Modeling and Identification of a Nuclear Reactor

with Temperature Effects and Xenon Poisoning

Attila Gábor, Csaba Fazekas∗, Gábor Szederkényi, Katalin M. Hangos

Process Control Research Group

Computer and Automation Research Institute,
Hungarian Academy of Sciences
H-1518, P.O. box 63, Budapest, Hungary
tel: +36 1 2796000, fax: +36 1 4667503
e-mail: [email protected]

Abstract

This paper presents the modeling and identification procedure for

a VVER-type pressurized water reactor. The modeling goal is to pro-
duce a mathematical description in nonlinear state-space form that is
suitable for control-oriented model analysis and preliminary controller
design experiments. The proposed model takes temperature effects
and Xenon poisoning into consideration and thus it is an extension
of formerly published simpler model structures. Real transient mea-
surement data from the plant has been used for the identification that
is based on standard prediction error minimization. It is shown that
the model is fairly well identifiable and the newly inserted model com-
ponents significantly improve the quality of fit between the measured
and computed model outputs. Furthermore, the estimated parameter
values fall into physically meaningful ranges.

Keywords: process modeling, parameter estimation, nuclear reactors, power

systems
∗
corresponding author

1
1 Introduction
The gradually changing operating requirements, the ever stringent regula-
tions related to performance, effectiveness and safety often necessitate the
re-tuning or the re-design of different subsystems in nuclear power plants.
This is particularly important in the case of elderly plants, such as the Paks
Nuclear Power Plant (NPP) in Hungary, where the refurnishment of the
control system is necessitated by the expiring of its originally designed life-
time. The Paks NPP operates four pressurized water (VVER-440/213 type)
reactors with a total nominal electrical power of 1860 MWs.
The above mentioned important re-tuning or re-design tasks are sup-
ported by the improving quantity and quality of dynamic measurements as
a result of developing hardware-software environment and modern sensor
devices. In addition, it is well known from theory and engineering practice
that the application of advanced feedback control can dramatically improve
dynamical system properties often without the need to introduce significant
changes in the technology.
Earlier efforts in developing advanced control schemes for nuclear reactors
of different type, or for other key equipments of nuclear power plants focus on
the application of soft-computing technologies (e.g. neural networks, fuzzy
controllers) that do not need a reliable dynamic model of the plant to be
controlled (see e.g. [3], [4], [18]).
For the dynamic analysis or controller synthesis, however, reliable dy-
namic models are needed where the level of detail and descriptive capability
very much depend on the exact modeling goal [12]. Most modern controller
design methods and the corresponding analysis techniques require that the
mathematical model of the system is in the form of (a preferably low num-
ber of) ordinary differential equations [14]. Unfortunately, the traditionally
available and commonly used dynamic models for nuclear power plants are
much too complex and detailed for control purposes, see e.g. the papers
[11, 15, 2, 21] applicable for NPPs with pressurized water reactors.
There are a few papers in the literature that use simple low dimensional
dynamic models to design controllers or to analyze dynamic properties of the
reactor or other equipments, such as the steam generator in nuclear power

2
plants, see e.g. [17], [22]. However, the models and the dynamic properties
of the modelled system are much dependent on the type of the nuclear power
plant they belong to, that is, on the fact whether the plant operates pres-
surized water reactors (PWRs) or boiling water reactors, etc. The studies
about the integrated control of pressurized water reactors (PWRs) are often
based on linear state-space ([5]) or input-output models ([19]). However,
these models usually do not give insight into the most important physical
processes and sub-systems.
Therefore, a simple dynamic model in physical coordinates and the cor-
responding parameter estimation procedure for the primary circuit dynamics
of VVER-type pressurized water reactors was presented in [9]. The primary
uses of this model are control oriented dynamic model analysis and high level
controller design. However, the reactor sub-model of this primary circuit
model was too much simplified that has caused a mismatch in the predicted
and observed signals related to the nuclear reactor itself. In this paper our
aim is to extend this reactor model with temperature feedbacks and the dy-
namic of delayed neutron emitting nuclei. Beside the new model structure,
the parameter estimation procedure is also presented using measured plant
data.
The paper is organized as follows. In the second section, the extended
model in state space form is presented. The third section describes the
parameter estimation method and the measurements. The fourth section
contains the results of the parameter estimation, while the conclusions can
be found in the fifth section.

