Journal of the Korean Nuclear Society

Volume 36, Number 5, pp. 420 429, October, 2004

Estimation of the Nuclear Power Peaking Factor

Using In-core Sensor Signals

Man Gyun Na, Dong Won Jung, and Sun Ho Shin

Chosun University
375 Seosuk-dong, Dong-gu, Gwangju 501-759, Korea
[email protected]

Kibog Lee
Korea Atomic Energy Research Institute
150 Deokjin-dong, Yuseong-gu, Daejeon, Korea, 305-353
[email protected]

Yoon Joon Lee

Cheju National University
1 Ara-il-dong, Jeju-do, 690-756, Korea
[email protected]

(Received April 9, 2004)

Abstract

The local power density should be estimated accurately to prevent fuel rod melting. The local
power density at the hottest part of a hot fuel rod, which is described by the power peaking
factor, is more important information than the local power density at any other position in a
reactor core. Therefore, in this work, the power peaking factor, which is defined as the highest
local power density to the average power density in a reactor core, is estimated by fuzzy neural
networks using numerous measured signals of the reactor coolant system. The fuzzy neural
networks are trained using a training data set and are verified with another test data set. They
are then applied to the first fuel cycle of Yonggwang nuclear power plant unit 3. The
estimation accuracy of the power peaking factor is 0.45% based on the relative 2 error by
using the fuzzy neural networks without the in-core neutron flux sensors signals input. A value
of 0.23% is obtained with the in-core neutron flux sensors signals, which is sufficiently accurate
for use in local power density monitoring.
Key Words COLSS, fuzzy neural network, in-core flux sensor, local power density, power
peaking factor

1. Introduction distribution monitoring in operating power

reactors is a prerequisite to ensure that various
Detailed 3-Dimensional (3D) core power safety limits imposed on the fuel pellets and fuel

420
Estimation of the Nuclear Power Peaking Factor M. G. Na, et al 421

clad barriers such as the local power density (LPD) especially for an unsymmetrical axial power
and the departure from nucleate boiling ratio distribution. The inaccuracy arises because they
(DNBR) are not violated during reactor operation. are unable to take into account the core operation
The ratio of the expected DNB heat flux to the history and transient situations caused by operator
actual fuel rod heat flux at a particular time during action such as control rod insertion/withdrawal
an incident is called the DNBR. Most commercial and boration/dilution or xenon transient.
power reactors have some types of fixed or Regarding the protection and monitoring systems
movable in-core detectors and ex-core detectors. of the Korea Standard Nuclear Power Plant
These are equipped with an on-line or off-line core (KSNP), the calculation of LPD and DNBR
power or flux distribution monitoring program to constitutes two major functions of Core Protection
estimate the 3D power distribution by combined Calculator (CPC) and COLSS, which each play a
use of the detector signals and pre-calculated role in protection and monitoring systems. COLSS
monitoring constants supplied at the core design monitors the operating limits of a reactor core
stage. For example, Yonggwang PWR nuclear including LPD and DNBR and provides related
power plant unit 3 (YGN-3) [1] has self-powered information to operators. COLSS is a program
rhodium fixed in-core neutron detectors installed that runs in the Plant Monitoring System (PMS)
at 45 fuel assembly (FA) locations on five axial computer, which helps plant operators to monitor
levels. The CECOR code [2] and Core Operation the Limiting Conditions for Operation (LCOs)
Limit Supervisory System (COLSS) [3] of specified in the technical specifications. However,
Combustion Engineering (CE) convert the rhodium COLSS carries out only a monitoring function
detector signals to detector box powers using pre- related to the operating limit of a core and does
determined constants. They then determine the not provide nuclear reactor protection functions.
uninstrumented FA powers using pre-calculated On the other hand, CPC, which provides nuclear
coupling coefficients (CC) defined as the inverse reactor protection functions, calculates faster than
ratio of the power of a given FA to the average COLSS but generates more conservative values.
power of the four surrounding FAs at each Therefore, CPC provides lower DNBR and higher
detector level. The detailed FA axial power LPD values than COLSS. COLSS periodically
distribution is also determined by fitting the five adjusts CPC based on operating variables that are
detector box powers along each FA by a five- accurately calculated by COLSS, including power
mode Fourier series. level, reactor coolant system flow, etc.
The CANDU-type Wolsung nuclear power plant LPD should be estimated accurately to prevent
unit 1[4] has fixed in-core vanadium detectors fuel rods from melting. LPD at the hottest part of
installed at 102 core locations. The CANDU on- a hot fuel rod, which can be explained by the
line flux mapping system [5] converts the 102 power peaking factor (Fq), is more important than
vanadium detector signals to thermal fluxes at the the local power density at any other position in a
detector locations and then maps out the 3D flux reactor core. DNBR studies have been extensively
distribution by least-squares fitting of the measured performed [6-12]. Meanwhile, very little LPD
thermal fluxes to a linear expansion of pre- research has been conducted using artificial
calculated flux modes. These methods using the intelligence methods that have been extensively
pre-calculated coupling coefficients or weighting used in a variety of engineering problems.
constants run fast but they are inaccurate, Therefore, the objective of this work is to predict
422 J. Korean Nuclear Society, Volume 36, No. 5, October 2004

