Electrical Power and Energy Systems 152 (2023) 109203

Contents lists available at ScienceDirect

International Journal of Electrical Power and Energy Systems

journal homepage: www.elsevier.com/locate/ijepes

Optimization method for short circuit current reduction in extensive

meshed LV network
David Topolanek a, *, Vit Krcal a, Ladislav Foltyn b, Pavel Praks b, Jan Vysocky c,
Renata Praksova b, Giuseppe Prettico d, Gianluca Fulli d
a
Brno University of Technology, Department of Electrical Power Engineering, Brno, Czech Republic
b
VSB–Technical University of Ostrava IT4Innovations National Supercomputing Center, Ostrava, Czech Republic
c
VSB–Technical University of Ostrava, Electrical Power Engineering Department, Ostrava, Czech Republic
d
European Commission Joint Research Centre, 21027 Ispra, VA, Italy

A R T I C L E I N F O A B S T R A C T

Keywords: This paper focuses on the analysis of suitable optimization methods applied to large meshed low-voltage net­
Meshed network works. The introduced methods aim to minimize the SC-current contribution by simultaneously fulfilling well-
Optimization defined operational constraints. The high number of binary variables (134) used in the worst case to deter­
Short circuit
mine the possible network configurations generates 1040 possible solutions. For this reason, the paper focuses
Reconfiguration
Distribution network
only on methods capable of reducing the necessary steady-state calculations to a tractable size. Depending on the
Evolutionary algorithm case under study, the deterministic method turns to be faster than the other ones at the price of reaching only a
local minimum. In contrast, if a longer computation time is tolerated, then evolutionary algorithms succeed in
finding the global optimum.

1. Introduction network topology reflects the process of gradual electrification bound

with the historical urban development. Traditionally, due to the limited
The decarbonisation process relies, among others, on the widespread number of measuring devices deployed, it was impossible to build
integration of Distributed Energy Resources (DER) connected to Distri­ effective numerical models and expert optimization methods verified
bution Networks (DN). A common practice of DN operators for and tested against real operational data [5]. Currently, it is vital to
increasing DER hosting capacity is to decrease the short-circuit (SC) design robust optimization algorithms, which can find the global opti­
impedance at the connection nodes. This is achieved by permanently mum in a short time even for these complex meshed networks. Nowa­
closing selected disconnectors [1]. As a result, radial networks are days, especially appropriate optimization tools verified in meshed DNs
turned into rings or meshed topologies, which are more difficult to are missing. These tools should effectively assess key operation aspects,
operate. In this case, the support of accurate numerical optimization such as transformer-taps setting, power-line sections power flow distri­
tools becomes fundamental [2]. bution, SC-currents evaluation, etc. The presentation of an effective
Decreasing the SC impedance is beneficial from the DER imple­ optimization tool that could be useful in large meshed electrical net­
mentation point of view, but there are also some negative impacts on the works, is the goal of this paper. Regarding both practical and demon­
network operation. The shortcomings of meshing the network topology strative aspects, the study is focused on finding suitable network
rest in impaired network simplicity, complicated fault localization, and configurations, which leads to minimizing the SC currents while
increased magnitude of SC currents [3]. Although meshed networks are respecting defined operational constraints. This goal is considered
capable of hosting densely concentrated and high-power loads and rather significant when LV network operation safety is taken into ac­
sources, their operation is hardly possible without the support of accu­ count, while until now, this issue has been often overlooked due to the
rate numerical models and suitable optimization tools [4]. Very large absence of proper simulation tools.
and dense networks – in radial or (partly) meshed configurations – are
typical for old city centres of big cities, where the (mostly cable-based)

* Corresponding author.
E-mail address: topolanek@vut.cz (D. Topolanek).

https://doi.org/10.1016/j.ijepes.2023.109203
Received 23 November 2022; Received in revised form 11 April 2023; Accepted 27 April 2023
Available online 9 May 2023
0142-0615/© 2023 Elsevier Ltd. All rights reserved.
D. Topolanek et al. International Journal of Electrical Power and Energy Systems 152 (2023) 109203

