You are on page 1of 75

Reconfiguration of Distribution Systems to Reduce Power

Loss using Ant Colony Optimization and Simulated


Annealing

by
Daniel Barrios Castilla
Thesis of 60 ECTS credits submitted to the Department of Engineering

at Reykjavík University in partial fulfillment


of the requirements for the degree of
Master of Science (M.Sc.) in Electric Power Engineering
Supervisor:
Elias August
Assistant Prof., Reykjavík University, Iceland
Advisors:
Laurentiu Anton, Lecturer, Reykjavík University, Iceland,
Kári Hreinsson, Specialist in innovation and technological development, Veitur, Iceland

June 2021
2
Abstract

This study researches the optimal reconfiguration to minimize power losses of the test grid, the
CIGRE network, and a grid owned by the Reykjavik area utility company Veitur, called A3-A5.
It is implemented in Python, where the grid’s data is obtained and processed with Pandapower,
a software-based on python that allows characterization of distribution grids and running power
flows. The optimization is executed with two techniques: Simulated annealing (SA) and ant
colony optimization (ACO).
SA is a well know and widely used optimization technique for reconfiguration problems. It
is an implementation based on keeping radiality and minimizing power losses. Also, it evaluates
every option with the cooling schedule and finally accepts the best reconfiguration.
ACO works with a focus on processing the grid as a graph. The implementation is based
on graph theory that allows establishing a rule framework that assures the constraints and fast
execution and searches for the best reconfiguration option.
Reliability, power quality, and end-user are direct beneficiaries of the reconfiguration of the
distribution grids. In addition, reconfiguration is the first step for future implementations to
improve the grid.

1
A mi madre que siempre me ha insistido que tocar la puerta es no entrar, pero definitivamente
cuando se abre siempre es para tener nuevas experiencias y una oportunidad de transformar.
- Daniel Barrios Castilla.

2
Acknowledgements
I would like to thank my supervisor, Elias August, and my advisors, Kári Hreinsson and Lau-
rentiu Anton, for their valuable input, help and patience throughout the process. This thesis
wouldn’t be here without them.
In addition I want to thank my mother, father and sister for everything. I wouldn’t be here
without them.
Also my class mates, the dysfunctional squad, and all my new Icelandic friends. I wouldn’t
have had as much fun without them.

3
Contents

1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The Power Distribution System . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2.1 Topology of Power Distributed Network and Reconfiguration Network . . 3
1.2.2 Distribution Losses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.3 The Veitur A3-A5 Network . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Optimization Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.1 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.2 Thesis Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4 Simulated Annealing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4.2 Simulated Annealing Algorithm . . . . . . . . . . . . . . . . . . . . . . . 15
1.5 Ant Colony Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.6 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.6.1 ACO algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.7 Power Flow Analysis and Power Loss Estimation . . . . . . . . . . . . . . . . . . 23
1.7.1 Power Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.7.2 Pandapower Power Flow Method . . . . . . . . . . . . . . . . . . . . . . 25

2 Methodology 29
2.1 Base Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.1.1 The CIGRE Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.1.2 The Veitur Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2 Distribution System Reconfiguration . . . . . . . . . . . . . . . . . . . . . . . . 33
2.3 The Reconfiguration Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.1 Encoding Using Graph Theory . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.2 Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4
2.3.3 Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3.4 Incidence Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.3.5 Network Connectivity and Radiality Check . . . . . . . . . . . . . . . . . 38
2.4 SA Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.5 ACO Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3 Results 43
3.1 Computational Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2 Parameter Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2.1 Parameter Values SA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2.2 Parameter Values ACO . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.3 Changes in Power Losses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3.1 SA Results - the CIGRE Network . . . . . . . . . . . . . . . . . . . . . . 45
3.3.2 ACO Results - the CIGRE Network . . . . . . . . . . . . . . . . . . . . . 47
3.3.3 SA Results - The Veitur Network . . . . . . . . . . . . . . . . . . . . . . 48
3.3.4 ACO Results - the Veitur Network . . . . . . . . . . . . . . . . . . . . . 50
3.4 Changes in Voltage Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.5 Plotting the Veitur Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.6.1 Pandapower, Power Flow and Python . . . . . . . . . . . . . . . . . . . . 56
3.6.2 Simulated Annealing Versus Ant Colony Optimization . . . . . . . . . . 56
3.6.3 Reaching the Target Optimum . . . . . . . . . . . . . . . . . . . . . . . . 56
3.6.4 Future Work in Optimization of Distribution Network . . . . . . . . . . . 57

