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

net/publication/355986356

Optimization of the loading pattern of the PWR core using genetic algorithms
and multi-purpose fitness function

Article in Nukleonika · November 2021

DOI: 10.2478/nuka-2021-0022

CITATIONS READS

2 248

3 authors, including:

Wojciech Kubinski Piotr Darnowski

Warsaw University of Technology Warsaw University of Technology
12 PUBLICATIONS 20 CITATIONS 56 PUBLICATIONS 152 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Piotr Darnowski on 08 November 2021.

The user has requested enhancement of the downloaded file.

NUKLEONIKA 2021;66(4):147151
doi: 10.2478/nuka-2021-0022 ORIGINAL PAPER

Optimization of the loading pattern Wojciech Kubiński ,

Piotr Darnowski ,
of the PWR core using genetic algorithms Kamil Chęć

and multi-purpose fitness function

Abstract. The study demonstrates an application of genetic algorithms (GAs) in the optimization of the first
core loading pattern. The Massachusetts Institute of Technology (MIT) BEAVRS pressurized water reactor
(PWR) model was applied with PARCS nodal-diffusion core simulator coupled with GA numerical tool to
perform pattern selection. In principle, GAs have been successfully used in many nuclear engineering problems
such as core geometry optimization and fuel configuration. In many cases, however, these analyses focused on
optimizing only a single parameter, such as the effective neutron multiplication factor (keff), and often limited to
the simplified core model. On the contrary, the GAs developed in this work are equipped with multiple-purpose
fitness function (FF) and allow the optimization of more than one parameter at the same time, and these were
applied to a realistic full-core problem. The main parameters of interest in this study were the total power peak-
ing factor (PPF) and the length of the fuel cycle. The basic purpose of this study was to improve the economics
by finding longer fuel cycle with more uniform power/flux distribution. Proper FFs were developed, tested, and
implemented and their results were compared with the reference BEAVRS first fuel cycle. In the two analysed
test scenarios, it was possible to extend the first fuel cycle while maintaining lower or similar PPF, in comparison
with the BEAVRS core, but for the price of increased initial reactivity.

Keywords: Genetic algorithms • Fuel loading • Loading optimization • PWR • Nuclear reactor technology

Introduction

Most of the nuclear power reactors operate on the

basis of cycles that involve periodically replacing
part of the fuel to introduce additional reactivity to
compensate for the loss of fissile material and the
creation of fission products. In a new cycle, some of
the assemblies with high burnup are replaced with
the fresh ones, and some of the old assemblies are
set in a new configuration (the so-called shuffling).
This process requires reactor shutdown and opening
W. Kubiński, K. Chęć
the vessel, which results in enhanced costs due to
Warsaw University of Technology
operations and interruption in power production.
Faculty of Physics
The selection of the first core loading pattern and
Koszykowa 75 Str., 00-665, Warsaw, Poland
further shuffling schemes are therefore important in
E-mail: [email protected]
terms of the economy of the core’s operation [1, 2].
The core usually contains several hundred fuel
P. Darnowski assemblies that differ in their isotopic composition,
Warsaw University of Technology enrichment, amount of burnable absorbers (BAs),
Institute of Heat Engineering and/or other parameters, which may additionally
Nowowiejska 21/25 Str., 00-665 Warsaw, Poland vary with time. For a typical pressurized water re-
actor (PWR), the number of fuel assemblies used
Received: 18 January 2021 is in the range of 120–250 [1]. Assuming only 10
Accepted: 7 February 2021 different types of assemblies and 200 of them in the
0029-5922 © 2021 The Author(s). Published by the Institute of Nuclear Chemistry and Technology.
This is an open access article under the CC BY-NC-ND 4.0 licence (http://creativecommons.org/licences/by-nc-nd/4.0/).
148 W. Kubiński, P. Darnowski, K. Chęć