2 Reactor model
There are a few simple dynamic models of nuclear reactors reported in the
literature that qualify to be suitable for control-oriented studies (see [6],
[13]) but they have mainly been developed for training and/or demonstration
purposes and they are not validated against real measured data. Below a
simple dynamic model of the nuclear reactor in a pressurized water NPP is
presented based on first engineering principles.

3
2.1 An oversimplified model

Earlier, we developed different versions of a dynamic model for the primary

circuit of a VVER-type nuclear power plants [8, 9, 10]. The domain of these
former models included the dynamic behavior in normal operating mode
together with the load changes between the day and night periods, which is
approximately the 80 − 100% thermal power range. The reactor sub-model
of our primary circuit model was a time-dependent, point kinetic model with
a single type of delayed neutron emitting nuclei whose concentration was in
a quasi steady-state [7, 16]. The effect of the control rod position on the
reactivity was approximated by a quadratic function, i.e. the original simple
reactor model was the following:

dN p1 z 2 + p2 z + p3
= N +S (1)
dt Λ
where N [%] is the neutronflux, z [m] is the control rod position, S [%/s] is
a virtual neutron source and pi , i = 1, 2, 3 are estimated parameters. This
model suffers from the following shortcomings.

• The neutron flux is independent of the temperatures, i.e. it does not

contain a temperature feedback from the temperature of the moderator
and/or the fuel.

• To be able to reproduce steady-states correctly with only one differen-

tial equation, a virtual neutron source term S has been included. This
is accepted and used in the literature, but in a more detailed model
the introduction of other elements with clear physical meaning would
be desirable.

• The concentrations of the delayed neutron emitting nuclei are assumed

to be in quasi steady-state, that is far from being realistic.

• The model does not describe the measured trends in the neutron flux
in the neighborhood of the steady states and it has some inaccuracies
when simulating load increase following a load decrease transient.

Process knowledge suggests that taking into consideration some additional

physical details, a more accurate model could be obtained with a bit higher

4
but still manageable number of state variables. Therefore, we extend the
simple reactor model in Eq. (1) with temperature feedbacks, the dynamics
of delayed neutron emitting nuclei and Xenon poisoning.

2.2 Modeling Assumptions

In order to have a low order dynamic model of the reactor the following
simplification assumptions have been made.

R1 The reactor is considered as a spatially homogeneous lumped parame-

ter system. Therefore, the physical reactor model is a time-dependent,
non-linear single-group model [16].

R2 The dynamic model of the reactor is derived from the point kinetic
equations.

R3 Only a single "average" group of the delayed neutron emitting nuclei

is assumed.

R4 The reactor is composed of the fuel, the moderator and the control rod
as modelling elements (balance volumes).

R5 The reactivity dependence on the rod position is assumed to be quadratic.

R6 The reactivity dependence on the temperatures is assumed to be linear.

R7 The reactivity dependence on the Xenon concentration is assumed to

be linear.

R8 The boron concentration is regarded to be constant during the simu-

lation together with the reactivity coefficients.

R9 The mass flow rate of the moderator is assumed to be constant.

R10 The heat loss of the reactor is neglected.

The input of the model is the control rod position z [m] and the temper-
ature of the water entering the reactor Tin [◦ C]. The outputs of the model
are the neutron flux N [%] and the average temperature of the moderator

5
Tm [◦ C]. It is important to note that both the inputs and the outputs are
measured variables.
The applied variables and the parameters with their units and definitions
can be found in Table 1. The nominal values of the variables at the 100%
operation mode are denoted by a subscript 0 .
We have to note that the structure of the fuel rod is assumed to be homo-
geneous, i.e. the mass and the specific heat of the revetment of the fuel rod
are taken into account in the mass and the specific heat of the fuel, respec-
tively. The reason of this simplification is that the change of temperature of
the revetment is proportional to the changes of the temperature of the fuel.
The rod effect has been modeled by a quadratic function. A widely used
sinusoidal rod effect function might be more accurate, but its parameter
dependence is less advantageous from an identification point of view.
Only a single averaged group of delayed neutron emitting nuclei has been
assumed after our preliminary analysis of a model version with six groups [8],
where we have found that their effects were negligible. The effective decay
constant has been calculated as follows [7]:
P6
βi
λC = P6i=1 β (2)
i
i=1 λC,i

2.3 Model Equations

The state equations are derived from the dynamic conservation balances for
the neutron and delayed neutron emitting nuclei, as well as for the internal
energy of the fuel and the moderator.

