You are on page 1of 17

See

discussions, stats, and author profiles for this publication at:


http://www.researchgate.net/publication/223578846

Axial offset control of PWR nuclear reactor


core using intelligent techniques
ARTICLE in NUCLEAR ENGINEERING AND DESIGN FEBRUARY 2004
Impact Factor: 0.95 DOI: 10.1016/j.nucengdes.2003.11.002

CITATIONS

READS

10

262

5 AUTHORS, INCLUDING:
Mehrdad Boroushaki

M.J. Yazdanpanah

Sharif University of Technology

University of Tehran

38 PUBLICATIONS 192 CITATIONS

150 PUBLICATIONS 571 CITATIONS

SEE PROFILE

SEE PROFILE

N. Sadati
University of British Columbia - Vanco
278 PUBLICATIONS 833 CITATIONS
SEE PROFILE

Available from: Mehrdad Boroushaki


Retrieved on: 10 October 2015

Nuclear Engineering and Design 227 (2004) 285300

Axial offset control of PWR nuclear reactor core


using intelligent techniques
Mehrdad Boroushaki a, , Mohammad B. Ghofrani a , Caro Lucas b ,
Mohammad J. Yazdanpanah b , Nasser Sadati c
a

Department of Mechanical Engineering, Sharif University of Technology, P.O. Box 11365-9567 Tehran, Iran
Department of Electrical and Computer Engineering, Tehran University, P.O. Box 14395-515 Tehran, Iran
c Department of Electrical Engineering, Sharif University of Technology, P.O. Box 11365-9363 Tehran, Iran

Received 9 December 2002; received in revised form 27 August 2003; accepted 3 November 2003

Abstract
Improved load following capability is one of the main technical performances of advanced PWR (APWR). Controlling the
nuclear reactor core during load following operation encounters some difficulties. These difficulties mainly arise from nuclear
reactor core limitations in local power peaking, while the core is subject to large and sharp variation of local power density during
transients. Axial offset (AO) is the parameter usually used to represent of core power peaking, in form of a practical parameter.
This paper, proposes a new intelligent approach to AO control of PWR nuclear reactors core during load following operation.
This method uses a neural network model of the core to predict the dynamic behavior of the core and a fuzzy critic based on
the operator knowledge and experience for the purpose of decision-making during load following operations. Simulation results
show that this method can use optimum control rod groups maneuver with variable overlapping and may improve the reactor
load following capability.
2003 Elsevier B.V. All rights reserved.

1. Introduction
Abbreviations: ANNs, artificial neural networks; AO, axial
offset; APE, average power error; APWR, advanced pressurized
water reactor; BOC, beginning of cycle; CAOC, constant axial
offset control; CGD, center of gravity defuzzifier; CRG, control
rods group; DYNCO, a dynamical core calculation code for VVER
reactor; CODA, computerized operator decision aids; COSS, computerized operator support system; MLP, multi-layer perceptron
neural network; NARX, nonlinear auto regressive with exogenous
inputs neural network; NN, neural network; OL, overlapping between the control rods groups; RNN, recurrent neural network;
RMLP, recurrent multi-layer perceptron neural network; P, thermal
power of the reactor core; PIE, product inference engine; PIL, error
between predicted and lower limit of the core AO; PIR, error between predicted and higher limit of the core AO; PT, error between
predicted and desired core thermal power; SF, singleton fuzzifier
Coresponding author. Tel.: +98-913-2063587;
fax: +98-21-6013128.
E-mail address: boroushaki@mehr.sharif.edu (M. Boroushaki).

Controlling the nuclear reactor core during load


following operation is an important area in nuclear
engineering, particularly in advanced PWRs. Nuclear reactor core is a complex nonlinear system and
multivariable nature with high interactions between
their state variable. Any maneuver of the control rod
groups (CRGs), can induce unintended time-space Xe
oscillations, resulting in large local power peaking.
Such a complexity cannot be duly represented by
simple few points models (Kuan et al., 1992). Modern
intelligent techniques, using neural networks (NNs)
and fuzzy systems, allow to satisfactory handle such
a problem (Akin and Altin, 1991; Na and Upadhyaya,
1998).