2. State of the art analysed meshed DN. This is a real issue which is considered very
relevant by DSOs. In general, a way to reduce the SC-current magnitude
As most of the DNs have radial topology, few publications can be for individual network nodes is to increase the impedance magnitude of
found in the literature that assesses the subject. The meshed topology the individual branches of that network. According to the physical tool
introduces certain constraints to the problem of electrical network used to increase the impedance of individual branches of the network,
analysis and optimization, and many of the analysis and optimization SC-current minimizing methods can be divided into two groups: 1.
methods developed for radial DNs cannot be used for them without methods considering the SC-current limiters installation into the
appropriate modifications. Due to the high complexity of meshed to­ network, and 2. methods considering disconnecting individual parts of
pology, the analysis and optimization of meshed electrical networks are the network, i.e., methods using bus splitting and area separation.
more computationally intensive than the analysis and optimization of Methods from the first group are presented in many papers, e.g., in
radial networks. This problem calls for developing very effective [21–25]. The SC-current limiter (or the fault-current limiter) is a special
methods capable of finding the solution in a short time. Methods of this network element of a nonlinear character. When a current correspond­
kind can be applied in real-life operating conditions and are considered ing to the normal operating state of the network flows through the SC-
more interesting by grid operators. current limiter, the magnitude of the impedance of this network
Selected papers proposed various approaches addressing the speci­ element is small. However, when a fault occurs here and the magnitude
ficities of optimization problems with meshed networks. To overcome of the current flowing through the limiter increases to the fault levels,
the problem under study, the optimization method in [6] uses bacterial the element impedance value increases by orders of magnitude. While in
foraging oriented by particle-swarm optimization. The goal of this [21–24] the methods presented here are described and analysed only
optimization method is to balance unbalanced meshed DNs. In [7], the partially, [25] provides a more detailed description, through an iterative
same problem is tackled by using the graphical model of nodal voltages. mixed-integer nonlinear programming method and a detailed analysis of
The presented algorithms determine the topology of the analysed the method-solving abilities (local analysis includes tests with IEEE
network and detect topology changes by using the algebraic and struc­ testing network models, namely the 9-bus and IEEE 30-bus, and with the
tural properties of these graphical models. In [8], a graph analysis model of a real North American system).
approach is used. In the paper, a graph-theory-based algorithm de­ This paper presents one of the methods from the second group. Bus
termines the minimum breakpoint set of a meshed DN. Another algo­ splitting and area separation are inexpensive and quickly applicable
rithm based on the network-graph analysis is described in [9], where a tools for SC-current limitation, and many DSOs prefer to use these
method for the optimal restoration of generation units powering a methods to control the SC-current-limitation process. To the best of our
meshed DN is presented. The method presented in the paper uses the knowledge, currently, only a few conference papers present partially
load-transfer algorithm based on the feeder-incidence diagram. In such methods without providing a comprehensive analysis, e.g.
[10,11], meshed-topology issues are addressed by using special mathe­ [26–29]. To cope with this gap, we created two new SC-current limi­
matical algorithms. In [10], a special method for the unit commitment tation methods using bus splitting and area separation, and analysed
and economic dispatch problem for an isolated meshed electrical these methods within a benchmark, containing these two local and two
network is presented. In [11], a framework for analysing the resilience other methods using evolutionary optimization frameworks (specif­
of meshed networks exposed to extreme conditions, with integrated ically, Pymoo and HEBO). The first local method is deterministic and the
microgrids, is presented. Within this framework, the Monte Carlo second one is based on a genetic algorithm (GA) specified to manage
method is used. Statistical methods are also used in an algorithm pre­ DNs.
sented in [12]. The local multistage algorithm solves the problem of
meshed-DN-optimal-expansion planning where the DN-equipment- 3. Problem formulation
reliability assessment is explicitly implemented as a constraint. Some
optimization methods overcome the high difficulty of the meshed DN This paper aims to compare methods that can be effectively used for
analysis by dividing the entire problem into a series of smaller problems solving optimization problems in highly meshed DNs. Problems in these
and then, they solve these smaller problems in a distributed manner networks are characterized by big computational complexity and
using a multi-agent system (MAS). Such methods are presented in [13]- require finding the global optimum. As mentioned in the introduction,
[16]. the presented optimization problem focuses on the SC-current reduction
MAS-based optimization methods presented in [13,14] divide the in the network, while respecting defined operational constraints and SC-
entire meshed DN into several subnetworks. Each local agent controls current ratings. This optimization problem can be described by an
reactive power flows across one subnetwork and the work of all agents objective function (see Section 3.3) and constrained by operational
brings the optimal voltage values across the entire controlled DN. In­ conditions that are defined in Section 3.2.
dividual agents of a MAS presented in [15] are controlled by the dual- The meshed topology of DNs on MV and LV levels is quite rare and
consensus alternating-direction method of multipliers. The local MAS- therefore, not many studies or projects deal with this kind of network.
based control method solves the economic-dispatch problem in All standard IEEE test feeders (13-bus, 34-bus, 37-bus, 69-bus, and 123-
meshed microgrids, and the realization of the local control process is bus models) have radial topology, are small and represent MV systems.
very undemanding in terms of the amount of data transferred. A MAS- The 8500-node test feeder is designed for testing tools on large net­
based optimization method presented in [16] also solves the works, but its structure is also radial. For the representation of urban
economic-dispatch problem. The local method controls a group of meshed networks, the 342-node LV networked test system was designed
meshed microgrids that are interconnected. Some optimization methods [30,31], but this system has an insufficient number of nodes compared
overcome difficulties related to the meshed DNs using the game-theory to the intended complex test network. Since no large test meshed
framework. Examples of such methods are found in [17–20]. Optimi­ network (with over 1 000 nodes) is available, a real-based LV cable
zation methods in [17–19] minimize the power losses in meshed DNs network of a European city was chosen with the scope of testing the
powered by many DER units. An optimization method in [20] uses a proposed optimization methods. This network is shortly introduced in
demand-response program to decrease the DN’s operation cost during section 3.1. The specifications of this test DN are available to other re­
the DN’s peak-load periods. Besides the game-theory methods, this searchers to confront the results of other optimization methods exposed
optimization method uses artificial neural networks to surplus the issues to the operational condition of the large meshed DN [32].
related to the meshed DNs.
The optimization problem discussed in this paper aims at the mini­
mization of the SC-current magnitudes for the individual nodes of the

2
D. Topolanek et al. International Journal of Electrical Power and Energy Systems 152 (2023) 109203

Fig. 1. Topology of test DN.

3.1. Test LV meshed network This condition also ensures that there are no customers disconnected
from the network since the voltage in these nodes would be equal to
The meshed network model represents an underground cable zero.
network placed in the collectors of an old city centre. Furthermore, there
are no DERs (e.g. photovoltaic or wind power plants) due to the limited 3.2.2. Branch loading constraint
space and historical concept of the network. The network consists of 1 The branch elements (i.e., lines and DTs) cannot get overloaded in
268 nodes, 1 498 lines and 79 MV/LV distribution transformers (DT) any normal operation state. The set limitations should respect the
that are placed in 55 MV/LV underground DT stations (one or two DTs in physical state and are usually defined in the range from 70 % to 90 % of
each DT station connected to LV network). In the LV network, there are the rated values to provide a safety margin. For the sake of clarity, the
also 55 manually controlled bus couplers (BC) placed at MV/LV DT loading limits were set to 70 % for all the branch elements (lines and
stations or in some bigger switchgears. In the European context, this DTs) in the presented case.
network could be considered as one of the largest meshed LV networks.
loading percentj,k ≤ 70% (2)
The network topology is depicted in Fig. 1.
where loading_percent is the relative current loading j is the number of
3.2. Meshed network operation constraints lines (1498 lines) and k is the number of LV/MV transformers (79
transformers).
To ensure reliable and safe network operation, the constraints
specified by current network codes, standards or by DSOs assessments 3.2.3. Upstream active power flow constraint
need to be met. In this study, four main operational constraints are Upstream active power flow over MV/LV DTs, also known as reverse
applied. The first three are considered hard ones, meaning that these flow, is not allowed for the study therefore condition (3) has to be met.
conditions must be satisfied for every feasible solution. The SC-currents This constraint prevents the supply of a part of the HV system from the
constraint, instead, is considered soft since its violation can be possible LV system. For this purpose, the DSO also uses overflow protections that
and it is penalized within the objective function (see section 3.3). disconnect the given MV/LV DT when the value of upstream active
power flow crosses the specified limit.
3.2.1. Node voltage constraint
pLV k ≤ 0 (3)
According to the voltage quality standard EN 50160, the per unit LV
phase voltage vi related to nominal voltage must be in the range where pLV is the active power flow at the low voltage transformer node
0.1 ≤ vi ≤ 1.1 (1) (direction from LV network to MV) of the k-th MV/LV transformer.

where i is the number of LV node of the network (1268 nodes).