4 Conclusion 58
List of Figures

1.1 Medium voltage distribution system topology [4]. . . . . . . . . . . . . . . . . . 2


1.2 A sample 7-bus radial distribution network [9]. . . . . . . . . . . . . . . . . . . . 3
1.3 Meshed MV network based on a ring-main structure [10]. . . . . . . . . . . . . . 4
1.4 Types and factors affecting distribution losses [2]. . . . . . . . . . . . . . . . . . 8
1.5 The Veitur A3-A5 network blue highlighted (physical system). . . . . . . . . . . 9
1.6 Optimization methodologies of distribution network reconfiguration [25]. . . . . 11
1.7 A brief description of the main benefits and weakness [25]. . . . . . . . . . . . . 13
1.8 Aspects of search pattern during annealing process [35]. . . . . . . . . . . . . . . 17
1.9 Flowchart simulated annealing algorithm [36] . . . . . . . . . . . . . . . . . . . . 18
1.10 Experimental setup for the double bridge experiment. Branches have different
lengths [38]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.11 Flowchart of an ant colony system algorithm [40]. . . . . . . . . . . . . . . . . . 22
1.12 Pandapower uses PYPOWER to solve the power flow problem. . . . . . . . . . . 25
1.13 Line modeled with the π − equivalent circuit. . . . . . . . . . . . . . . . . . . . 26
1.14 Pandapower electric bus modeled. . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.15 Pandapower electric load modeled. . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.16 Pandapower electric switch modeled. . . . . . . . . . . . . . . . . . . . . . . . . 28

2.1 Medium voltage distribution network. . . . . . . . . . . . . . . . . . . . . . . . . 30


2.2 Initial configuration of the Veitur network. . . . . . . . . . . . . . . . . . . . . . 32
2.3 Power systems representations. (a) Single line diagram; (b) positive sequence
network diagram; (c) oriented connected graph [46]. . . . . . . . . . . . . . . . 36
2.4 Tree and cotree of the oriented connected graph [46]. . . . . . . . . . . . . . . . 37
2.5 Basic loops the oriented connected graph [46]. . . . . . . . . . . . . . . . . . . . 37
2.6 The element-node incidence matrix for the graph shown in Fig. 2.4 [46]. . . . . 38
2.7 IEEE 33-bus system divided by loops and described in the rule 1 2.3.5 [47]. . . . 39
2.8 Table with the common branch vectors described in rule 2 2.3.5 [47]. . . . . . . 39

6
2.9 Table with the prohibited group vectors and islanded principal interior nodes
2.3.5 [47]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.10 Edge selection of the ant in loop vector Lm [47]. . . . . . . . . . . . . . . . . . . 41

3.1 Optimal configuration - CIGRE Network by SA . . . . . . . . . . . . . . . . . . 46


3.2 Optimal configuration - the Veitur network using SA . . . . . . . . . . . . . . . 49
3.3 Optimal configuration - the Veitur Network by ACO. . . . . . . . . . . . . . . . 51
3.4 The CIGRE network voltage profile. Distribution network 15-bus before recon-
figuration and after reconfiguration by SA and ACO. . . . . . . . . . . . . . . . 53
3.5 The Veitur network voltage profile. Distribution network 188-bus before recon-
figuration and after reconfiguration by SA. . . . . . . . . . . . . . . . . . . . . . 54
3.6 Veitur network voltage profile. Distribution network 188-bus before reconfigura-
tion and after reconfiguration by ACO. . . . . . . . . . . . . . . . . . . . . . . . 55

7
List of Tables

1.1 Analogy between physical and simulated annealing. . . . . . . . . . . . . . . . . 16

2.1 Characterization of the CIGRE network. . . . . . . . . . . . . . . . . . . . . . . 29


2.2 Initial power flow of the CIGRE network. . . . . . . . . . . . . . . . . . . . . . . 31
2.3 Characterization the Veitur network. . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4 Initial power flow of the Veitur network. . . . . . . . . . . . . . . . . . . . . . . 33