0029-5493/$ see front matter 2003 Elsevier B.V. All rights reserved.
doi:10.1016/j.nucengdes.2003.11.002

286

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

Nomenclature
A
B
di
E(m)
FC
hi (m)
I
l
m
M
oj
R
t
U
u(m)
V
wij
x
x
Xe(m)
y
y
Y(m)
Z1

fuzzy set in input space of U


fuzzy set in output space of V
desired value of neuron i
error of neural network during the
training at time interval m
fuzzy critic output
position of control rods group
number i at the time interval m
normalized axial offset of the
reactor core
fuzzy IF-THEN rule number
time interval
total number of time intervals
during a load following transient
activation function of neuron i
set of real numbers
time
input space of fuzzy inference engine
input signal of neural network
at time interval m
output space of fuzzy inference engine
weight from neuron j to i
input variable of fuzzy system
real valued point (vector) in x
Xenon concentration of reactor core
at time interval m
output variable of fuzzy system
real valued point (vector) in y
output signal of neural network
at time interval m
Z-transform of a taped-delay-line

Greek letters
i
error signal of neuron i

learning rate
Al (x)
membership function of the lth
rule in IF section
Bl (y)
membership function of the lth rule
in THEN section
(m)
reactivity of the reactor core at the
time interval m
Subscripts
E
expected value of the reactor
core power

max
min
NN
r
RE

maximum
minimum
neural network output
fraction of real to nominal value of
reactor core power
real value of the reactor core power

In load following mode, the reactor has to track the


load changes while considering the core limitations
in local power peaking and safety margins. Axial offset (AO) is the parameter usually used to represent
complex three-dimensional phenomena of core power
peaking, in form of a practical parameter. In previous
control systems, usually crisp logics have been used to
control AO and the thermal power. However, in constant axial offset control (CAOC) strategy, in which the
reactor AO is maintained within predetermined limits,
via suitable maneuver of control rods, the variation of
AO versus reactor relative power during transients has
a fuzzy nature. Thus, using fuzzy control based methods is more suited to this type of problem and may
improve the plant control capability.
In recent years, many intelligent control methodologies have been proposed for controlling different
complex plants (e.g., steam generator). Most of these
methodologies have used generalization capabilities
of the neural network and decision-making abilities
of fuzzy systems (Narendra and Parthasarathy, 1990;
Akin and Altin, 1991; Cho and No, 1997; Na and
Upadhyaya, 1998). In this research, we proposed a
new intelligent approach to AO control of PWR nuclear reactors core during load following operation.
We used a recurrent neural network (RNN) for
modeling and fast prediction of a PWR nuclear reactor
core dynamics. This model was trained by an accurate
three-dimensional core calculation code. This model
can capture three-dimensional dynamics effects of the
core (i.e., the effects of spatial power distribution and
shadowing of the control rods, on core dynamics) in
a large rang of thermal power variation. We used a
fuzzy system, based on the operator knowledge and
experience for decision-making during load following
operations. The intelligent core controller uses the
fuzzy system and the neural network core model in
a control algorithm. This controller proposes a suitable CRGs maneuver and variable overlapping for

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

the next time interval (Boroushaki, 2003a, b). Simulation results show that this controller may improve
the responses, comparing to other control systems.

2. Artificial neural networks and fuzzy systems


In this section, some concepts related to artificial
neural networks (ANNs) and fuzzy systems are briefly
reviewed.
2.1. Artificial neural networks
ANNs consist of a great number of processing
elements (neurons), connected to each other. The
strengths of the connections are called weights. For
the modeling of physical systems, a feed-forward
multi-layered network is commonly used. It consists
of a layer of input neurons, a layer of output neurons and one or more hidden layers. In a multi-layer
perceptron MLP (Fig. 1), there is no connection
between the neurons in a given layer, so that the information is transferred from the (l 1)th layer to
the lth one. External data enter the network through
the input nodes and, through nonlinear transformations, output data are generated by the output nodes
(Yegnanarayana, 1999).
In ANNs, the knowledge lies in the interconnection
weights between neurons. Therefore, learning process
is an important characteristic of the ANN methodology, whereby representative examples of the knowledge are iteratively presented to the network, so that it
can integrate this knowledge within its structure (training phase).
In most applications of MLP, the weights are determined by means of the back-propagation algorithm,

Output

Inputs

Input
Laye

Hidden
Layer

Output
Layer

Fig. 1. A multi-layer perceptron (MLP) neural network.

287

which minimizes a quadratic cost function by a gradient descent method. During the training phase, the
weights are successively adjusted based on a set of
inputs and the corresponding set of desired output targets. First, the inputs are presented to the network and
propagated forward to determine the resulting signal at
the output neurons. The difference between the computed output vectors and the desired output targets represents an error that is back-propagated through the
network in order to adjust the weights. This process is
repeated and the learning continues until the desired
degree of accuracy is achieved (Haykin, 1999).
According to the back-propagation algorithm, when
an input is presented to the network, the activation of
each neuron is determined by:

J

oi = yi wij oj
(1)
j=0

where oi is the activation of unit i, yi is the activation


function of unit i, wij is the weight from unit j to i,
and J is the total number of inputs (excluding the bias)
applied to neuron i. The synaptic weight wi0 (corresponding to the fixed input) equals the bias applied to
the neuron i. Back propagation is then invoked to update all the weights in the network according to the
following rule:
wij (n + 1) = (i oj ) + wij (n)