3
D. Topolanek et al. International Journal of Electrical Power and Energy Systems 152 (2023) 109203

calculated SC current, as shown in (5):

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1 ∑nLV
MIk = (ΔSCCRi )2 (5)
nLV i=1

( )
ΔSCCRi = max 0, I ′′ki − SCCRi

where Ik“ is the initial symmetrical SC-current of the i-th node, and
SCCRi is the SC-current rating of i-th node. The main goal for using this
penalty function is to minimize the rate of exceeding the SCCR in the
monitored nodes. Used penalization deliberately ignores the costs of
upgrading/modifying the equipment, since the possibility of increasing
the SCCR depends on the extent of exceeding the SCCR and on the type
of switchgears or fuse boxes.
Value MIk of the objective function (5) corresponds to an evaluation
index (score). In the optimization procedure, complementary parame­
ters are evaluated for each feasible solution to better assess the perfor­
mance of the tested methods. These parameters are:

- Over_SCCR: the total number of nodes, where the Ik“ is higher than
SCCR,
- Ik_max: maximum value of Ik“ in LV network,
- NOSC: number of evaluations of the SC function used for obtaining
the solution.

3.4. Reconfiguration variables

To reduce the SC-current contribution, changes in the network

configuration can be implemented. The changes in the network config­
uration will cause an increase in the network SC impedance. From the
DSO perspective, the easiest way to reconfigure the network is to use the
Fig. 2. Flowchart of the DOA.
installed switches. Although it is a common practice that dispatchers use
their best experience and judgement to carry out operations in the DNs,
3.2.4. SC-currents constraint this approach does not usually lead to optimal solutions, especially when
To meet the safety criteria of network operation, the initial sym­ dealing with complex networks, such as those used in this study.
metrical SC-current, calculated in accordance with standard IEC 60909 Therefore, in this paper, the optimization is achieved through network
[33], must be lower than the SC-current ratings (SCCR) of the reconfiguration utilizing selected optimization methods. A novel
switchgears reconfiguration state is obtained by changing the operation status (ON/
Ik˝ i ≤ SCCRi (4) OFF) of selected elements in the network. For this paper, two sets of
binary variables defined by switchable elements (SE) are considered in
where Ik˝ is the initial symmetrical SC-current of the i-th node, and SCCRi the following:
is the SC-current rating of the i-th node.
In the case when the SCCR is exceeded, the short-circuit current - Case A: switching of BCs (55 binary variables),
capability of the given switchgear has to be increased, which requires - Case B: switching of BCs and DTs (55 + 79 = 134 binary variables).
considerable additional costs. Therefore, the reduction of the SC-current
magnitude below the SCCR limit is the objective of the optimization From the number of binary variables, it is inevitable to use effective
problem. This is the reason why this constraint is considered as soft and methods, since the brute-force approach would result in an immense
the penalization of not meeting this condition is defined by Eq. (5). number of possible combinations. In case A, for instance, this corre­
Another possible constraint that can be applied to this optimization sponds to 255 ≈ 3.6⋅1016 alternative configurations while for case B, the
problem is the N-1 criterion. The n-1 criterion requires that the system number of combinations increases up to 2(55+79) ≈ 2.2⋅1040 (including
can fulfil all local constraints during unexpected failure or outage of a also unacceptable operating configurations).
single system component. Due to the robustness of the analysed meshed
network and due to the immense calculation time, the N-1 criterion was 4. Applied methods
not evaluated by optimization methods in this study.
To reduce the NOSC, four methods were chosen: one deterministic
method partially using a heuristic approach and methods based on meta-
3.3. Optimization method objective heuristics, specifically evolutionary algorithms.

The optimization problem is considered with a single objective to 4.1. Deterministic optimization algorithm (DOA)
minimize the SC currents in LV nodes by applying appropriate network
configuration when all operation constraints defined in the section This algorithm uses a heuristic approach supplied with deterministic
above are respected. From the operational point of view, it is key to rules based on detailed knowledge of the network operation. Knowledge
lower the SC currents below the SCCR values defined for each node. The of network operation coming from the DSO’s experiences is used to
quantification of the objective function is calculated as the root mean design a suitable algorithm procedure capable of reaching the optimal
square of excess of SCCR in all respected LV nodes nLV. The excess is configuration.
considered zero for the nodes, where the SCCR value is higher than the The process follows the same principles of heuristic greedy

4
D. Topolanek et al. International Journal of Electrical Power and Energy Systems 152 (2023) 109203

Fig. 4. Convergence of the optimization methods – CASE A.

Fig. 5. Convergence of the optimization methods – CASE B.

Table 1
Case A results.
Optimization method DOA HEBO SGA Pymoo

Constraints compliance Yes Yes Yes Yes

MIk (kA) 4.0510 4.0510 4.0510 4.0510
Over_SCCR (-) 164 164 164 164
Ik_max (kA) 72.89 72.89 72.89 72.89
Fig. 3. Flowchart of the DOA – selection of the most feasible element. NOSC (-) 1 326 450 1379 1 100
Time (h:mm) 00:04 1:30 00:04 00:04
algorithms where at each step the optimal choice is made. This means
that the SE, whose disconnection leads to a minimal value of MIk score, is
chosen and switched off. In the case of disconnecting both the BC and Table 2
the DT (Case B), a deterministic rule is used: After the DT is switched off, Case B results.
the LV BC on its secondary side must be switched on. Clearly, this applies Optimization method DOA HEBO SGA Pymoo
only to DTs that are working in parallel in some substations. The flow­
Constraints compliance Yes Yes Yes Yes
chart depicting the algorithm procedure is shown in Fig. 2 and Fig. 3.
MIk (kA) 0.5210 0.4528 0.2869 0.2589
The main program of SEs disconnection is shown in Fig. 2, where the Over_SCCR (-) 19 30 15 14
algorithm starts for all SEs switched on. The main loop is terminated if Ik_max (kA) 51.07 49.04 46.95 44.45
the SC condition is satisfied (MIk = 0) or there are no more SEs left to NOSC (-) 3681 2000 40 178 40 000
switch off. The inner loop (Fig. 3) examines each feasible configuration Time (h:mm) 0:11 56:30 1:55 2:24

respecting subsequent disconnection of selected SEs respecting the

deterministically defined priority of DTs being switched off only if all configuration is feasible (constraints specified in section III.B) are met),
possible manipulations with the BC have been used. For each SE, the the SC currents are calculated, the objective function (5) is evaluated
switching operation is carried out and the load flow results of this hy­ and the MIk is stored. From all the possible SEs, the one leading to the
pothetical state are checked for operation constraints. If the lowest value of MIk is chosen for switching off. This reflects the