core (20 for each type), one can get 200!/(10 · 20!)
 10355 possible configurations. Finding the optimal
core fuel loading scheme is, therefore, an extremely
complex problem and requires a special approach.
One of the possible techniques for solving the
above problem may be the use of genetic algorithms
(GAs). GAs are optimization tools based on genet-
ics and Darwin’s theory of evolution [2, 3]. GAs
are based on a population of encoded chromosomes
(set of characters) containing information about the
optimized system. These chromosomes evolve over
time and approach the optimal solution. Due to the
use of random elements (mutation, crossing), GAs
are resistant to get stuck within a local optimum, Fig. 2. The BEAVRS core loading pattern for Cycle 1
which is a significant advantage in solving complex, (based on [12]).
multidimensional problems.
GAs have been successfully used in many fields, and then mirrors it symmetrically to build the whole
such as pattern recognition, data mining, and image core. The assumption of symmetry allows a signifi-
processing [4]. Additionally, they were also used cant reduction in the number of possible solutions
in the field of nuclear engineering. In the available and decreases the time needed for optimization.
literature, loading pattern [1], core design [5, 6], BA
arrangement [7, 8], online refueling [9], and thorium
loading [10] were considered. However, due to the PWR core model
required computing power, these optimizations were
typically simplified and focused on one parameter In this work, the 1000 MWe PWR defined in MIT
or used significantly simplified models. BEAVRS benchmark was applied [11]. The first
fuel cycle core design is presented in Fig. 2, and we
limited the choice of fuel assemblies to nine types
Genetic algorithm which were present during this cycle. It covers fuel
assemblies with enrichment of 1.6%, 2.4%, and 3.1%
GA operates on generations of chromosomes (usu- and the number of BA rods per assembly equals 0,
ally 50–100 per population) for which three main 6, 12, 15, 16 or 20. BAs, made of borosilicate glass,
genetic operators are used: selection, crossover, and are placed in control rods guide tubes in assemblies
mutation [2, 3]. without control rods. Neutronic calculations were
Selection is related to the assessment of the chro- performed with the PARCS core simulator [13, 14].
mosome. For this purpose, a fitness function (FF) is The model, validation, and test details are described
defined, which is to be maximized during the opera- in [15], and core definition and detailed design are
tion of the algorithm. Based on the FF value, chromo- available in benchmark definition document [11].
somes are selected for further GA steps. In general,
the higher the FF value, the greater the probability of
that chromosome to survive. Crossover is the process Power peaking factor
of gene exchange between chromosomes. Usually,
two chromosomes (parents) are selected randomly; In this work, we have introduced the PPF into our
they exchange subarrays between each other and algorithm and FFs, as it was not studied in our previ-
create new specimens (offspring). Mutation involves ous research [12]. It is the parameter that describes
random replacement of a given gene with another. quantitatively the uniformity of the heat sources
In this study, the chromosome is defined as 1/4 (also neutron flux) in the core [16]. The total nuclear
of the PWR core containing a set of numbers that PPF is defined by Eqs. (1) and (3) [15].
reflects the current configuration of the core (Fig. 1). max heat flux in the core
The algorithm operates on 1/4 of the configuration (1) PPF  Pxyz 
average heat flux in the core

and it can be divided into radial and axial parts:

(2) Pxyz  Pxy Pz

average heat flux of the hot channel

average heat flux of all channels
max heat flux of the hot channel

average heat flux of the hot channel

Fig. 1. Chromosome representing 1/4 of the optimized

core.
Optimization of the loading pattern of the PWR core using GAs and multi-purpose fitness function 149

(3) Pxyz  Pxy Pz 

 q ''  r  dz
H HC
different parts of the core, both radially and axially.
Lower PPF will lead to more uniform fuel depletion
1 and uniform coolant temperature distribution both
N
  q ''(r)dz
NC H in axial and radial directions, which can lead to more
C
economical use of the fissile material in the core.
max  q ''  rHC  