the power peaking factor in a reactor core using as follows:

the measured signals (in particular, including in-
core neutron sensor signals) of the reactor coolant (1)
system by applying fuzzy neural networks
according to the operating conditions. Neural where xj is the input variables to the fuzzy neural
networks have been extensively and successfully network (j=1, 2, ..., m; m = number of input
applied to a variety of engineering problems. The variables), Aij the membership functions for the
fuzzy neural networks should be optimized to antecedent of the i-th rule and j-th input (i = 1, 2,
achieve good monitoring performance of the local ..., n; n = the number of rules), and y i the output
power density. of the i-th rule.
The output and input data employed are the In Eq. (1), the if part is fuzzy linguistic, while the
power peaking factor value in the reactor core and then part is crisp. Usually is a
numerous operating conditions, which are polynomial in the input variables, but it can be any
characterized by reactor power, core inlet function as long as it can appropriately describe
temperature, pressurizer pressure, coolant flowrate the output of the fuzzy inference system within the
of a reactor core, axial shape index, in-core fuzzy region specified by the antecedent of the
neutron sensor signals, and a variety of control rod rule. In this work, a symmetric Gaussian
positions. The F q value in the reactor core is membership function is used. The output of an
i
predicted by the developed fuzzy neural networks arbitrary i-th rule, f , consists of the first-order
using these various operating condition data as the polynomial of inputs, as given in Eq. (2).
inputs to the fuzzy neural networks. The proposed
power peaking factor estimation algorithm is (2)
verified by using the nuclear and thermal data
acquired from numerical simulations of YGN-3. where

2. Fuzzy Neural Networks qij= the weighting value of the j-th input on the i-th
rule output,
In this work, neural networks, which are most ri = the bias of the i-th output.
popular for function approximation, are combined
with fuzzy logic to predict the power peaking The output of a fuzzy inference system with n
factor for various operating conditions. A system rules is a weighted sum of all the fuzzy rule
that consists of a fuzzy inference system outputs. The estimated signal from the fuzzy
implemented in the framework of a neural inference system is given by:
network is generally called an adaptive network-
based fuzzy inference system (ANFIS) or a fuzzy (3)
neural network [13]. The training of the fuzzy
neural network is accomplished by a hybrid where
method combined with a back-propagation
algorithm and a least-squares algorithm. Also, a
first-order Sugeno-Takagi type [14] fuzzy inference
system is used where the i-th rule can be described
Estimation of the Nuclear Power Peaking Factor M. G. Na, et al 423