(2)

where n is the iteration number, is the learning rate


and i is the error signal for unit i. The error signal k
for an output unit k is calculated from the difference
between the desired value dk and actual value yk for
that unit, while the error signal h for a hidden unit h
is a function of the error signals of those units in the
next higher layer connected to unit h and the weights
of those connections.
The back-propagation process will generally converge to a minimum that satisfies the criterion imposed by the user which usually renders the
 sum of the
squares of the error of the output signals, k (dk ok )2
less than a predetermined value. In this work, the bias
of all neurons was set to 1.
RNN are neural networks with one or more feedback connections. Given an MLP as the basic building
block, we may have feedback from the output neurons of the MLP or from the hidden neurons of the
network to the input layer. When the MLP has two or

288

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

Input
u (m)
Z-1

u (m-1)
Z-1

u (m-2)
u (m-p+2)
u (m-p+1)

Z-1

Output

Multilayer
Perceptron
(MLP)

y (m+1)

y (m-q+1)
Z-1

y (m-q+2)
y (m-1)
Z-1

y (m)
Z-1

Fig. 2. Nonlinear autoregressive with exogenous inputs (NARX) structure.

more hidden layers, the possible forms of feedbacks


expand even further. The recurrent networks have a
rich repertoire of architectural layouts (Medsker and
Jain, 1999).
Fig. 2 shows the architecture of a type of RNN,
based on a MLP. In this figure, the Z1 represents the Z-transform of a single taped-delayed-line.
The model has a single input that is applied to a
tapped-delay-line memory of p units. It has a single output that is fed back to the input via another
tapped-delay-line memory of q units. The content of
these two tapped-delay-line memories are used to feed
the input layer of the MLP. The present value of the
model input is denoted by u(m), and the corresponding value of the model output is denoted by y(m + 1);
that is, the output is ahead of the input by one time
unit. Thus, the signal vector applied to the input layer
of the MLP, consists of a data window made up as
present and past values of the plant inputs, representing exogenous inputs originated from outside the
network and delayed values of the model outputs, on
which the model outputs is regressed. This recurrent
network is referred to as a nonlinear auto regressive
with exogenous inputs (NARX) model. The dynamic

behavior of the NARX model is described by:


y(m + 1) = F(y(m) . . . y(m q + 1), u(m), . . . ,
u(m p + 1))

(3)

where F is a nonlinear function of its arguments.


2.2. Fuzzy systems
Fuzzy systems are knowledge-based or rule-based
systems (Wang, 1997; Zimmermann, 1996). The inputs and outputs of a fuzzy system are linguistic
variables (i.e., low, high, etc.), which are defined
via membership functions. The heart of a fuzzy system is a knowledge base consisting of the so-called
fuzzy IF-THEN rules. A fuzzy IF-THEN rule is an
IF-THEN statement in which some words are characterized by continuous membership functions. The
starting point of constructing a fuzzy system is to obtain a collection of fuzzy IF-THEN rules from human
experts or based on domain knowledge. These rules
use the membership functions for different input and
output of the system to determine the state of output parameters for any state of the input parameters.

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

Fuzzy Rule Base


x in U

y in V
Fuzzifier

Fuzzy sets
in U

Defuzzifier

Fuzzy Inference
Engine

Fuzzy sets
in V

Fig. 3. The basic configuration of a fuzzy system with fuzzifier


and defuzzifier.

The next step is to combine these rules into a single


system. Different fuzzy systems use different principles for this combination. There are two types of
fuzzy systems that are commonly used in the literature: TakagiSugenoKang (TSK), and expert fuzzy
systems with fuzzifier and defuzzifier. In this work,
we used second type of fuzzy systems. The basic
configuration of a fuzzy system with fuzzifier and
defuzzifier is shown in Fig. 3.
The fuzzy rule-base represents a collection of fuzzy
IF-THEN rules as follows:
IF x1 is A1 and . . . and xn is An, THEN y is B
where Ai and B are fuzzy sets in Ui R and V R,
and x = (x1 , x2 . . . , xn )T U and y V are the input
and output (linguistic) variables of the fuzzy system,
respectively.
The fuzzy inference engine combines these fuzzy
IF-THEN rules into a mapping from fuzzy sets in the
input space U Rn to fuzzy sets in the output space
V R based on fuzzy logic principles. Therefore,
fuzzy inference engine combines the rules in the fuzzy
rule-base into a mapping from fuzzy set A in U to
fuzzy set B in V. Since in most applications, the input
and output of the fuzzy systems are real-valued numbers, we must construct interfaces between the fuzzy
inference engine and the environment. The interfaces
are the fuzzifier and defuzzifier depicted in Fig. 3. The
fuzzifier is defined as a mapping from a real-valued
point x U Rn to a membership grade representing fuzzy set A in U. The defuzzifier is defined
as a mapping from fuzzy set B in V R (which is
the output of the fuzzy interface engine) to crisp point
y V . Conceptually, the task of the defuzzifier is
to specify a point in V that best represents the fuzzy
set B .