1
q ''  rHC  dz
H H Simulations
max  q ''  rHC  
 Two simulations were performed and a reference
1 BEAVRS calculation. The first one applies the algo-
 q ''(r)dz
HNC NC H rithm using a simplified form of FF. In this part, the
only goal of the algorithm was to minimize the PPF
where NC is the number of cooling channels, H is the that determines the non-uniformity of the power dis-
active height of the core, and rHC is the location of tribution. Thus, the FF took a simple form (Eq. (4)):
the hot channel. The hot channel is defined as the
channel with highest heat flux and enthalpy rise [17]. (4) FF1 = 1/PPF
In the PARCS code [13, 14], the neutronic solu-
tion is based on large nodes with XY dimensions where PPF is given by using Eq. (3).
similar to assemblies 20 cm × 20 cm × 20 cm. The In the second part, it was decided to use multi-
so-called pin-power reconstruction is necessary to -purpose FF and optimize two parameters of the
find detailed location of hot channels. In this work, core’s operation: PPF and cycle length. Therefore,
this approach was not applied, as it is beyond the the goal was also to minimize the PPF, but at the
scope of this study. The code estimates hot channel same time, extend the length of the cycle. In this
on the basis of the available nodalization and an as- part, FF took the form (Eq. (5)):
sembly is treated as a single cooling channel in the
context of Eqs. (1)–(3). The PARCS calculates all (5) FF2 = d/PPF
peaking parameters, but in this report we focused
on the optimization of the total PPF only. where d is the length of the given cycle (days).
Typically the PPF should be minimized to avoid For each of the FFs, 500 generations containing
large discrepancies in neutron flux/power between 100 chromosomes were performed (50 000 simula-
tions in total). The mutation level in both cases was
2%. Figure 1 shows the relative change in the FF
over generations.
From Fig. 3, one can see that FF increases very
fast at the beginning of the simulation; then, the
changes are smaller. After about 300 steps, regard-
less of the case, the FF reaches a maximum value.
Then, due to the mutation, the algorithm does not
converge to a specific value but oscillates around it.

Results

As a result of the simulations, two optimized core

configurations were obtained containing fuel as-
Fig. 3. Relative change of FF over generations for cases semblies from the BEAVRS benchmark. Figure 4
1 and 2. shows a comparison of the obtained configurations
(a) BEAVRS (b) Case 1 (c) Case 2

d = 333.6 days, PPFmax = 1.88 d = 451.7 days, PPFmax = 1.82 d = 512.5 days, PPFmax = 1.89
Fig. 4. BEAVRS core configuration (a), and optimal configurations obtained for Case 1 (b) and Case 2 (c).
150 W. Kubiński, P. Darnowski, K. Chęć

(a) BEAVRS BOC (b) Case 1 BOC (c) Case 2 BOC

(d) BEAVRS EOC (331 days) (e) Case 1 EOC (451 days) (f) Case 2 EOC (511 days)

Fig. 5. Radial power distribution (normalized to average) at the BOC and the EOC for the BEAVRS core (a, d), Case 1
core (b, e) and Case 2 core (c, f).