2.3.1 Neutron Dynamics

According to the assumptions R1, R2 and R3, the neutron dynamics and
the delayed neutron emitting nuclei dynamics are described by the following
equations [16]:
!
dN N β
=β ρ−1 +C
dt Λ Λ
dC
=λC (N − C) (3)
dt

6
 Table 1: Variables and parameters
Identifier M. u. Definition
z(t) m Rod position
N (t) % Neutron concentration
C(t) % Concentration of the delayed
neutron emitting nuclei
nI (t) cm−3 Iodine concentration
nX (t) cm−3 Xenon concentration
ρ(t) $ Reactivity
Tf (t) ◦C Temperature of the fuel
Tm (t) ◦C Average temperature of the moderator
Tout (t) ◦C Temperature of the water leaving the reactor
Tin (t) ◦C Temperature of the water entering the reactor
φ0 cm−2 s−1 Initial equilibrium neutron flux
Λ s Average generation time
Σf cm−1 Macroscopic fission cross section
β - Fraction of delayed neutron group
λC s−1 Decay constant of the
delayed neutron emitting nuclei
λI s−1 Decay constant of Iodine
λX s−1 Decay constant of Xenon
σX cm2 Microscopic absorption cross section
YI - Iodine yield
YX - Xenon yield
αf $/◦ C Temperature coefficient of the fuel
αm $/◦ C Temperature coefficient of the moderator
A m2 Area of a fuel rod
U W m−2 K −1 Heat transfer coefficient between the fuel
and the moderator
Mf cpf J/K Heat capacity of the fuel
Mm cpm J/K Heat capacity of the moderator
mp kg/s Mass flow rate of the moderator
F W/% Reactor heat power per 1% of neutron flux
$ $
p0 , p1 , p2 $, m , m2 Rod parameters

7
 The reactivity depends on the temperatures, the control rod position and
the Xenon concentration (assumptions R5, R6 and R7) [4]:

ρ = αf (Tf − Tf 0 ) + αm (Tm − Tm0 )

+ p2 z 2 + p1 z + p0
σX
+ (nX − nX0 ) (4)
βΣf
where αf (Tf − Tf 0 ) describes the temperature feedback of the fuel, αm (Tm −
Tm0 ) describes the temperature feedback of the moderator, p2 z 2 + p1 z + p0
σX
is the effect of the rod to the reactivity and βΣf (nX − nX0 ) is the effect of
the Xenon.

2.3.2 Equations of Thermodynamics

Energy balances are constructed for the fuel and the moderator (assumption
R4). The energy balance for the fuel is

Mf cpf dTf = −U A(Tf − Tm )dt + F N dt (5)

where Mf cpf dTf is the inner energy change of the fuel due to the temperature
change, −U A(Tf − Tm )dt is the transferred heat to the moderator and F N
is the heat power of the reactor per unit time. To describe the temperature
of the fuel Eq. (5) is transformed to the
dTf UA F
=− (Tf − Tm ) + N (6)
dt Mf cpf Mf cpf
form.
The energy balance for the moderator is

Mm cpm dTm = U A(Tf − Tm )dt + mp cpm Tin dt −

−mp cpm Tout dt (7)

where Mm cpm dTm is the inner energy change of the moderator due to the
temperature change, mp cpm Tin dt is the energy of the inlet mass flow of the
moderator and mp cpm Tout dt is the energy of the outlet mass flow of the
moderator (assumption R9). Applying a similar transformation as before we
obtain that
dTm UA mp
= (Tf − Tm ) − (Tout − Tin ) (8)
dt Mm cpm Mm

8
Let us group the parameters and introduce the following notations
UA F
A1 = A2 = (9)
Mf cpf Mf cpf
UA mp
A3 = A4 = (10)
Mm cpm Mm
With this we can transform (6) and (8) into the following form
dTf
= −A1 (Tf − Tm ) + A2 N (11)
dt
dTm
= A3 (Tf − Tm ) − A4 (Tout − Tin ) (12)
dt
However, the steady state determines the following relationships among these
parameters:

0 = −A1 (Tf 0 − Tm0 ) + A2 N0

0 = A3 (Tf 0 − Tm0 ) − A4 (Tout0 − Tin0 )

From these we can express A2 and A4 as:

A1 (Tf 0 − Tm0 )
A2 =
N0
A3 (Tf 0 − Tm0 )
A4 =
(Tout0 − Tin0 )
Furthermore, the expression Tout0 − Tin0 in the denominator of A4 can be
Tout0 +Tin0
transformed into 2 (Tm0 − Tin0 ) using the equation Tm0 = 2 .