289

There are a variety of choices in the fuzzy inference engine, fuzzifier, and defuzzifier modules.
Specifically, we can propose five fuzzy inference engines (product, minimum, Lukasiewicz, Zadeh, and
Dienes-Rescher), three fuzzifiers (singleton, Gaussian and triangular), and three types of defuzzifiers
(center-of-gravity, center average, and maximum).
Not all of the 45 possible combinations proved equally
useful for using in our fuzzy system. In the rest of
the paper, we shall only report the results obtained by
one of the more suitable fuzzy systems.
In this paper, a fuzzy system with a singleton fuzzifier, a product inference engine and a center of gravity
defuzzifier (SF-PIE-CGD), has been used. In a singleton fuzzifier (SF), the membership function A (x) of
a fuzzy set A has a value equal to 1 at point x0 and
equal to 0 for points other than x0 . The product inference engine (PIE) is based on Mamadani method.
The Mamadani product inference calculates the maximum value of the products of membership functions
within IF-THEN rules (Wang, 1997). Therefore, the
membership function of a fuzzy set B in V, may be
calculated at the output of the fuzzy engine:
 n


M

B (y) = Maxl=1
Al (x )Bl (y))
(4)
i=1

where l is the rule number, x is the input vector to the


system, Al (x) and Bl (y) are membership functions
i
of the lth rule in the IF and THEN sections, respectively (Wang, 1997). The system output is calculated
through a center of gravity defuzzifier (CGD). This
defuzzifier specifies the y as the center of the area
covered by the membership function of B , that is:

yB (y)dy

(5)
y =
s

s B (y)dy

where s is the conventional integral.


3. Axial offset control in PWRs during load
following operations
A nuclear power reactor core controller has to control the core thermal power along with the safety limitation on local power peaking during load following
operation. Local power peaking in nuclear reactors,
is a complex three-dimensional phenomena, resulting

290

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

from different reactor parameters (fuel loading, power


level, temperature distribution, position of CRGs, fuel
burn up, spatial Xe oscillations, etc.). To simplify this
phenomenon, local power peaking is usually divided
into two radial and axial components. While radial
power peaking is usually flattened (via optimum fuel
loading/reloading pattern) once at the beginning of cycle (BOC), the axial power peaking is continuously
changing by perturbations created by CRG manoeuvres. Axial offset (AO), is the parameter usually used
to determine the core power peaking. This parameter
is defined as:
PT PB
AO =
(6)
PT + P B
where PT and PB represent the fraction of thermal
power generated in the top and bottom halves of the
core, respectively (Sipush et al., 1976). This parameter,
which could be easily measured on-line, via ex-core
instrumentation, still reduces the complexity of the
problem and provides an efficient practical mean to
control the reactor. Thus, the main challenge of the
reactor control, during load following operation, is to
maintain the axial power peaking (represented by AO)
within certain limits, about a reference target value.
The limitations on the core AO can be analysed in
Pr I coordinates, where I is the normalized AO
defined by:
I = AO Pr =

PT P B
Pr
PT P B

(7)

where Pr is relative core thermal power (the ratio of the


actual core thermal power to nominal thermal power).
In CAOC strategy, the limitations on the core AO
value can be shown by two parallel lines in Pr I
coordinates (Sipush et al., 1976). This means that the
core working conditions in Pr I coordinates must
lie within a certain band (e.g., 5%) during any power
transient (Fig. 4). The chosen target AO would occur
at full power, equilibrium Xe, and all rods out. This
control strategy protects the reactor from any divergent
Xe oscillation and would ensure safe operation of the
reactor during load following transients. Figs. 5 and 6
exhibit CAOC strategy in a 100 to 50% load following
operation in a typical VVER-1000 (Yousefpour and
Ghofrani, 2000). As shown in these figures, I exceeds the upper margin in absence of AO control and
is maintained within I band with AO control.

Relative power (Pr)


100

Target value of I 5% I

-5%

+5%

Normalized axial offset (l)


Fig. 4. Principles of CAOC strategy, using normalized AO (I).

4. Intelligent approach to AO control


To implement CAOC strategy in real nuclear reactor, the operator can use CRGs and/or boric acid
as control agents. Although the use of boric acid
does not affect the power distribution in the core,
but would cause some delays in the reactor response.
Using CRGs will lead to a faster reactor response but
any CRG maneuver affect the reactor power distribution, i.e., the core AO. Moreover, there is a tendency
in some cases to limit the use of the boric acid, to
avoid accumulation of liquid wastes. Thus, the CRGs,
with different overlapping, constitute the main control agent for the load following operation. It is to
be noted that in some cases, the operator may allow
the core AO to leave the band for a short period, provided that it returns back to the band, by the effect of
thermal feedbacks and/or Xe build up.
The variation and control of AO during transients
has a fuzzy nature. Thus the process can be suitably
modeled and implemented by an intelligent controller,
using a fuzzy critic. We designed an intelligent controller, which can use the knowledge and experiences
of the operator in a fuzzy decision maker to select the
best maneuvers of the CRGs. This intelligent core controller (Boroushaki, 2003a, b) includes: a NARX core
model, a control rod groups (CRGs) maneuver generator, a fuzzy critic and an optimum CRGs maneuver
finder. The NARX is updated with real plant data at
any time interval, for capturing any process dynamics not included in the training set. The fuzzy critic
considers all of the possible CRGs maneuvers, and
proposes the optimum CRGs maneuvers and overlap-