Table 1. Characteristics of the obtained configurations and the BEAVRS core

Average enrichment No. of BA Initial Cycle length

PPFmax
(%) rods keff (days)
BEAVRS 2.36 1268 1.08 1.88 333.6
Case 1 2.76 1392 1.13 1.82 451.7
Case 2 2.78 1356 1.13 1.89 512.5
(Figs. 4b,c) with the BEAVRS core configuration Table 2. The maximal Pxy values for the BEAVRS core,
(Fig. 4a). One can see that assemblies with greater Case 1 and Case 2 for BOC and EOC
enrichment are arranged at the edges of the core
for each configuration to minimize the fall of the BEAVRS Case 1 Case 2
flux at the outer boundary. Then in the center Pxymax BOC 1.33 1.26 1.34
are alternately arranged assemblies with different Pxymax EOC 1.21 1.26 1.17
enrichments but usually containing a greater amount
of burnable absorber (BA) rods to flatten the dis- towards the outer boundary. The maximum value of
tribution of the flux in the central part of the core. Pxy is 1.33 at the BOC and 1.21 at the end. In Case 1
The characteristics of the obtained configurations and Case 2 configurations, the higher Pxy values are
and the BEAVRS core are presented in Table 1. It spread throughout the core, resulting in relatively
contains an indicator of the amount of fissile mate- flatter power distribution. In Case 1, the maximum
rial used (the average enrichment calculated as the value of Pxy is 1.26 for both the beginning and the
sum of the enrichments in the assemblies divided EOC. For Case 2, the maximum initial Pxy is 1.34,
by their number), the number of BA rods, the initial which drops to 1.17 at the EOC.
keff, the maximum value of PPF throughout the cycle
(PPFmax), and the length of the cycle.
As the parameter optimized for both cases Conclusions
of the algorithm’s operation was PPF, it was decided
to plot the normalized radial power distribution at As part of the study, two optimizations were per-
the beginning and end of the cycle (EOC) for the formed, the first of which aimed at minimizing the
BEAVRS and the cores obtained by the algorithm PPF and the second at minimizing the PPF, while ex-
(Fig. 5a–f). tending the fuel cycle. As a result of the simulations,
As can be seen from Fig. 5, the power distribu- two configurations were obtained, both character-
tion for BEAVRS is characterized by an initial peak ized by a longer fuel cycle compared to the original
in the center of the core that spreads at the EOC configuration. In Case 1, the maximum PPF value
 Optimization of the loading pattern of the PWR core using GAs and multi-purpose fitness function 151