where

The back-propagation algorithm is a general

method for recursively training fuzzy neural
networks. It uses a gradient descent method. The
gradient descent method tunes the antecedent
y is the output data vector, q is the parameter
parameters (the center position and sharpness of
vector, and the matrix W consists of the input data
membership functions) so that the predefined
and the membership function values. The equation
objective functions E is minimized. In order to
for minimizing the cost function is as follows:
train an antecedent parameter aij, the following
iterative calculation is used:
y = Wq . (6)
(4)
A series of the output of the fuzzy neural network
where , i = 1, 2,..., n, j = 1, 2, is represented by the N (m+1)n-dimensional
matrix W and the (m+1)n - dimensional parameter
..., m, t=0,1,2,... and a is a learning rate for a
vector q. The parameter vector q in Eq. (6) is solved
parameter a. The gradient descent method is very
by using the pseudo-inverse of the matrix W.
stable when the learning rate is small but
susceptible to local minimum.
3. Application to Power Peaking Factor
If we fix the antecedent parameters of the fuzzy
Estimation
inference system by the back-propagation
algorithm, the resulting fuzzy neural network is
The proposed algorithm was applied to the first
equivalent to a series of expansions of some basis
fuel cycle of the Yonggwang unit 3 PWR plant.
functions. This basis function expansion is linear in
The used data were obtained by running the
its adjustable parameters, as shown in Eq. (3);
MASTER [15] code based on some assumptions.
y=wTq, since wT has been determined by the back-
The MASTER (Multipurpose Analyzer for Static
propagation algorithm. Therefore, we can use the
and Transient Effects of Reactor) reactor analysis
least-squares method to determine the remaining
code developed by KAERI (Korea Atomic Energy
parameters (consequent parameters qij and ri ). If a
Research Institute) is a nuclear analysis and design
total number of N input-output training data are
code that can simulate the PWR and BWR core in
given, from Eq. (3) the consequent parameters are
1-, 2-, and 3-dimensional geometry. The
chosen to minimize the following cost function:
MASTER code was designed to have a variety of
capabilities such as static core design, transient
(5)
core analysis, and operation support.
424 J. Korean Nuclear Society, Volume 36, No. 5, October 2004

Fig. 1. Fixed Rhodium In-core Detector Location of YGN-3

The data obtained from simulations of the indicated as instrument numbers 16, 20, 23, and
MASTER code comprise a total of 21875 input- 26 at 3 axial levels in Fig. 1).
output data pairs(x1, x2, x23 , yr . The data The ranges of the input and output signals used
are divided into both training data sets and test for training in this work are described in Table 1.
data sets. These data sets are then divided into The fuzzy neural networks are trained for two
two kinds of data with positive axial shape index kinds of data sets, the positive (relatively high
(ASI) and negative ASI. The training data set power at the top part of the reactor core) ASI
comprise one-third of the acquired input-output cases and the negative ASI cases. This results in
data pairs and the test data set comprises two- smaller errors compared with using only one
thirds of the total data. x1 through x23 represent summed data set.
the reactor power, core inlet temperature, coolant The selected number of rules of fuzzy neural
pressure, mass flowrate, axial shape index, 12 in- networks is 6 for both the positive ASI cases and
core neutron sensor signals, R1, R2, R3, R4, R5, the negative ASI cases in order to prevent
and P control rod positions, and yr is the power underfitting and overfitting problems. The
peaking factor (Fq) in the reactor core. R1 through antecedent parameters such as membership
R5 and P are the names of the control rod groups. function parameters are optimized by the back-
The used in-core detector signals are located on propagation method and the consequent
the central part of the core (a total of 12 in-core parameters qij and r are optimized by the least-
sensor signals including instrument locations squares method.
Estimation of the Nuclear Power Peaking Factor M. G. Na, et al 425

(a) Actual Fq Histogram (a) Actual Fq Histogram

(b) Error Histogram Between Actual Fq and (b) Error histogram Between Actual Fq and
Estimated Fq (without SPND Signals) Estimated Fq (without SPND Signals)

(c) Error Histogram Between Actual Fq and Fq (c) Error Histogram Between Actual Fq and
Estimated (with SPND Signals) Estimated Fq (with SPND Signals)
Fig. 2. Estimation Performance of Fuzzy Neural Fig. 3. Estimation Performance of Fuzzy Neural
Networks for Test Data with Positive ASI Networks for Test Data with Negative ASI
426 J. Korean Nuclear Society, Volume 36, No. 5, October 2004

Table 1. Input and Output Single Ranges

Input signals Nominal values Ranges

Reactor power (%) 100% 80 ~ 103

Inlet temperature 295.8 290.5~ 301.7

Pressure (bar) 155.17 131.0 ~ 160.0

Mass flowrate (kg/m2-sec) 3565.0 2994.6 ~ 4135.4

Axial shape index - 0.597 ~ -0.534

Simulated in-core detector signals - 7.4 ~ 322.0
(12 different positions)
R1 control rod positions (cm) - 0 ~ 381