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

291

Fig. 5. Simulation results of a daily load following operation in a typical VVER-1000 without AO control.

ping, for the next time interval. This is an innovative


method for controlling PWR nuclear reactor core, and
may improve the responses, compared to other control
approaches.

5. Intelligent core controller structure


Fig. 7 shows the structure of the designed intelligent
core controller. The control algorithm used includes

the following steps:


Step 1: Defining: the expected core thermal power
(PE ); the maximum and minimum overlapping between the CRGs (OLmin , OLmax ) and
the maximum and minimum of allowable core
axial offsets (AOmin , AOmax ) during the next
time interval. This step is performed by the
Modules 1 and 2, which constitute the inputs
of the controller.

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

Relative power

292

1.0

Preference
Pactual

0.5

Control rods group


position (cm)

0.0
0

400

12

15

18

21

24

18

21

24

18

21

24

CRG # 8
CRG # 9

300
200

CRG # 10

100
0
0

12

15

0.06
0.04

Upper margine

0.02

0.00

-0.02
-0.04
-0.06

Lower margine

-0.08

Boron concentration(g/kg)

-0.10
0

12

15

2.0
1.5
1.0
0.5
0.0
0

12

15

18

21

24

Time(hr)
Fig. 6. Simulation results of a daily load following operation in a typical VVER-1000 with AO control.

Step 2: Generating all possible CRGs maneuvers (h8,


h9 and h10 indicate the positions of the CRGs
numbers 8, 9 and 10, respectively), taking into
account the initial CRGs positions, and the
maximum and minimum overlapping between
the CRGs. This is performed by the Module 3.
Step 3: Implementing each of CRGs maneuvers
to a fast core model and predicting the
core response during the next time interval
(Module 4).
Step 4: Finding optimum response of the core model
and the related CRGs maneuver (Modules
57).

Step 5: Implementing the best CRGs maneuver to


the real plant (Module 8).
Step 6: Updating the core model, using the response
of the real plant (see Section 6).
The above steps are executed at each time interval
during the transient. Detailed descriptions of the other
modules are explained in following sections.
In this research, the proposed intelligent approach
has been implemented to a Russian PWR (VVER) core
at beginning of cycle. We have used DYNCO code
for simulation of the plant. DYNCO is dynamic code
designed to simulate Russian 3000 MWt PWR core

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

STEP 1
2- Expected core thermal
power (PE).
t
STEP 2

1- Maximum and minimum of


control rod groups overlapping
(O.L), and core axial offsets
(A.O).

A.Omin

A.Omax

PE

O.L max

O.L min

3- Control rod groups (CRGs)


manoeuvre generator.
O.L

STEP 3

h8-h9-h10

4- Core dynamic behavior predictor


( a fast core model)
STEP 4
t

A.ONN

PNN

5- Plant prediction analyzer.

PIR

PIL

PT

6- Fuzzy critic.

FC

7- Optimum control rod


groups manoeuvre finder.
STEP 5
O.L

h8-h9-h10

8- DYNCO code (as a real


VVER reactor core).
Fig. 7. Intelligent core controller structure.

Updating the core


model

STEP 6

293

294

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

(VVER type 320) by means of 3260 nodes distributed


symmetrically in the core. In this type of reactor, ten
groups of control rod plus boric acid, are used to control the reactor power level and spatial power distribution in the core, while satisfying the safety criteria
(Yousefpour and Ghofrani, 2000). DYNCO can show
the effects of movement of any CRGs on the core parameters, as a function of time. Criticality of the plant
is maintained by 1000 ppm boron acid concentration
and the reactor is controlled by using three CRGs 8,
9 and 10 (other groups are fixed at the top of the core
as safety rods).
We used the C++ language to build the intelligent core controller (Fig. 7) and the DYNCO code
to represent the real plant. To evaluate validity of the
results of the core controller, during different load
followings, we defined an average power error (APE)
parameter as:
M