5
D. Topolanek et al. International Journal of Electrical Power and Energy Systems 152 (2023) 109203

Fig. A1. Histogram of SC currents in the network for three referential Fig. A3. Histogram of lines and MV/LV transformers loading for three refer­
configurations. ential configurations.

Adaptation Evolution Strategy1. The Pymoo package was primarily

created for continuous optimization problems. On the contrary, our
electricity network optimization task represents a discrete problem, as
the algorithm deals with discrete states of the electricity network. It is
because network switches are represented by discrete states (ON/OFF).
Consequently, only the GA could successfully solve our optimization
problem. A general GA tutorial can be found in [35].

4.3. Specific version of GA (SGA)

The next implemented method is based on a specific version of the

GA derived for electrical network optimization problems. The GA
already has provided valuable results in the domain of electrical
network optimization [36]. The advantage of the GA, in general, is that
no a priori information about the optimized problem is needed. The
greatest disadvantage of the GA is its high amount of required function
evaluations. Fortunately, the GA can use the advantages of parallel
computing, by which the computational time can be significantly
reduced; see [34,36].
In GA, the key piece is the chromosome: an array of 0 and 1, which is
Fig. A2. Histogram of bus voltages for three referential configurations. generated randomly. In our implementation, this array represents the
states of BCs and of DTs present in the network under study. Their value
aforementioned greedy algorithm principle. can be either 1 (switched on) or 0 (switched off). Each chromosome thus
The advantage of this sequential approach is that it reflects real corresponds to a given configuration of the network. To reduce the
operation procedures when the elements are switched off sequentially as computing time, the “check procedure” dealing with unconnected
well. Furthermore, the operation constraints are evaluated in every step network topology configurations was also implemented. The check
of the procedure. The analysis carried out according to this principle procedure reconnects the unconnected topology by randomly switching
could also indicate the impact of each successive step. The disadvantage on the elements with off-state until a completely connected network is
may be the risk of reaching only the local optimum. In this case, the best obtained. After that, the subsequent computations are performed for the
possible operation configuration may not be found. configurations, load flow, and SC. The fitness value of the SGA is pre­
sented by MIk. The values obtained by the load flow are used to verify
that the operational constraints are satisfied (section III.B). If not, the
4.2. Pymoo optimization framework algorithm calculates a “distance value” using the Euclidean norm to
determine how badly are the operational constraints violated for the
The Pymoo framework was successfully used for our optimization given configuration. All chromosomes then proceed to the crossover
task. The Pymoo is an open-source Python library shared under the step, where the parents are picked by the tournament method. The al­
Apache License 2.0 and offers various single- and multi-objective opti­ gorithm also detects duplicate configurations within each generation
mization algorithms [34]. The advantage is that Pymoo supports parallel which are consequently regenerated randomly to keep diversity in the
processing of the function evaluation. Thus, the optimization process is population. Such operation results in additional NOSC evaluations in the
accelerated on parallel computing architectures. This can help DSOs current generation, but the population size remains the same. It can be
optimize large meshed networks in an acceptable calculation time.
The Pymoo library offers a variety of optimization algorithms, such
as Differential Evolution, GA, Pattern Search and Covariance Matrix 1
https://Pymoo.org/algorithms/list.html#nb-algorithms-list.

6
D. Topolanek et al. International Journal of Electrical Power and Energy Systems 152 (2023) 109203

Table B1 Table B1 (continued )

