.
Hornik, K., Stinchcombe, M., White, H., 1990. Neural Networks 3, 551560.
Kim, H.C., Chang, S.H., 1997. Ann. Nucl. Energy 24, 14371446.
Loretz, R. et al., 1996. CE-CES-4 Rev. 11-P. Combustion Engineering, Inc.
G.-C. Lee et al. / Annals of Nuclear Energy 29 (2002) 805819 819

Moon, S.K., Chang, S.H., 1994. Nuc. Eng. Design 150, 151161.
Ober, T.G., Stork, G.H., Rickard, I.C., Gasper, J.K., 1977. Nuc. Sci. Eng. 64, 605623.
Pasquenza, J.P., Church, J.F., Foster, R.G., Humphries, J.R., Kadak, A.C., 1975. CENPD-169-P. Com-
bustion Engineering, Inc.
Ricotti, M.E., Zio, E., 1999. Rel. Eng. Sys. Safety 64, 5971.
Rumellhart, D.E., 1986. Parallel Distributed Processing, Vol. 1. MIT Press, Cambridge MA.
Sarle, W.S., ed., 1997. Neural Network FAQ, (ftp://ftp.sas.com/pub/neural/ FAQ.html).
Terney, W.B., Marks, G.H., Williamson Jr., E.A., Ober, T.G., 1975. Trans. Am. Nucl. Soc. 22, 682683.
Terney, W.B., Williamson Jr., E.A., Ober, T.G., 1976. Trans. Am. Nucl. Soc. 24, 429430.

You might also like