APE =

1 
|PrRE (t(m)) PrE (t(m))|
M

(8)

m=1

where M is the total number of time intervals,


PrRE (t(m)) and PrE (t(m)) are the real and the expected core thermal powers at the mth time interval,
respectively.
5.1. Prediction of the core dynamic behaviors using
NARX
Fig. 8 shows the neural network designed for identification of the plant (Module 4). This model includes a
RNN type NARX with 33-45-30-4 structure. We used
four reactor dynamic state variables, Xenon (Xe), core
reactivity (), thermal power (P) and AO for identification of long-term core dynamics with a time step equal
to 300 s (Kerlin et al., 1977). This structure includes:
two hidden layers, each composed of 45 and 30 neurons; 4 output units considered for Xe, P, and AO;
and 33 input units, foreseen for the present and two
past delayed CRG positions of the CRGs numbers 8, 9,
and 10 (h8 , h9 and h10 ), plus the present and five past
delayed NARX outputs. This model predicts the core
dynamics behavior much faster than the DYNCO code
(about 800 times), with an acceptable error. Although,
a 33-78-78-78-4 structure for NARX has been already
used by Boroushaki et al.(2002, 2003a, b), further researches showed that a smaller structure (33-45-30-4)

is possible and improves the response and learning


time of the model.
This designed NARX had been already trained by an
off-line batch learning, using 64 transients generated
by the DYNCO code (Boroushaki, 2002 and 2003b).
These transients include all possible core dynamic
conditions by movement of CRGs during a fixed period of 10,000 s. This model will then be updated by an
on-line learning, using the plant responses in each time
interval (step 6). The starting point of the batch learning is fixed at the beginning of the load following transient, using the initial off-line batch learning weights
matrix. The on-line learning is accomplished after acquisition of a new data set for capturing any process
dynamics, which was not included in the training set.
We have not used boric acid as a control agent,
to avoid sharp increase in number of required training transients and training time. The detailed description and characteristics of this model are provided by
Boroushaki et al. (2002, 2003b).
5.2. Plant prediction analyzer
Each of CRGs maneuvers generated in Module 3,
causes a different transient in the reactor core (variation of the AO and power). We need a suitable method
to analyze the predicted plant response with regard to
the core AO limitation and to find the optimum CRG
maneuver. We defined two parameters PIR and PIL,
to compare the predicted core model responses with
AO limitation:
PIR = k1 Imax (Pr [(m+1)])INN (Pr [t(m + 1)])
(9)
PIL = k2 INN (Pr [(m+1)]) Imin (Pr [t(m + 1)])
(10)
where t(m) and t(m + 1) are the mth and the next
time intervals, Pr [t(m+1)] is the relative core thermal
values at the (m + 1)th time interval, Imax (Pr [t(m +
1)]) and Imin (Pr [t(m + 1)]) are the maximum and
minimum limitations of the core I at the (m + 1)th
time intervals, INN is the predicted core I for a CRGs
maneuver and k1 , k2 are positive constant coefficients
that are used for scaling the parameters. Each of these
two parameters can be positive, zero or negative. If
PIR and PIL were both positive then the predicted core
working condition would be within the AO limitation

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

295

h8(t)
h9(t)

Nuclear
Reactor Core

h10(t)

Xe (t+1)

-1

-1

-1

h10(t -p)
h9(t -p)
P (t+1)

h8(t -p)

Multi Layer
Perceptron
(MLP)

(t-q)
A.O(t -q)

33-45-30-4

P(t -q)

A.O(t+1)
Xe(t -q)

Z -1

Z -1

Z -1

Z -1
(t+1)

-1

-1

-1

-1

E (t+1)
(t+1)

A.O(t+1)

P (t+1)

Xe (t+1)

Fig. 8. Nonlinear autoregressive with exogenous inputs (NARX) structure designed for reactor core modeling.

band. If one were positive and the other negative, the


predicted core working condition would be out of the
allowable AO limitation band.
We defined another parameter (power error) between the predicted and the expected core power as
follows:
PT = k3 PrNN (t(m + 1)) PrE (t(m + 1))

(11)

where PrNN (t(m + 1)) and PrE (t(m + 1)) are the predicted and the expected relative core thermal powers
at the (m + 1)th time interval and k3 is a positive
constant coefficient that is used for scaling the parameter. This parameter can be positive, zero or negative.
If PT were zero, the predicted core thermal power
would be matched with the expected one, during the
next time interval.

296

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

Table 1
Twenty-four fuzzy critic rules
Rule number
1
2
3
4
5
6
7
8

PIR
P
P
P
P
P
P
P
P

PIL
P
P
P
Z
Z
Z
N
N

PT
P
Z
N
P
Z
N
P
Z

r
Z1
Z4
Z1
Z1
Z4
Z1
Z4
Z1

Rule number
9
10
11
12
13
14
15
16

PIR
P
Z
Z
Z
Z
Z
Z
Z

Module 5 calculates three parameters of PIR, PIL


and PT for each of the transients predicted by the
NARX core model.
5.3. Fuzzy critic
The Module 6 is a fuzzy system designed for analyzing each of the transient generated by CRG maneuver. This module examines any core model response
with regard to the core AO limitations and the power
error. This fuzzy system contains a singleton fuzzifier,
a product inference engine and a center of gravity defuzzifier (SF-PIE-CGD). Inputs of the fuzzy system
include PIR, PIL and PT parameters, and output of defuzzifier is a crisp value that shows the suitability degree of the input parameters (core model responses).
The most important part of this fuzzy system is the
fuzzy rule-base, which should be written by an expert
(the operator) using his knowledge and experience for
decision-making at any core states, during load following operations. Table 1 includes 27 rules, defined
to this purpose. The inputs of the fuzzy rule-base are
Gaussian membership as follows:

if x 0.3

N(x) = 1  
2 
x + 0.3
if x > 0.3

N(x) = exp
0.2

PIL
N
P
P
P
Z
Z
Z
N

PT
N
P
Z
N
P
Z
N
P

r
Z4
Z1
Z4
Z1
Z1
Z4
Z1
Z4

Rule number
17
18
19
20
21
22
23
24

PIR
Z
Z
N
N
N
N
N
N

PIL
N
N
P
P
P
Z
Z
Z

PT
Z
N
P
Z
N
P
Z
N

r
Z1
Z4
Z4
Z1
Z4
Z-4
Z1
Z4

put FC are also Gaussian membership functions as:


 
 
yd 2
Zd (y) = exp
(15)
0.25
where d is equal to 1 and 4.
These membership functions are used in fuzzy
IF-THEN rules. Twenty-four rules in Table 1 cover
all different possible state with three inputs PIR, PIL
and PT. These IF-THEN rules are used in inference
engine of the fuzzy critic to determine the status of
the output parameter FC , in any states of the input
parameters PIR, PIL and PT. The fuzzy critic output
(FC ) shows the suitability degree of CRGs maneuvers. The Gaussian membership functions in (15) will
imply the crisp output value of FC vary between 1
and +4.

6. Case study and discussion

(14)

Figs. 9 and 10 show the results of the designed intelligent core controller with on-line learning in two
different cases. The overlap between CRGs number 8,
9, 10 has been limited between 0 and 40% and the allowable core I to 17% 5% (17% corresponds
to the AO at nominal power). Selection of the time
interval for controlling the core during a transient depends on the: execution time of the control algorithm
and maximum rate of power change during load following. In the following simulations, a time interval
of 300 s was selected. The constants k1 , k2 and k3 in
(9), (10) and (11) were selected as 10, 10 and 5, respectively.

where x is one of the three parameters: PIR, PIL and


PT. The membership functions of the fuzzy critic out-

Case 1: The first load following considered was a


16-2-4-2 transient shown in Fig. 9. This transient includes a slow decrease of the core

(12)
 x 2
Z(x) = exp
0.2

P(x) = 1


x 0.3 2
P(x) = Exp
0.2

(13)
if x 0.3
if x < 0.3

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

297

Fig. 9. Core controller results for a 16-2-4-2 load following, using on-line learning: (a) expected and real core thermal powers; (b) core
I limitations and real coreI; c) control rod groups (CRGs) number 8, 9 and 10 positions; (d) core working conditions and core I
limitations in Pr I coordinates; (e) CRGs overlap value; (f) fuzzy critic output value.

298

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

Fig. 10. Core controller results for a 16-8 load following, using local on-line learning: (a) expected and real core thermal powers; (b) core
I limitations and real coreI; (c) control rod groups (CRGs) number 8, 9 and 10 positions; (d) core working conditions and core I
limitations in Pr I coordinates; (e) CRGs overlap value; (f) fuzzy critic output value.

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

power from 100 to 50% during 2 h, constant


power at 50% for 4 h, and slow return to full
power during 2 h. The total transient time
was 35,800 s divided in 119 (=35,800/300)
time intervals. This case resulted in an APE
equal to 43.3 MWt (less than 1.5% of nominal power).
Case 2: The second load following considered was a
16-8 transient shown in Fig. 10. This transient
includes a fast decrease of the core power
from 100% to 70% with a rate of 2%/min,
constant power at 70% for 8 h, and fast return to full power with a rate of 5%/min. The
total transient time was 34,600 s, divided in
115 (=34,600/300) time intervals. This case
resulted in an APE equal to 28.1 MWt (less
than 1% of nominal power).
As can be seen in Figs. 9 and 10, the intelligent controller succeeded to keep the AO within the specified
bands in both cases, using solely CRGs maneuvers,
i.e., without using Boric acid as a control agent. It
should be noted that Boric acid can play an important
role in controlling the reactor power level, as shown
in Fig. 6. In case we allow the use of Boric acid, the
intelligent controller will better track the load and the
APE may still be reduced.
We had two limitations to compare our results with
other real case studies. First, load following mode of
control has not to date been implemented to VVER
plants. Therefore, we had to limit ourselves to transient
simulation using DYNCO code. Second, we had not
access to western benchmarked case studies and relevant computer codes. However, we strongly believe
that our proposed method, allowing the use of variable overlapping CRGs maneuvers, may improve the
load following capability of modern PWRs. Because
variable overlapping of CRGs would add a new input
control agent to the control system and this would in
turn increase the degree of freedom and controllability
of the plant.
The total execution time of the designed intelligent
core controller on a Pentium IV 1.4 MHz PC falls to
less than 115 s, from which 25 s is spent for on-line
learning and the remaining 90 s for other algorithm
steps. Computation time may be considerably reduced
by using parallel processing of the control algorithm,
and/or faster computers.