3.1 Parameters of the proposed SA (the CIGRE network). . . . . . . . . . . . . . . 45


3.2 Parameters of the proposed ACO (the CIGRE network). . . . . . . . . . . . . . 47
3.3 Feasible individuals provided by ACO - the CIGRE network. . . . . . . . . . . . 48
3.4 Parameters of the proposed SA (the CIGRE network). . . . . . . . . . . . . . . 49
3.5 Parameters of the proposed ACO (the Veitur network). . . . . . . . . . . . . . . 50
3.6 Feasible individuals provided by ACO - the Veitur network. . . . . . . . . . . . . 52

8
Chapter 1

Introduction

1.1 Motivation
For the last 30 years, power distribution network reconfiguration has been solved with differ-
ent perspectives and objectives. Nowadays, digital transformation and electrical systems are
integrated and there is a trend to decentralized electrical systems. Networks require better re-
liability, energy efficiency, and service to the end-user. There is a possibility to apply better
and more complex optimization techniques to systems that are combined with renewable energy.
Several challenges in terms of load and power losses can be faced with reconfiguration because it
has been appropriately studied to bring a base of knowledge which is enough to propose feasible
solutions and assure reliability, power quality, and stable service.
Processing power losses is a continuous challenge because minimizing power losses is an issue
that has many approaches. However, in this thesis, the approach that was considered is to
minimize applied optimization methodologies that allow for the voltage profile to be improved
and keep parameters of the power distribution network in a stable range.

1.2 The Power Distribution System


The power distribution system lies between power generators and end-users. It is a part of a
complex and dynamic power delivery system. It starts with transmission lines carrying power by
means of high voltage, as that is the economical way. Later, the voltage is reduced and reaches
specific levels when it is distributed to end-users. Levels of voltage are classified accordingly [1],
[2].
Distribution is where most losses occur due to the low level of voltage and lack of efficiency in
the transportation. These power losses are costly for utility companies. It is possible to minimize

1
these losses by applying several methodologies [3]. This thesis explores network reconfiguration
through two different optimization techniques.

Figure 1.1: Medium voltage distribution system topology [4].

Figure 1.1 shows the structure of the distribution network medium voltage (MV) distribution
system. The power source was the primary sub-transmission HV system, 110 kV. The network
has two 110/10 kV transformers that are connected to bus I of station A, and their low sides
connect to bus III and IV of switching stations by overhead (OH) lines. There are three OH.
The first is three-sectionalized and single radial and the other two are three-sectionalized and
two-tie connections. The long OH line exemplifies a feeder in a rural area with moderate load
density and the other two exemplify feeders in urban districts with medium load density [4].

2
1.2.1 Topology of Power Distributed Network and Reconfiguration
Network

Radial Distributed Network

A radial distribution system is a standard often used in distribution system configurations.


There are radial systems that supply only one radial feeder (cable or overhead lines) but also
multi-radial systems that supply power to more than one radial feeder [5], [6].
In general, the radial distribution system has a structure with feeders that go from a substa-
tion to lateral feeders. From there a variety of end-users get service [7].
A radial distributed network is reliable because the radial structure brings advantages of low
current and operation control. However, the reliability is limited as the end-users can experience
more power outages than others. Thus, network reconfiguration has been repeatedly applied to
increase reliability. Network radiality applies to maintaining a power flow to all end-user loads
on a line. Reconfiguration of a power distribution network ensures network radiality [8].

Figure 1.2: A sample 7-bus radial distribution network [9].

Figure 1.3 shows a radial distribution network that has a common power source with end-user
loads coming off at single nodes along the line.

Meshed Distributed Network

A meshed structure is a ring-main network with all switches in a closed position. A ring-main
network is set connecting the line ends back to the feeding station. Often ring-main networks
are operated by open interruption points at defined spots on each line, which provides a simple
operation including a switchable reserve capacity, depending on the loading of the lines [10].
This configuration is characterized by:

• Clear and simple structure.

• Manageable operation under standard operating requirements.