2.3.3 Equations of Poisoning

In power reactors we cannot neglect the Xenon poisoning, because the du-
ration of load changes (4-6 hours) between the day and night is close to the
half life times of the Xenon (9 hours) and the Iodine (6 hours). The con-
centration change of Xenon is also relevant in shorter terms, because it has
great microscopic absorption cross section. To describe the concentration
change of Xenon we also have to describe the Iodine concentration [7, 16]:
dnI N
= YI Σ f φ0 − λI nI (13)
dt N0
dnX N
= YX Σ f φ0 + λI nI − λX nX
dt N0
N
− σX n X φ0 (14)
N0

9
where YI Σf NN0 φ0 and YX Σf NN0 φ0 is the effect of the Uranium fission, λI nI
is the Iodine atoms decay to Xenon in one cm3 per time unit, −λX nX −
σX nX NN0 φ0 is the decrease of the Xenon concentration because of the decay
and neutron absorption. From now on, we use the following simplifications
in the notations: X = nX /Σf and I = nI /Σf .

2.4 State Space Form

In order to be able to apply standard identification and controller design

methods, the above engineering model equations have been transformed into
a state-space model form.

State equations

dN N
= β αf (Tf − Tf 0 ) + αm (Tm − Tm0 ) + (15)
dt Λ
!
2 σX β
+ p2 z + p1 z + p0 + (X − X0 ) − 1 + C
β Λ
dC
= λC (N − C) (16)
dt
dTf Tf 0 − Tm0
= −A1 (Tf − Tm ) + A1 N (17)
dt N0
dTm Tf 0 − Tm0
= A3 (Tf − Tm ) − A3 (Tm − Tin ) (18)
dt Tm0 − Tin0
dI N
= YI φ0 − λI I (19)
dt N0
dX N N
= YX φ0 + λI I − λX X − σX X φ0 (20)
dt N0 N0
Output equation
y = [N, Tm ]T (21)

3 Parameter identification
This section is devoted to the parameter identification including the descrip-
tion of the measurements and that of the parameter estimation method.

10
3.1 Measurements

Measured data from unit 1 of the Paks Nuclear Power Plant in Hungary were
collected for parameter estimation purposes. To extract as much dynamic
information as possible, load change periods were selected for identification.
The measured data that are needed for the identification included the
neutron flux N , and the average temperature of the moderator Tm as out-
puts, the rod position z and the temperature of the inlet moderator Tin as
inputs. The data source was the Verona system (see [2]) that is a reactor
monitoring system storing also reactor data. The stored values are uniformly
sampled, the sampling time is 10 s.
The time-span of the raw measurements was between 2 and 72 hours.
The selected data sequences had to contain steady state values together with
power increase/decrease without any significant disturbances and operating
mode changes. After the investigation of the measured data, a time interval
of 7 hours was chosen for parameter estimation.
It is important to note that these data are passive measured data from
the nuclear power plant under closed-loop control where the excitation was
provided only by the load changes.

3.2 Identification Method

First, the parameters of the model have been grouped based on the knowledge
of their values.

• Known parameters. They are the parameters of nuclear processes, β,

YI , YX , λC , λI , λX . The value of these parameters is in principle
known from the literature, but because of the model simplification,
the estimated value can be different from the literature value. We
accept ±5% differences.

• Partially known parameters. A reliability domain is given to their

value, they are φ0 , σX , αm , αf , Λ, A1 , A3 and the rod parameters.

The value of these parameters have been estimated from measurements using
a constraint during the estimation: the estimated value of the parameters

11
 Rod position Temperature of the inlet moderator
0.3 266

Temperature [°C]
0.2 265.5
Position [m]

0.1 265

0 264.5

−0.1 264

−0.2 263.5
0 2 4 6 8 0 2 4 6 8
Time [h] Time [h]
Neutron flux Average temperature of the moderator
102 281

Temperature [°C]
100
280
98
Flux [%]

96 279

94
278
92

90 277
0 2 4 6 8 0 2 4 6 8
Time [h] Time [h]

Figure 1: The measured signals in unit 1.

must be in their reliability (physically meaningful) domain. The domains

were determined using available technical documentation and through con-
sultations with reactor operation experts of the plant and the Institute of
Nuclear Techniques, Budapest University of Technology and Economics. Ta-
ble 2 shows the reliability domain of the estimated parameters in a separate
column.
In addition to the parameters, the initial value of the non-measured state
variables, such as the concentrations of the delayed neutron emitting nuclei
C, the Xenon X and Iodine I, as well as the fuel temperature Tf has also
been estimated, i.e. they were considered as additional partially known
parameters. Particularly, X(0) and I(0) have a significant long term effect
on the system dynamics that is not vanishing through the whole examined
operation interval.
The above parameter estimation problem is basically an optimization
problem with objective function fobj which is bound constrained to keep
some estimated parameter values in a physically meaningful range.