R2 control rod positions (cm) - 0 ~ 381

R3 control rod positions (cm) - 0 ~ 381

R4 control rod positions (cm) - 0 ~ 381

R5 control rod positions (cm) - 0 ~ 381

R12 control rod positions (cm) - 0 ~ 381

P control rod positions (cm) - 0 ~ 381

Output signal Nominal value Range

Fq - 1.930 ~ 4.066

Table 2. Power Peaking Factors Calculated by the Fuzzy-Neural Networks

Training data Test data

Relative Relative 2 Relative Relative 2
maximum error(%) maximum error(%)
error(%) error(%)
Positive axial
Without shape index 1.354 0.370 2.510 0.371
in-core Negative axial
shape index 2.109 0.508 4.089 0.522
sensor
signals Total 2.109 0.444 4.089 0.453
Positive axial
With shape index 1.043 0.202 1.081 0.205
in-core Negative axial
sensor shape index 1.032 0.245 1.647 0.261
signals Total 1.043 0.225 1.647 0.235

Figure 2 shows the power peaking factors for relative 2-sigma error is 0.37% and its maximum
~8500 test cases and their estimation error error is 2.51% (see Table 2). If the in-core neutron
histogram (without and with in-core detector flux sensor signals are used, the relative 2-sigma
signals) for test data with positive ASI. If the in- error is 0.21% and its maximum error is 1.08%.
core neutron flux sensor signals are not used, the Figure 3 shows the power peaking factors and
Estimation of the Nuclear Power Peaking Factor M. G. Na, et al 427

Table 3. Comparison of the Power Peaking Factors

Proposed
Proposed Proposed
ASI MASTER Algorithm
Power Algorithm Algorithm COLSS
value (target) (with SPND)1)
(without SPND) (with SPND)

0.081 80 1.968 1.967 1.969 2.133 2.064

0.094 90 1.959 1.956 1.958 2.135 2.063

0.069 100 1.952 1.952 1.951 2.137 2.069

0.073 103 1.949 1.949 1.949 2.138 2.072

-0.525 80 2.778 2.773 2.776 3.000 2.886

-0.504 90 2.718 2.724 2.718 2.961 2.830

-0.483 100 2.663 2.667 2.664 2.918 2.777

-0.520 103 2.646 2.649 2.649 2.905 2.762

1) Values calculated under the assumption that all 12 in-core sensor signals are over-measured 5% more
largely than actual values

their estimation error histogram (without and with 5). Here, z denotes the axial position of the
in-core detector signals) for test data with negative reactor core and F xy is a plane-wise (radial
ASI. If the in-core neutron flux sensor signals are direction) peaking factor. In the COLSS method,
not used, the relative 2-sigma error is 0.52% and the Fxy values are prepared and provided at the
its maximum error is 4.09%. If the in-core neutron design stage according to a variety of control rod
flux sensor signals are used, the relative 2-sigma configurations. For example, for the control rod
error is 0.52% and its maximum error is 1.65%. configurations of Fig. 5, each Fxy for 3 different
If we consider the relative 2-sigma error together regions is selected by a table lookup scheme from
for both test data sets with positive and negative the F xy values prepared at the design stage.
ASIs (see Figure 4 and Table 2), the relative 2- However, in the MASTER code, the plane-wise
sigma error is 0.45% without the in-core sensor F xy values at the real core state are used to
signals and 0.23% with the in-core sensor signals. calculate the F q value. Therefore, if the fuzzy
It is known that the use of self-powered neutron neural networks of the proposed method
detector (SPND) signals reduces the estimation accurately estimate the target values, the
error by more than two times compared to not proposed method always provides a lesser or equal
using the SPND signals. Fq value relative to that of the COLSS method,
Table 3 shows other test results to compare the and the COLSS method is always equally or
Fq values with current COLSS methodology. The excessively conservative relative to the proposed
Fig values calculated by the COLSS method are method.
obtained by multiplying the core average axial Also, the rightmost values are the power peaking
power PID(z) to the F xy values of the factors calculated under the assumption that all 12
corresponding regions and by then selecting the in-core sensor signals are over-measured by 5%
maximum values of the multiplication (refer to Fig. more than the actual values, which is a very severe
428 J. Korean Nuclear Society, Volume 36, No. 5, October 2004

(a) Actual Fq Histogram Fig. 5. The Pseudo Hot Pin Axial Power
Distribution of COLSS

assumption. Even the power peaking factors for

these severe cases are lower than those of
COLSS. Thus, the proposed method secures a
larger operation margin than the current COLSS
method.
It is important to verify the fuzzy neural networks
for test data that has not been used in the training
stage. It is known that the 2-sigma error calculated
by the fuzzy neural networks for the test data is
(b) Error Histogram Between Actual Fq and similar to the relative 2-sigma error for the training
Estimated Fq (without SPND Signals) data (see Table 2). Therefore, if the fuzzy neural
networks are first trained using data for a variety
of operating conditions, they can accurately
estimate power peaking factors for any other
operating conditions.