299

Finally, the following additional points are to be


considered in practical application of the proposed
method: The RNN can be trained by data recorded
from the plant load following operations during a sufficient time period; the control algorithm may be executed by parallel processing or by a very fast computer, and the boric acid may be added as a core control agent.

7. Conclusions
In this research, we tried to develop a new method
to tackle one of the important problems of modern
PWRs, i.e., improvement of load following capability of the plant, using advanced intelligent controllers.
The proposed method represents an innovative approach for identification and control of complex nonlinear plants (i.e., nuclear reactor core).
The proposed intelligent core controller may improve the plant maneuverability during load following
operations, using variable overlapping of CRGs, even
without use of boric acid as a control agent. The result
of the cases studied based on CAOC strategy shows
that the intelligent controller succeeded to control the
core AO within the specified bands during sever load
following operation (case 2 above). One of the potential applications of this method may be in design and
development of computerized operator decision aids
(CODA) or support system (COSS).
The drawbacks of this method are mainly: the need
for big amount of data for training of RNN; the relevant long training time and complexity of the controller structure. Further steps, i.e., uncertainty analysis, stability analysis, use on a reactor simulator, are to
be undertaken toward practical application in nuclear
power plants.

References
Akin, H.L., Altin, V., 1991. Rule-based fuzzy logic controller for a
PWR-type nuclear power plant. IEEE Trans. Nucl. Sci. 38 (2),
883890.
Boroushaki, M., Ghofrani, M.B., Lucas, C., 2002. Identification of
a nuclear reactor core (VVER) using recurrent neural networks.
Ann. Nucl. Energy 29 (10), 12251240.
Boroushaki, M., Ghofrani, M.B., Lucas, C., Yazdanpanah, M.J.,
2003a. An intelligent nuclear reactor core controller for load

300

M. Boroushaki et al. / Nuclear Engineering and Design 227 (2004) 285300

following operations, using recurrent neural networks and fuzzy


systems. Ann. Nucl. Energy 30 (1), 6380.
Boroushaki, M., Ghofrani, M.B., Lucas, C., Yazdanpanah, M.J.,
2003b. Identification and control of a nuclear reactor core
(VVER) using recurrent neural networks and fuzzy systems.
IEEE Trans. Nucl. Sci. 50 (1), 159174.
Cho, B.H., No, H.C., 1997. Design of stability and performance
robust fuzzy logic gain scheduler for nuclear steam generators.
IEEE Trans. Nucl. Sci. 44 (3), 14311441.
Haykin, S., 1999. Neural Networks a Comprehensive Foundation.
Prentice-Hall, New Jersey, pp. 732783.
Kerlin, T.W., Zwingelstein, G.C., Upadhyaya, B.R., 1977.
Identification of nuclear systems. Nucl. Technol. 36, 737.
Kuan, C.C., Lin, C., Hsu, C.C., 1992. Fuzzy logic control of
steam generator water level in pressurized water reactors. Nucl.
Technol. 100, 125134.
Medsker, L.R., Jain, L.C., 1999. Recurrent Neural Networks
Design and Applications. CRC Press, Boca Raton, pp. 133146.

Na, M.G., Upadhyaya, B.R., 1998. A neuro-fuzzy controller for


axial power distribution in nuclear reactors. IEEE Trans. Nucl.
Sci. 45 (1), 5967.
Narendra, K.S., Parthasarathy, K., 1990. Identification and control
of dynamical systems using neural networks. IEEE Trans.
Neural Networks 1 (1), 427.
Sipush, P.J., Ginsberg, A.P., Morita, T., 1976. Load following
demonstrations employing constant axial offset power
distribution control procedures. Nucl. Technol. 31, 1231.
Wang, L.X., 1997. A Course in Fuzzy Systems and Control.
Prentice-Hall, New Jersey.
Yegnanarayana, B., 1999. Artificial Neural Networks. Prentice-Hall
of India, New Delhi, pp. 265270.
Yousefpour, F., Ghofrani, M.B., 2000. Improvement of the axial
power distribution control capabilities in VVER-1000 reactors.
Ann. Nucl. Energy 27 (10), 949975.
Zimmermann, H.-J., 1996. Fuzzy Set Theory and Its Application.
Kluwer Academic Publishers, London, pp. 203240.

You might also like