12
 For the evaluation of fobj , the simulation of the system dynamics with
some parameter vector θ is required which is a computationally expensive
operation. This means that the numerical approximation and evaluation of
the gradient of fobj requires much computational effort and moreover, it can
often be unreliable because of the noise of some measurements. These facts
motivated us to choose a simple yet effective numerical optimization method
that does not need the computation of the gradient of the objective function.
The Parameter Estimation Tool of the Matlab is applied to implement
the identification algorithm. The applied identification method is the Pattern
search/Nelder-Mead search method. It is an optimization-based parameter
estimation method based on the Nelder-Mead simplex method [20]. The ob-
jective function is the SSE (sum of square of error) which measures the data
fit in terms of the 2-norm between the measured and the model-computed
output signals, i.e.
v
uR T RT
u (N̂ (t) − N (t))2 dt (T̂m (t) − Tm (t))2 dt
0
fobj = t RT + 0 RT (22)
N 2 (t)dt 2
0 0 Tm (t)dt

where N and Tm are the measured outputs, N̂ and T̂m are the model-
computed (simulated) output signals and T denotes the time-span of the
measurement/simulation.
The brief operational principle of the Nelder-Mead algorithm is the fol-
lowing (for details, see [20]). A simplex is the convex hull of n + 1 vertices
in an n-dimensional space. The method starts from an initial working sim-
plex which is created using the given initial parameter value. The algorithm
then performs a sequence of transformations (that can be reflection, expan-
sion, contraction or shrink) of the working simplex, to decrease the objective
function values at the vertices. The algorithm is terminated when the size of
the simplex is sufficiently small, or when the function values at the vertices
are close to each other in some norm. In each iteration step, the algorithm
typically needs only one or two objective function evaluations which is quite
low compared to most other methods.
It is important to note that the simplex search algorithm (similarly to
many nonlinear optimization techniques) does not guarantee that the ob-
tained point is a global minimum on the whole parameter domain. There-

13
fore it is very important to use as much prior information about the modeled
process as possible to choose proper initial parameter values for the method.
To use the simplex method, suitable initial values are needed. They are
taken from [1] and from the discussion with nuclear power plant experts.

4 Results
The parameter estimation was based on the state-space model equations
(15)-(20). The input variable to the model was the rod position z(t) and the
input temperature Tin (t), the model output variables were the neutron flux
N (t) and the moderator temperature Tm (t).
The values of the parameters β, YI , YX , λC , λI and λX were assumed
to be known from the literature, while A1 , A3 , αf , αm , φ0 , σX and Λ were
estimated using Nelder-Mead simplex algorithm.
The applied measured signals can be seen in Fig. 1. We have to note that
the rod position is measured as the difference between the nominal position
of the rod and the current rod position. If the rod is inserted from there,
then the rod position becomes positive.
The estimated parameter values can be seen in Table 2, while the neutron
flux and the temperature of the moderator fitting can be seen in Fig. 2 and
Fig. 3, respectively.
The quality of the parameter estimation is investigated by the analysis
of normalized error function as a function of each parameter, and by the
analysis of the normalized error function as a function of some pairs of pa-
rameters. The normalized error function is defined as in Eq. (22). Some
typical results can be seen in Figs. 4 and 5. The circle shows the estimated
values in Fig. 4.

4.1 Discussion

The analysis of the fit In Fig. 2 one can see the measured neutron flux,
together with the simulated one of our new model.
The steady states are reproduced very well, and the the dynamics of the
measurements are also described in an excellent way. This is particularly

14
Table 2: Values of estimated parameters and known constants
Identifier Estimated value Acceptable domain

φ0 1.3 · 1013 [1013 , 1014 ]

σX 2.805 · 10−18 [2.8 · 10−18 , 3.2 · 10−18 ]

αf −5.362 · 10−3 [−5.5 · 10−3 , −3.8 · 10−3 ]

αm −2.075 · 10−2 [−3.5 · 10−2 , −1.8 · 10−2 ]

A1 0.1056 [0.1, 1]

A3 0.8757 [0.1, 1]

p0 0.0401 [−0.1, 0.1]

p1 −0.44 [−1, −0.1]

p2 −0.966 [−1, −0.1]

Λ 2.18 · 10−5 [1.5 · 10−5 , 3.5 · 10−5 ]

λI 2.849 · 10−5

λX 2.150 · 10−5

λC 7.728 · 10−2

β 0.0065

YI 6.39 · 10−2

YX 2.2 · 10−3

15
 Neutron flux
102

100
Measured
Flux [%] 98 Simulated

96

94

92

90
0 1 2 3 4 5 6 7
Time [h]

Figure 2: The measured and the simulated neutron flux in unit 1.