Configuration of switches and MV/LV transformers in the initial, case A wining Index Switches MV/LV transformers
and case B winning state.
Initial case A case B Initial case A case B
Index Switches MV/LV transformers
72 1 1 0
Initial case A case B Initial case A case B 73 1 1 0
0 1 1 1 1 1 1 74 1 1 0
1 1 0 1 1 1 0 75 1 1 1
2 1 0 1 1 1 1 76 1 1 0
3 1 0 1 1 1 1 77 1 1 1
4 1 0 1 1 1 0 78 1 1 1
5 1 0 0 1 1 0
Note: The symbol 1 stands for a closed switch and a transformer in operation,
6 1 0 0 1 1 1
whereas the symbol 0 stands for an open switch and a transformer out of
7 1 0 0 1 1 0
8 1 0 0 1 1 0 operation.
9 1 0 1 1 1 0
10 1 0 1 1 1 0 interpreted in a way that the survival probability of duplicate configu­
11 1 0 0 1 1 1
ration is zero. The new population consists of:
12 1 0 0 1 1 0
13 1 0 1 1 1 0
14 1 0 1 1 1 1 • 10 % of the population size are chromosomes with the lowest fitness
15 1 0 1 1 1 1 value, which satisfies the operational constraints (if they exist).
16 1 0 0 1 1 1 • Up to 70 % (to 80% if there are less than 10% lowest values) of the
17 1 0 0 1 1 0
18 1 1 1 1 1 1
population chromosomes are created by a crossover with mutation
19 1 0 0 1 1 1 step. Parents are selected by the tournament method. Each parent is
20 1 1 1 1 1 0 selected from 8 competitors. The selected individual has the lowest
21 1 0 1 1 1 1 fitness values (if the operational constraints are satisfied) or it is
22 1 0 0 1 1 0
closer to the feasible set if both competitors violate the constraints.
23 1 0 0 1 1 0
24 1 0 0 1 1 1 The crossover is performed uniformly using the “flip a coin” tech­
25 1 0 0 1 1 0 nique. If there is a “head” for the given gene, it will be crossovered.
26 1 0 0 1 1 0 • 10 % of the population size is randomly regenerated to avoid local
27 1 0 0 1 1 0 extremes.
28 1 0 0 1 1 0
29 1 0 0 1 1 0
30 1 0 0 1 1 0 The rest 10 % of the population are chromosomes created by an
31 1 0 0 1 1 0 additional mutation step, in which the best configuration is randomly
32 1 1 1 1 1 1 reconfigured to perform variations around the best individual. At first, it
33 1 0 0 1 1 0
chooses if it will be turning on or off the elements. BCs and DTs sepa­
34 1 0 1 1 1 0
35 1 0 0 1 1 0 rately. Then, it picks as many indices as the number lesser or equal to
36 1 0 0 1 1 1 half of all zeros/ones in the chromosome. The zeros if it will be turning
37 1 0 0 1 1 1 on the elements, and the ones if it will be turning off the elements.
38 1 0 0 1 1 0
39 1 0 0 1 1 0
40 1 0 0 1 1 0
4.4. Hebo
41 1 0 0 1 1 0
42 1 0 0 1 1 0 Finally, the Heteroscedastic Evolutionary Bayesian Optimization
43 1 0 0 1 1 1 algorithm (HEBO) was used for our DN optimization tasks. The HEBO
44 1 0 0 1 1 0
algorithm won the black-box optimization challenge of NeurIPS 2020
45 1 0 0 1 1 1
46 1 0 0 1 1 1 and is described in more detail in [37]. According to the authors of [37],
47 1 0 0 1 1 1 the HEBO algorithm aims to be fast in terms of the number of evalua­
48 1 0 1 1 1 0 tions of the function f.
49 1 0 0 1 1 0 The HEBO algorithm integrated into the Ray Tune library [38] was
50 1 0 0 1 1 1
51 1 0 0 1 1 0
used in our numerical experiments. The Ray Tune is a Python library for
52 1 0 1 1 1 0 experiment execution and distributed hyperparameter tuning. The Ray
53 1 0 0 1 1 1 Tune provides users unique access to various optimization libraries.
54 1 0 0 1 1 0 Consequently, the user can easily access various search algorithms on
55 1 1 0
different networked computers (i.e., on distributed systems).
56 1 1 0
57 1 1 1
58 1 1 0 5. Results and method comparison
59 1 1 0
60 1 1 0 All the presented methods were applied to the meshed DN problem
61 1 1 1
62 1 1 0
previously introduced. For load flow and SC-current calculations, pan­
63 1 1 1 dapower package was used [39]. The SC currents were calculated ac­
64 1 1 1 cording to IEC 60909 standard [33]. The loads were set to the annual
65 1 1 0 maximum derived from the yearly load profile of the modelled network
66 1 1 1
and the initial state of selected switching components is set to switched-
67 1 1 1
68 1 1 0 on.
69 1 1 0 The methods must find a configuration of the network with mini­
70 1 1 1 mum value MIk while all following constraints must be satisfied:
71 1 1 1

• lines and DTs loading must be below or equal to 70 %,

7
D. Topolanek et al. International Journal of Electrical Power and Energy Systems 152 (2023) 109203

• voltage magnitude (per unit) of all nodes must be between 0.9 and the SGA performs lesser deviation in each iteration than the Pymoo
1.10, algorithm.
• all nodes must be supplied, Case A was designed only for a fundamental comparison of the
• no upstream active power flow must occur. methods because its solution in this setting is quite straightforward.
From the knowledge of the presented network operation, it can be
For measuring the computational complexity of individual methods, concluded that the minimum MIk is reached when all the BCs, whose
the required NOSC and total computational time are compared. Nu­ disconnection does not lead to supply interruption, are switched off.
merical experiments were performed on the laptop Acer Nitro 5, CPU However, even this simple solution can validate the correct imple­
Intel(R) Core(TM) i5-9300H CPU @ 2.40 GHz (4 CPUs) with 16 GB RAM mentation of applied optimization methods. To increase the complexity
and Linux Mint. For both study cases, A and B, the comparison of of the optimization problem and to further compare the performance of
methods performance is presented using convergence graphs (Figs. 4 the methods, case B was carried out.
and 5). For the evolutionary methods (SGA, Pymoo, HEBO), the algo­
rithms were run 10 times to evaluate their stochastic behaviour. 5.2. Results for case B