4. Conclusions

In this work, fuzzy neural networks have been

developed and applied to the estimation of the
power peaking factor in the reactor core. The
fuzzy neural networks are trained by using the data
set prepared for training (training data) and
(c) Error Histogram Between Actual Fq and
verified by using another data set different
Estimated Fq (with SPND Signals)
Fig. 4. Estimation Performance of Fuzzy Neural (independent) from the training data. In addition,
Networks for All Test Data (Includin two fuzzy neural networks are trained for two
Positiveand Negative ASI Data)
Estimation of the Nuclear Power Peaking Factor M. G. Na, et al 429

kinds of data sets, divided into both positive ASI Nuclear Power Plant Simulator," Annals of
and negative ASI, respectively. The developed Nuclear Energy, 25, pp. 409-420, (1998).
fuzzy neural networks were applied to the first fuel 8. S. Han, U. S. Kim, and P. H. Seong, "A
cycle of the Yonggwang unit 3 PWR plant. The Methodology for Benefit Assessment of Using
relative 2-sigma error of the estimated power In-Core Neutron Detector Signals in Core
peaking factor is 0.2349% when in-core neutron Protection Calculator System (CPCS) for
flux detector signals are used and 0.4527% when Korea Standard Nuclear Power Plants
they are not used. The use of SPND signals as (KSNPP)," Annals of Nuclear Energy, 26, pp.
input signals to the fuzzy neural networks reduces 471-488, (1999).
the estimation error by about two times compared 9. M. G. Na, "Application of a Genetic Neuro-
to when the SPND signals are not employed. In Fuzzy Logic to Departure from Nucleate
summary, the fuzzy neural network is sufficiently Boiling Protection Limit Estimation," Nuclear
accurate to be used in power peaking factor Technology, 128, pp. 327-340, (1999).
monitoring. 10.M. G. Na, "DNB Limit Estimation Using an
Adaptive Fuzzy Inference System," IEEE
References Trans. Nucl. Sci., 47, pp. 1948-1953,
(2000).
1. Final Safety Analysis Report for YGN Unit 3 & 11.W. K. In, D. H. Hwang, Y. J. Yoo, and S. Q.
4, Korea Electric Power Company. Zee, "Assessment of Core Protection and
2. W. B. Terney, J. L. Biffer, C. O. Dechand, A. Monitoring Systems for an Advanced Reactor
Josson, and R. M. Versluis, "The C-E CECOR SMART," Annals of Nuclear Energy, 29, pp.
Fixed In-core Detector Analysis System,"Trans. 609-621, (2002).
Am. Nucl. Soc. 44, 542 (1983). 12.G. C. Lee, W. P. Baek, and S. H. Chang,
3. "Overview Description of the Core Operation "Improved Methodology for Generation of
Limit Supervisory System (COLSS)," CEN-312- Axial Flux Shapes in Digital Core Protection
P, Revision 01-P, ABB Combustion Systems," Annals of Nuclear Energy, 29, pp.
Engineering Inc., Nov. (1986). 805-819, (2002).
4. Final Safety Analysis Report for Wolsung Unit 13.J.-S. R. Jang, "ANFIS: adaptive-network
1, Korea Electric Power Company. -based fuzzy inference systems," IEEE
5. Tang, T. L., et al., "Analytical Design of the Transactions on Systems, Man, and
CANDU-600 On-line Flux Mapping System," Cybernetics, 23, pp 665-685, (1993).
TDAI-152, Atomic Energy of Canada Limited 14.T. Takagi and M. Sugeno, "Fuzzy Identification
(1978). of Systems and Its Applications to Modeling
6. H. C. Kim and S. H. Chang, "Development of and Control," IEEETrans.System,Man,Cybern.,
a Back Propagation Network for One-Step 1, pp. 116-132, (1985).
Transient DNBR Calculations," Annals of 15.B. O. Cho, et al., MASTER-2.0: Multipurpose
Nuclear Energy, 24, pp. 1437-1446, (1997). Analyzer for Static and Transient Effects of
7. J. K. Lee and B. S. Han, "Modeling of Core Reactors. KAERI, KAERI/TR-1211/99,(1999).
Protection and Monitoring System for PWR