Average temperature of the moderator

280.5

280
Temperature [°C]

279.5 Measured
Simulated
279

278.5

278

277.5

277
0 1 2 3 4 5 6 7
Time [h]

Figure 3: The measured and the simulated temperature of the moderator in

unit 1.

visible at the end of the power increase after 5 hours of examined operation,
where the drifting is caused by the poisoning [9]; here our model fits well,
too. In addition, the time constants of the new model correspond well to the
dynamics of the real system shown by the measurements.
The fit of the moderator temperature (see Fig. 3.) is a bit worse, but
the maximal error is below 0.2o C that is well within the acceptable range.
Here again, the dynamics of the system shown by the measurements are
reproduced very well.

16
 Parameter Sensitivity Parameter Sensitivity
200 200

180 180
Error function [%]

Error function [%]

160 160

140 140

120 120

100 100

80 80
90 95 100 105 110 60 80 100 120 140
p [%] A1 [%]
1

a, b,
Parameter Sensitivity 7 Parameter Sensitivity
x 10
200 2

180
Error function [%]
Error function [%]

1.5
160

140 1

120
0.5
100

80 0
90 95 100 105 110 60 80 100 120 140
α [%]
f
σX [%]

c, d,

Figure 4: Normalized error values vs. parameter. Circles show the estimated
values.

Analysis of the estimated values The estimated values are acceptable,

they are in their physically meaningful reliability domains.
The parameters can be classified based on the shape of the error function
as follows.

• p0 , p1 , p2 , αf , αm (Fig. 4 a, and c,).

The error value as a function of the parameter is similar to a quadratic
function, therefore a unique minimum exists. One can see that the
estimated parameter values are close to the minimum value.

• A1 , A3 (Fig. 4 b,).
The error value function as a function of the parameter is close to a
constant. It contains a lot of local minima but their values are similar.
This means that these parameters cannot be estimated properly, the
result is not sensitive to these parameters.

17
 −0.4

22
5% 175
−3

675
x 10 −0.5

%
%
225%

468 24%
%
1.8 −0.6

Normalized error value

3
1.6

10
−0.7

5%
1.4

175
125

125%
p

%
1.2

%
−0.8
1

225% 175%
468
0.8
−0.9

%
675
0.6

22
%

5%
−0.5

974
0.4 −1

32
4%
%
0.2
−1 −1.1
−0.5 −0.4 −0.5 −0.4 −0.3 −0.2
−0.3 −0.2 p2
p
p 1
1

a,
−3
x 10 −3
x 10
14
−2.5 9%
470
%
29
12 01
% 10
−3 22
1%
Normalized error value

48
−3.5

224175%
10

%
5%
−4 12
8

4%

%
22
αf

105
−4.5

5%
12
6

5%
−5

17
4
−5.5

1%
5%
17

48
2 −6
4%
22

−6.5
−0.5 −0.4 −0.3 −0.2
−3 −4 −6 −0.4 −0.2
x 10
αf p1 p1

b,

Figure 5: Normalized error values vs. two parameters and their contours.

• σX (Fig. 4 d,).
The error value function is asymmetric. On one side its gradient is
high, while in the other side it is much lower. A unique minimum
exist, that can be relatively well determined.

Based on the analysis of the error function as a function of two parameters

(see Fig. 5) one can see that the estimation of some variable pairs are strongly
correlated. For example, one can see from Fig. 5 a, and b, that there is a
unique minimum of the error function as a function of p1 and p2 , as well as
that of p2 and αm but there is a correlation between these two parameters,
i.e. they cannot be estimated totally independently.
The uncertainty of the well-identifiable parameter estimates was also
studied by analyzing the parameter dependence of the error function. For
those parameters (and initial conditions), where the shape of the error func-
tion permitted, the minimal and maximal parameter values corresponding
to the 110% of the minimal error function value were determined. Fig. 6
shows the results for αf , αm , p0 and σx . The obtained intervals (which were