5.1. Results for case A For the SGA, and Pymoo, the following settings were considered:

For the SGA, and Pymoo, the following settings were considered: • population size is 1 000 (the best convergence rate within the
reasonable computation time),
• the algorithm stops if there is the same minimum value in 5 • maximum number of generations is 40 (acts as computation time
consecutive generations (based on several tests), constraint),
• the population size is 100 (based on several tests), • the algorithm can stop earlier if there is the same minimum value in 5
• the Pymoo version 0.6.0 with integer random sampling, tournament consecutive generations,
selection, simulated binary crossover, and polynomial mutation was • the Pymoo version 0.5.0 with integer random sampling, tournament
used (based on several tests). selection, simulated binary crossover, and polynomial mutation was
used (based on pre-tests).
For the HEBO algorithm of the Ray Tune, the following setting was
considered: For the HEBO algorithm of the Ray Tune, the following setting was
considered:
• the number of samples is 450.
• the number of samples is 2 000 (user restriction taking into account
Table 1 summarizes the results of the optimization methods for the allowable computing time of two days on the laptop).
simple case study A. All variables shown in Table 1 are based on median
values of ten runs (SGA, HEBO, Pymoo). For case B, the results are summarized in Table 2. For the SGA and
All methods found the best configuration of the tested network and Pymoo algorithms, 10 runs of each were performed. The median of the
have equivalent results of MIk, Over_SCCR and Ik_max, see Table 1. NOSC needed to obtain the solution is shown in Table 2. The most
However, the winning method utilizing the NOSC is HEBO, in which the promising results are highlighted in bold.
minimum value of MIk = 4.0510 kA was obtained using only 450 SC Considering the MIk values, the best-performing method in case B is
calculations. On the other hand, the computational time of HEBO was the Pymoo with MIk = 0.2589 kA. The Pymoo found here the minimum
one hour and thirty minutes on the laptop Acer Nitro 5, see Table 1, line that was lower than the minimums found by the other methods. The
Time (h:mm). The Pymoo found the same minimum value MIk = 4.0510 Pymoo method provides the most promising results not only for MIk, but
using 1 100 SC calculations (11 generations) within 4 min. In the also for Over_SCCR and Ik_max. The Pymoo version 0.5.0 has a slightly
shortest run of the Pymoo, the minimum was obtained after 10 gener­ better convergence rate than version 0.6.0 in this case. The second-
ations. The SGA needs at least 12 generations (4 min) with approxi­ lowest score MIk = 0.2869 kA with Over_SCCR = 15 was obtained by
mately 1 379 NOSC to terminate the process. In the shortest run, the the SGA algorithm. The third score MIk = 0.4528 kA with Over_SCCR =
solution was obtained after 10 generations using 1 141 NOSC. The DOA 30 was calculated using the HEBO method which required only 2 000 SC
finds the minimum using 1326 NOSC within 4 min. evaluations (the HEBO method was stopped by the user due to the
Further, it ought to be said that the computation time of the HEBO maximal computation time restriction of 2 days). The DOA method
method grows exponentially with the increasing number of evaluations needs 3 681 NOSC, producing the worst score. Moreover, only the local
(due to internal overhead). For example, 1 000 SC evaluations within the minimum of MIk was reached as can be seen in Fig. 5. In Fig. 5, also the
HEBO method, take 570 min. The computational time is increased by 25th and 75th percentile for the SGA and Pymoo are depicted by filled
factor 6 compared to the 450 SC calculations above. To compare HEBO area. Further, the lowest value of all 10 Pymoo’s runs was MIk = 0.2214
with GAs, 1000 NOSC in one iteration using GA methods takes kA with Over_SCCR = 15. The lowest value found by SGA within 10 runs
approximately only 3 min for both. was MIk = 0.1865 kA with Over_SCCR = 10.
The comparison of individual method convergence is depicted in Analysing the time perspective of the best-performing methods, case
Fig. 4, where the resulting score MIk is drawn concerning the NOSC. The B was much more computationally intensive than case A. Unfortunately,
graphs of the evolutionary algorithms are based on the median value despite the low NOSC and rapid convergence, the HEBO method was not
within each iteration. The x-axis minimum is set to 100 NOSC to enable able to solve case B in a reasonable time. Although the HEBO method
visual comparison of all the methods, since the generation-defined found a minimum value of MIk = 0.4528 kA by only 2000 SC evalua­
methods (Pymoo, SGA), require 100 NOSC to get the first result. tions, the computation time for HEBO was extremely long: it took two
Moreover, the values between the 25th and 75th percentile of 10 days and 8.5 h to deliver results. On the other hand, although DOA
repeated runs are depicted by filled area for each evolutionary method. required 3 681 SC calculations, the computation time of DOA was only
As could be seen, the advantage of DOA rests in the certainty of 11 min. The low number of required NOSC and the high speed of the
getting the same result in each algorithm execution, while in contrast, DOA method imply that the simple DOA provides the fastest approxi­
the evolutionary methods are stochastic (characterized by the variability mation of minimized MIk for case B.
of results). As shown in Fig. 4, it is apparent that all methods converged For case B, the speed of both analysed GA methods was almost
to the same result (global minimum) in all 10 repeated runs. Moreover, equivalent: Evaluations of 40 000 SC took approximately 2 h.

8
D. Topolanek et al. International Journal of Electrical Power and Energy Systems 152 (2023) 109203

In comparison to all analysed methods, HEBO estimates the mini­ relationships which may be considered as potential competing interests:
mum using a very low NOSC. This would be a great advantage in the case David Topolanek reports financial support was provided by Technology
of extreme optimization tasks over large DNs [38], for example for an Agency of the Czech Republic.
electrical network expansion planning problem [40,41], where the
calculation time of networks evaluation is several times higher than the Data availability
calculation time of one iteration of HEBO method.
The comparison of the convergence of optimization methods for case Data will be made available on request.
B is shown in Fig. 5. The axes are provided in the logarithmic scale. The
figure shows a very fast convergence of HEBO and DOA methods. The Acknowledgements
figure also shows that the DOA method could not find a minimum as low
as the SGA, or the Pymoo. The best choice seems to be the Pymoo This research work has been partially carried out in the Centre for
providing a lower deviation than the SGA. Research and Utilization of Renewable Energy (CVVOZE). Authors
The result of the SC currents optimization is not only SC level gratefully acknowledge financial support from the Technology Agency
reduction, but also SC level redistribution, which can be seen in the of the Czech Republic (project No. TK02030039) and the Ministry of
histogram of short-circuit currents (Fig. A-1, Appendix A). The impact of Education, Youth and Sports of the Czech Republic through the e-INFRA
the optimization on SC currents, bus voltages, active power losses, and CZ (ID:90140). This work was also supported by the Doctoral grant
relative loading is presented in figures of Appendix A, where case A and competition VSB - Technical University of Ostrava, reg. no. CZ.02.2.69/
B corresponds to the best found (winning) configuration. Individual 0.0/0.0/19 073/0016945 within the Operational Programme Research,
states of handling switches and MV/LV transformers, which were Development and Education, under project DGS/TEAM/2020-017
applied for the presented initial, case A and case B configuration, are “Smart Control System for Energy Flow Optimization and Manage­
listed in Table B-1 (Appendix B). ment in a Microgrid with V2H/V2G Technology”.