was reduced from 1.88 to 1.82 and the cycle was ex- 4. Jayalal, M. L., Murty, S. A. V. S., & Baba, M. S. (2014).
tended by 118 days. In Case 2, a slightly higher PPF A survey of genetic algorithm applications in nuclear
(1.89) was obtained, while the cycle was extended fuel management. J. Nucl. Eng. Technol., 4(1), 45–62.
by 179 days. Of course, other core parameters have 5. Pereira, C. M. N. A., Schirru, R., & Martinez, A. S.
also changed. The cores in Cases 1 and 2 use more (2000). Genetic algorithms applied to nuclear reac-
fissile material (average enrichment 2.76% and tor design optimization. In Da Ruan (Ed.), Fuzzy
2.78% vs. 2.36% in the original configuration) hence systems and soft computing in nuclear engineering
also have a higher initial keff (both 1.13 vs. 1.08). (pp. 315–334). Springer. https://doi.org/10.1007/978-
However, all the cores use a similar number of BA 3-7908-1866-6_14.
rods (around 1300). The larger fissile inventory was 6. Martín-del-Campo, C., Palomera-Pérez, M. Á., &
the main reason of longer cycles in comparison to the François, J. L. (2009). Advanced and flexible genetic
BEAVRS core. Additional constraints with bounding algorithms for BWR fuel loading pattern optimization.
keff or fissile mass will be studied in future research. Ann. Nucl. Energy, 36(10), 1553–1559. https://doi.
The above-mentioned examples show the suc- org/10.1016/j.anucene.2009.07.013.
cessful operation of the algorithm. As part of the 7. Khoshahval, F., & Fadaei, A. (2012). Application of
simulations, it was possible to optimize the objec- a hybrid method based on the combination of genetic
tives included in the multi-purpose FF, i.e., flattening algorithm and Hopfield neural network for burnable
the power distribution and extending the cycle. To poison placement. Ann. Nucl. Energy, 47, 62–68.
propose a comprehensively optimized configuration DOI: 10.1016/j.anucene.2012.04.020.
that could be used in a nuclear power plant, future 8. Yilmaz, S., Ivanov, K., Levine, S., & Mahgerefteh, M.
simulations may include more extensive FFs that (2006). Application of genetic algorithms to optimize
would take into account other constraints such burnable poison placement in pressurized water
as maximum average enrichment, acceptable keff reactors. Ann. Nucl. Energy, 33(5), 446–456. DOI:
range, control rods worth, and/or other parameters, 10.1016/j.anucene.2005.11.012.
optimizing the core using many objectives at the 9. Huo, X., & Xie, Z. (2005). A novel channel selection
same time. method for CANDU refuelling based on the BPANN
and GA techniques. Ann. Nucl. Energy, 32(10),
1081–1099. DOI: 10.1016/j.anucene.2005.03.003.
Acknowledgments. The second author (PD) would like 10. Mishra, S., Modak, R. S., & Ganesan, S. (2009). Op-
to acknowledge the National Atomic Energy Agency timization of thorium loading in fresh core of Indian
(PAA, Polish regulatory body) for support of nuclear- PHWR by evolutionary algorithms. Ann. Nucl. Energy,
-engineering related research at Warsaw University of 36(7), 948–955. DOI: 10.1016/j.anucene.2009.03.003.
Technology. Especially, thanks to Dr Ernest Staroń 11. Massachusetts Institute of Technology. (2018).
and Szymon Suchcicki for very fruitful cooperation in BEAVRS – Benchmark for Evaluation and Validation
reactor physics. Special thanks are dedicated to Prof. of Reactor Simulations. Rev. 2.0.2. MIT Computa-
Tomasz Kozłowski and Dr Michał Pawluczyk for fruit- tional Reactor Physics Group. Available from https://
ful discussions. We would like to thank anonymous crpg.mit.edu/sites/default/files/css_injector_im-
reviewers for their comments and efforts towards ages_image/BEAVRS_2.0.2_spec.pdf.
improving our manuscript. 12. Kubiński, W., Darnowski, P., & Chęć, K. (2021). The
development of a novel adaptive genetic algorithm
for the optimization of fuel cycle length. Ann. Nucl.
ORCID Energy, 155, art. ID 108153. DOI: 10.1016/j.anu-
cene.2021.108153.
P. Darnowski http://orcid.org/0000-0003-1602-5379 13. Downar, T., Xu, Y., Seker, V., & Hudson, N. (2010).
W. Kubiński http://orcid.org/0000-0002-5805-367X PARCS v3.0 U.S. NRC Core Neutronics Simulator.
Theory manual. USNRC.
14. Downar, T., Xu, Y., Seker, V., & Hudson, N. (2010).
References PARCS v3.0 U.S. NRC Core Neutronics Simulator.
User manual. USNRC. Available from https://www.
1. Israeli, E., & Gilad, E. (2017). Novel genetic algo- nrc.gov/docs/ML1016/ML101610098.pdf.
rithms for loading pattern optimization using state-of- 15. Darnowski, P., & Pawluczyk, M. (2019). Analysis
the-art operators and a simple test case. J. Nucl. Eng. of the BEAVRS PWR benchmark using SCALE and
Radiat. Sci., 3(3), 1–10. DOI: 10.1115/1.4035883. PARCS. Nukleonika, 64(3), 87–98. DOI: 10.2478/
2. Holland, J. H. (1992). Adaptation in natural and nuka-2019-0011.
artificial systems: an introductory analysis with 16. US Nuclear Regulatory Commission. (2005). West-
applications to biology, control, and artificial intel- inghouse technology systems manual. Section 2.2
ligence. MIT Press. – Power distribution limits. Rev 0508. USNRC.
3. Goldberg, D. E., & Holland, J. H. (1988). Ge- Available from https://www.nrc.gov/docs/ML1122/
netic algorithms and machine learning. Guest ML11223A208.pdf.
editorial. Mach. Learn., 3(2), 95–99. DOI: 10.1007/ 17. Anglart, H. (2011). Applied reactor technology.
BF01920603. Stockholm: KTH Royal Institute of Technology.

View publication stats