18
 −6 −6
x 10 x 10
10 12

11
9

error function
10
error function

8 9

8
7
7
6
6

−5.6 −5.5 −5.4 −5.3 −5.2 −5.1 −0.024 −0.022 −0.02 −0.018
α −3 αm
f x 10

a, b,
−6 −6
x 10 x 10
9
8.5
8.5
8
8
error function

error function

7.5
7.5
7
7
6.5
6.5

6 6

5.5 5.5
0.039 0.04 0.041 0.042 2.802 2.803 2.804 2.805 2.806
p σ x 10
−18
0 X

c, d,

Figure 6: Error function dependence on parameters. Empty circles show

the estimated values, the horizontal dashed lines show 110% of the minimal
error.

symmetrized by a simple averaging) together with the estimated parameter

values and initial conditions are shown in Table 3. It is clearly visible from
the table that the reliability of the examined parameters is fairly good.

4.2 Model validation

For the validation of our model a time interval of 4 hours was chosen. The
measured data was also collected in unit 1. and close to the data used for the
identification. It is important because the values of the parameters change
through the campaign (the time period in a nuclear power plant between two
fuel-changes). The change of the moderator temperature coefficient (αm ) is
significant during the typically 48 weeks of full power operation. During this
time the boron concentration decreases from its original value to zero, and

19
 Table 3: Sensitivity of parameters and estimated initial values
Identifier Estimated value Symmetrized 110% interval

φ0 1.3018 · 1013 ±2.6 · 109

σX 2.8041 · 10−18 ±9.25 · 10−22

αf −5.362 · 10−3 ±1 · 10−4

αm −2.018 · 10−2 ±1.3 · 10−3

p0 0.0401 ±7.8 · 10−4

p1 −0.44 ±10−2

p2 −0.976 ±4.4 · 10−2

Tf 0 6.19 · 102 ±6.1

Xeinit 1.49 · 1016 ±2.9 · 1012

Iinit 2.927 · 1016 ±9.7 · 1012

this causes the decrease of αm from about −0.017 to −0.076. Certain other
parameters (β, rod parameters, αf ) are also changing slightly during the
campaign because the components of the fuel are changing, but their effect
on the model quality is less significant. In summary, we can say that the
reactor model can be considered time-invariant within a few weeks interval.
We only used an initial value estimation of the Xenon and Iodine concen-
trations for the validation. In Fig. 7 one can see the measured neutron flux
and the measured temperature of the moderator together with the simulated
one. While the simulated neutron flux fits very well to the measured data,
on the temperature graph, a maximum of 0.2◦ C difference can be seen.

5 Conclusions
A new extended reactor model for the control oriented modeling of the pri-
mary circuit of a nuclear power plant has been presented in this paper that
describes the dynamic behaviour of the reactor under normal operating con-
ditions and load changes. It is important to emphasize, however, that the

20
 Neutron flux
105
Measured
Flux [%]
100 Simulated

95

90

85
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [h]

Average temperature of the moderator

281
Measured
Temperature [°C]

280 Simulated

279

278

277
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [h]

Figure 7: The measured and the simulated neutron flux in unit 1.

model is not suitable (and not intended) for describing dynamics under non-
standard operating conditions, such as faults.
In the new reactor model, the reactivity depends on the control rod posi-
tion, the average temperature of the moderator, the temperature of the fuel,
and on the poison processes. The model parameters have been classified
appropriately, and the partially known and the unknown model parameters
together with the initial condition of the non-measurable state variables have
been estimated using a quadratic objective function and a nonlinear opti-
mization algorithm, namely, the Nelder-Mead simplex search method. The
necessary measurement data were collected from a unit of the Paks Nuclear
Power Plant, located in Hungary. The quality of estimates has also been
investigated by the analysis of the objective function.
The introduction of previously unmodeled effects resulted in the fact that
this more detailed model describes the system dynamics more precisely than
before, and a very good fit has been achieved even for the load changing
transients.

21
6 Acknowledgements
This research work has been partially supported by the Hungarian Scientific
Research Fund through grant no. K67625 and by the Control Engineering
Research Group of the Budapest University of Technology and Economics.
The third author is a grantee of the Bolyai János Research Scholarship
of the Hungarian Academy of Sciences.
The authors gratefully acknowledge the continuous support of the mem-
bers of the INC Refurbishment Project at the Paks Nuclear Power Plant.
The authors thank István Varga at the Systems and Control Laboratory of
the Computer and Automation Research Institute for his project manage-
ment work.
The authors thank Tibor Reiss at the Institute of Nuclear Techniques,
Budapest University of Technology and Economics for his support on the
reactor physics.