6. Conclusion Appendix A. . The impact of optimization on key parameters

Four optimization methods were applied to a unique large meshed The impact of realized optimization on SC current distribution, bus
DN model aiming to minimize SC currents, which is an important and voltages, relative loading of the lines and transformers is presented in
relevant aspect of operational safety in meshed systems. The numerical Fig. A1, Fig. A2 and Fig. A3 for the initial operation state, case A and
model was created based on a real meshed network by the authors and case B configurations corresponding to the best performing throughout
was provided to IEEE DataPort to be available for other researchers. For the tested methods. The network configuration of all these operational
solving such a large DN, the optimization methods were selected and states (initial, case A and case B) is defined in Appendix B. The total
modified to give an effective performance. The DOA was designed active power loss is 0.144 MW for initial state, 0.149 MW for case A and
especially for the SC-minimizing problem in the specific DN and the SGA 0.219 MW for case B respecting the total active power consumption in
was modified and configured to be suitable for solving various problems the network 15 MW.
in DNs. On top of these two developed methods, another two open-
source optimization frameworks were used for comparison [34,37]. Appendix B. . Configuration of switches and transformers in the
Two case studies, distinguishing the number of switching elements initial, case A winning and case B winning state
were carried out, both searching for the optimal network configuration.
Across both case studies, the SGA and Pymoo methods seem to have high Table B1
potential since both reach a good score in an acceptable time. The
expert-based DOA needs a very short computation time and a small References
amount of SC evaluations but with a considerable risk of reaching only a
local minimum as occurred in case B. On the other hand, the results of [1] Sahu SK, Ghosh D. Hosting capacity enhancement in distribution system in highly
trenchant photo-voltaic environment: a hardware in loop approach. IEEE Access
such quick approximation could be used as a starting point for the 2020;8:14440–51. https://doi.org/10.1109/ACCESS.2019.2962263.
evolutionary algorithms methods (SGA, Pymoo) to reduce the compu­ [2] Zou Z, Tang J, Liserre M, Wang Z, Cheng M. Modeling and stability analysis of
tational time. The HEBO method is the winner in both case studies (A meshed smart transformer-fed grid. In: CSEE Journal of Power and Energy Systems.
10.17775/CSEEJPES.2021.06600.
and B) in terms of convergence related to the NOSC, but the big disad­ [3] Wolter D, Zdrallek M, Stötzel M, Schacherer C, Mladenovic I, Biller M. Impact of
vantage of this method is its extremely long computation time which meshed grid topologies on distribution grid planning and operation. CIRED - Open
grows exponentially with the increasing number of SC evaluations (e.g., Access Proc J 2017;2017:2338–41. https://doi.org/10.1049/oap-cired.2017.0708.
[4] Foltyn L, Vysocký J, Prettico G, Běloch M, Praks P, Gianluca Fulli OPF, et al. OPF
1 000 SC evaluations take about 9.5 h; 2 000 SC evaluations took more solution for a real Czech urban meshed distribution network using a genetic
than 2 days on regular laptop). Therefore, HEBO is disqualified from the algorithm. Sustain Energy Grids Netw 2021;26:100437. https://doi.org/10.1016/j.
real application with such large DNs as the meshed DN presented here (1 segan.2021.100437. ISSN 2352-4677.
[5] Primadianto A, Lu C-N. A review on distribution system state estimation. IEEE
268 nodes).
Trans Power Syst Sept. 2017;32(5):3875–83. https://doi.org/10.1109/
Regarding the presented results, the recommendation for solving TPWRS.2016.2632156.
optimization problems in large meshed networks is to use the SGA, when [6] Hooshmand RRA, Soltani S. Fuzzy optimal phase balancing of radial and meshed
the solution speed is the priority (1:55 of SGA vs. 2:24 of Pymoo, see distribution networks using BF-PSO algorithm. IEEE Trans Power Syst Feb. 2012;27
(1):47–57. https://doi.org/10.1109/TPWRS.2011.2167991.
Table 2). The other feasible choice is Pymoo, which is the winner when [7] Deka D, Talukdar S, Chertkov M, Salapaka MV. Graphical models in meshed
the solution quality of Ik_max is the priority (46.95 kA of SGA vs. 44.45 kA distribution grids: topology estimation, change detection & limitations. IEEE Trans
of Pymoo, see Table 2). Concerning the computation time of HEBO, Smart Grid Sept. 2020;11(5):4299–310. https://doi.org/10.1109/
TSG.2020.2978541.
which is currently affected by an internal bottleneck of Bayesian opti­ [8] Vinnakota BV, Ababei C. Determination of the minimum breakpoint set of
mization, the use of HEBO would be reasonable only for networks, in directional relay networks based on k-trees of the network graphs. IEEE Trans
which the cost of one evaluation of steady state is very high (e.g., sys­ Power Deliv Oct. 2011;26(4):2318–23. https://doi.org/10.1109/
TPWRD.2011.2150249.
tems with an extremely high number of nodes). [9] Deng Y, Cai Le, Ni Y. Algorithm for improving the restorability of power supply in
distribution systems. IEEE Trans Power Deliv Oct. 2003;18(4):1497–502. https://
Declaration of Competing Interest doi.org/10.1109/TPWRD.2003.817805.

The authors declare the following financial interests/personal

9
D. Topolanek et al. International Journal of Electrical Power and Energy Systems 152 (2023) 109203

[10] Sokoler LE, Vinter P, Bærentsen R, Edlund K, Jørgensen JB. Contingency- programming. In: IEEE Transactions on Power Systems, vol. 31, no. 6; Nov. 2016.
constrained unit commitment in meshed isolated power systems. IEEE Trans Power p. 4776–83. 10.1109/TPWRS.2015.2507067.
Syst Sept. 2016;31(5):3516–26. https://doi.org/10.1109/TPWRS.2015.2485781. [26] Gavrilova A, Khalyasmaa A. Power system sectionalizing optimization using
[11] Liu X, Shahidehpour M, Li Z, Liu X, Cao Y, Bie Z. Microgrids for enhancing the genetic algorithm. In: 2019 16th Conference on Electrical Machines, Drives and
power grid resilience in extreme conditions. IEEE Trans Smart Grid March 2017;8 Power Systems (ELMA); 2019. p. 1–4. 10.1109/ELMA.2019.8771663.
(2):589–97. https://doi.org/10.1109/TSG.2016.2579999. [27] Wang H, Zhu L, Li Y, Zhang F, Liu D, Tang Y. Research on short-circuit current
[12] Li Z, Wu W, Tai X, Zhang B. A reliability-constrained expansion planning model for restraining method based on dynamic partitioning technology. In: 2017 IEEE
mesh distribution networks. IEEE Trans Power Syst March 2021;36(2):948–60. Manchester PowerTech; 2017. p. 1–5. 10.1109/PTC.2017.7981152.
https://doi.org/10.1109/TPWRS.2020.3015061. [28] Lizhong Z, Ye Y, Siyao L, Jingwei G. Research on short-circuit current suppression
[13] Yu L, Czarkowski D, de Leon F. Optimal distributed voltage regulation for based on equivalent sensitivity. In: 2020 International Conference on Urban
secondary networks with DGs. IEEE Trans Smart Grid June 2012;3(2):959–67. Engineering and Management Science (ICUEMS); 2020. p. 251–5. 10.1109/
https://doi.org/10.1109/TSG.2012.2190308. ICUEMS50872.2020.00062.
[14] Faiya BA et al. A Self Organizing Multi Agent System for Distributed Voltage [29] Jiaming W, Wenying L, Chang X, Jing L, Longzhen W. Study on optimization
Regulation. In: IEEE Transactions on Smart Grid. 10.1109/TSG.2021.3070783. strategy for short circuit current based on multilevel fuzzy comprehensive
[15] Lyu C, Jia Y, Xu Z. A novel communication-less approach to economic dispatch for evaluation model. In: 2011 4th International Conference on Electric Utility
microgrids. IEEE Trans Smart Grid Jan 2021;12(1):901–4. https://doi.org/ Deregulation and Restructuring and Power Technologies (DRPT); 2011. p.
10.1109/TSG.2020.3018622. 1350–54. 10.1109/DRPT.2011.5994106.
[16] Utkarsh K, Srinivasan D, Trivedi A, Zhang W, Reindl T. Distributed model- [30] Schneider K, Phanivong P, Lacroix. IEEE 342-node low voltage networked test
predictive real-time optimal operation of a network of smart microgrids. IEEE system. In: 2014 IEEE PES General Meeting | Conference & Exposition; 2014. p.
Trans Smart Grid May 2019;10(3):2833–45. https://doi.org/10.1109/ 1–5. 10.1109/PESGM.2014.6939794.
TSG.2018.2810897. [31] “Resources – IEEE PES Test Feeder.” Resources – IEEE PES Test Feeder, cmte.ieee.
[17] Amaris H, Molina YP, Alonso M, Luyo JE. Loss allocation in distribution networks org/pes-testfeeders/resources [Accessed 9 Feb. 2023].
based on Aumann-Shapley. IEEE Trans Power Syst Nov. 2018;33(6):6655–66. [32] Topolanek D, March 4. Meshed LV distribution network. IEEE Dataport 2022.
https://doi.org/10.1109/TPWRS.2018.2844740. https://doi.org/10.21227/kssw-1457.
[18] Sharma S, Abhyankar AR. Loss allocation for weakly meshed distribution system [33] IEC 60909-0:2016, Short-circuit currents in three-phase a.c. systems – Part 0:
using analytical formulation of shapley value. IEEE Trans Power Syst March 2017; Calculation of currents.
32(2):1369–77. https://doi.org/10.1109/TPWRS.2016.2571980. [34] Blank J, Deb K. Pymoo: multi-objective optimization in python. In: IEEE Access,
[19] Kumar H, Khatod DK. -value based approach for loss allocation in radial and vol. 8; 2020. p. 89497–509. 10.1109/ACCESS.2020.2990567.
weakly meshed distribution networks with distributed generation. In: IEEE [35] Whitley D. A genetic algorithm tutorial. Statistics and Computing; 1994. p. 65–85.
Transactions on Power Delivery. 10.1109/TPWRD.2021.3099207. 10.1007/BF00175354.
[20] Li Z, Wang S, Zheng X, de León F, Hong T. Dynamic demand response using [36] Foltyn L, Vysocký J, Prettico G, Běloch M, Praks P, Fulli G. OPF solution for a real
customer coupons considering multiple load aggregators to simultaneously achieve Czech urban meshed distribution network using a genetic algorithm. Sustain
efficiency and fairness. In: IEEE Transactions on Smart Grid, vol. 9, no. 4, p. Energy Grids Netw 2021;26. https://doi.org/10.1016/j.segan.2021.100437.
3112–21; July 2018. 10.1109/TSG.2016.2627140. [37] Cowen-Rivers A et al. HEBO: Heteroscedastic Evolutionary Bayesian Optimisation.
[21] Chen L, Huang M, Wu J, Wang D, Gan D. An optimal strategy for short circuit In: 34th Conference on Neural Information Processing Systems (NeurIPS 2020);
current limiter deployment. In: 2010 Asia-Pacific Power and Energy Engineering 2020.
Conference; 2010. p. 1-4. 10.1109/APPEEC.2010.5449483. [38] Wang Z, Liu Z, Kraiczy M, Bornhorst N, Wende-von Berg S, Braun M. Fast parallel
[22] Guo DH, Xiao XY. Optimal allocation of superconducting fault current limiter using quasi-static time series simulator for active distribution grid operation with
multi-objective estimation of distribution algorithm based on bacterial foraging. In: pandapower. In: CIRED 2021 - The 26th International Conference and Exhibition
2013 IEEE International Conference on Applied Superconductivity and on Electricity Distribution; 2021. p. 1602–6. 10.1049/icp.2021.1540.
Electromagnetic Devices; 2013. p. 250–53. 10.1109/ASEMD.2013.6780755. [39] Thurner L et al. Pandapower—an open-source python tool for convenient
[23] Mahmoudian A, Islam MR, Kouzani AZ, Parvez Mahmud MA. Optimal allocation of modeling, analysis, and optimization of electric power systems. In: IEEE
fault current limiter in distribution network with NSGA-II algorithm. In: 2020 IEEE Transactions on Power Systems, vol. 33, no. 6; Nov. 2018. p. 6510–21. 10.1109/
International Conference on Applied Superconductivity and Electromagnetic TPWRS.2018.2829021.
Devices (ASEMD); 2020. p. 1–2. 10.1109/ASEMD49065.2020.9276231. [40] Kisse JM, Braun M, Letzgus S, Kneiske TM. A GIS-based planning approach for
[24] Li D, Hao Z, Zhang Z, Wang Y. Optimization of fault current limiter configuration urban power and natural gas distribution grids with different heat pump scenarios.
based on improved niche genetic algorithm. In: 2019 IEEE 8th International Energies 2020;13:4052. https://doi.org/10.3390/en13164052.
Conference on Advanced Power System Automation and Protection (APAP); 2019. [41] Scheidler A, Thurner L, Braun M. Heuristic optimisation for automated distribution
p. 1461–64. 10.1109/APAP47170.2019.9224651. system planning in network integration studies. IET -Renew Power Gener 2018;12
[25] Yu P, Venkatesh B, Yazdani A, Singh BN. Optimal location and sizing of fault (5):530–8. https://doi.org/10.1049/iet-rpg.2017.0394.
current limiters in mesh networks using iterative mixed integer nonlinear

10