References
[1] The engineering toolbox, 2009. http://www.engineeringtoolbox.com/.

[2] F. Adorján, L. Bürger, A. Cserháti, I. Lux, M. Makai, J. Valkó, and

E. Végh. Experiences with the VERONA core monitoring system re-
cently installed at Paks NPP. Report KFKI-1985-96, Budapest, Hun-
gary, 1985.

[3] H. L. Akin and V. Altin. Rule-based fuzzy logic controller for a pwr-type
nuclear power plant. IEEE Trans. on Nuclear Science, 38(2):883–890,
1991.

[4] H. Arab-Alibeik and S. Setayeshi. Adaptive control of a PWR core using

neural networks. Annals of Nuclear Energy, 32:588–605, 2005.

[5] R.N. Banavar and U.V. Deshpande. Robust controller design for a nu-
clear power plant using H-infinity optimization. In Proc. of the 35th
Conference on Decision and Control, pages 4474–4479, Kobe, Japan,
December 1996.

22
 [6] A. Cammi, F. Casella, M. Ricotti, and F. Schiavo. Object-oriented
simulation for the control of the iris nuclear power plant. In IFAC
World Congress, Prague, 2005.

[7] G. Csom. Operation of nuclear power plants. Introduction into reactor

physics (in Hungarian). Press of the Techn. Univ. of Budapest, 1997.

[8] Cs. Fazekas, A. Gábor, and K. M. Hangos. Modeling and identification

of a nuclear reactor with temperature dependent reactivity. In European
Control Conference - ECC 2009, 2009.

[9] Cs. Fazekas, G. Szederkényi, and K.M. Hangos. Parameter estimation

of a simple primary circuit model of a vver plant. IEEE Transactions
on Nuclear Science, 55(5):2643–2653, 2008.

[10] Cs. Fazekas, G. Szederkényi, and K.M. Hangos. A simple dynamic model
of the primary circuit in VVER plants for controller design purposes.
Nuclear Engineering and Design, 237:1071–1087, 2007.

[11] C. D. Fletcher and R. R. Schultz. RELAP5/MOD3 Code Manual. User’s

Guidelines, vol. 5, Rev. 1. NUREG/CR-5535, INEL-95/0174. 1995.

[12] K.M. Hangos and I.T. Cameron. Process Modelling and Model Analysis.
Academic Press, London, 2001.

[13] IAEA. Pressurized water reactor simulator. IAEA-TCS-22, ISSN 1018-

5518, IAEA, Vienna, 2003.

[14] A. Isidori. Nonlinear control systems. Communications and Control

Engineering Series. Springer, Berlin ; New York, 3rd edition, 1995.

[15] A. Keresztúri, Gy. Hegyi, Cs. Marázcy, I. Panka, M. Telbisz, I. Trosztel,

and Cs. Hegedüs. Development and validation of the three-dimensional
dynamic code kiko3d. Annals of Nuclear Energy, 30(1):93–120, 2003.

[16] G. Kessler. Nuclear Fission Reactors. Springer-Verlag, Wien, New York,

1983.

23
[17] M. V. Kothare, B. Mettler, M. Morari, P. Bendotti, and C. M. Falinower.
Level control in the steam generator of a nuclear power plant. IEEE.
Trans. on Control Systems Technology, 8(1):55–69, 2000.

[18] M. G. Na. Design of a genetic fuzzy controller for the nuclear steam
generator water level control. IEEE Trans. on Nuclear Science, 45:2261–
2271, 1998.

[19] M.G. Na, D.W. Jung, S.H. Shin, J.W. Jang, K.B. Lee, and Y.J. Lee. A
model predictive controller for load-following operation of PWR reac-
tors. IEEE Transactions on Nuclear Science, 52:1009–1020, 2005.

[20] J.A. Nelder and R. Mead. A simplex method for function minimization.
Computer Journal, 7:308–313, 1965.

[21] E. Silvennoinen, K. Juslin, M. Hanninen, O. Tiihonen, J. Kurki, and

K Porkholm. The APROS software for process simulation and model
development. Technical Report 618, Technical Research Centre of Fin-
land, Espoo, Finland, 1989.

[22] P. Zanocco, M. Giménez, and D. Delmastro. Modeling aspects in lin-

ear stability analysis of a self-pressurized, natural circulation integral
reactor. Nuclear Engineering and Design, 231(3):283–302, 2004.

24