• Loading of lines under typical serving conditions up to 70%.

3
• Reserve for outages of more than one line.

• Maintenance costs are high [10].

Figure 1.3: Meshed MV network based on a ring-main structure [10].

Distributed Network Reconfiguration

A distributed network is the biggest network in terms of length inside the electric power delivery
system and it is therefore responsible for a large part of power losses. Switches in different
branches within the distributed network allow the network to optimize power and keep the
voltage and current within desired limits. The strategy of playing with the switches and their
states (close and open) and find the best optimal configuration tree is part of operations research.
Distribution networks have a two-fold design. They are interconnected networks, but to
operate them they are arranged into a radial structure (feeder) with switches that can be either
closed or open. Using graph theory in this thesis, the distribution network will be defined by
two elements; a set of nodes N and a set of branches B. It is possible to have two types of
nodes, a source node directly connected to the supply transformer or a sink node that is simply
the end-user. Meanwhile, the branch is a part that can have two states; loaded or unloaded.
A network’s radial distribution is proved when, inside the feeder section, each consumer point
connects to only one source node [11].
Two types of switches are used in primary distribution systems. There are normally closed
switches (sectionalizing switches) and normally open switches (tie switches). These two types are
designed for both protection and configuration management. Because there are many switching

4
combinations in the distribution system, network reconfiguration is a complicated combinatorial,
non-differentiable constrained optimization problem [12].
In general, the methods of network reconfiguration can be categorized into three groups:
Heuristic, numerical, and metaheuristic-based approaches[13]. This thesis will explore two meth-
ods; SA and ACO.

Applications of Reconfiguration

The main focus will be on the problem of distribution network reconfiguration for power loss
reduction. It is possible to observe that the methodology applied for power loss reduction has
many similarities with reconfiguration for load balancing and restoration service.

Distribution Network Reconfiguration for Load Balancing

The main issue in load balancing is basically moving certain load currents from the busy feeder
or transformer to another with a light load through switch status in the distribution feeders.
Keeping load currents on the power distribution network provides advantages, such as optimal
use of transformers and feeders. Also, this type of reconfiguration must satisfy minimum power
loss maintenance in the system. Reliability and the stability of the system will be improved [14].
Network reconfiguration for load balancing is a combinatorial problem that must produce a fea-
sible distribution system structure and maintain radiality. Reconfiguration has been completed
when the balance of load currents is optimal and the system’s radial structure is not displaced
[15].
Power losses in distribution networks can mathematically be expressed as follows:

n
X P 2 + Q2
ri i 2 i (1.1)
i=1
|Vi |
where Pi is active power, Qi is reactive power, ri is resistance of branch i, n is the total number
of branches in the system, and Vi is voltage of branch [14]. Real power and reactive power are
directly related to real and reactive loads connected in the network. It is possible to formulate
this problem to express a load unbalance for main transformers and feeders in indices that will
be useful for measuring the system and branch loading.
System load balancing index (LBsys ) is defined as

nb
1 X Si
LBsys = (1.2)
nb i=1 Simax

5
where nb is the total number of branches, Simax is the apparent power rating of branch i, and
Si is the apparent power in branch i.
Branch load balancing index (LBi ) is define as

Si
LBi = (1.3)
Simax
The optimization problem will be completed when the system load and all branch load indices
are equal [14].

Distribution Network Reconfiguration for Restoration Service

Reconfiguration for restoration service is a solution to mitigate the risk associated with elements
such as protection malfunction, insulation aging, and others that can initiate components failure
or weather factors. Immediately after the fault the idea is to use the tools available such as
circuit breakers, sectionalizers, and open tie-switches for the reconfiguration process. Their goal
is to restore the isolated load as soon as possible. The system reconfiguration can be performed
in two cases for a healthy network. Network distribution system reconfiguration aims to reduce
losses, switch operation and optimize voltage profile. Reconfiguration during faulty conditions
is one of the most complex problems. It is possible to tackle the formulation with a single or
multi objective function [16].
The out of service area can be expressed as one of the following representations:
Minimization of the number of out-of-service buses (Nbus ):

X
min( Nbus ) (1.4)

Minimization of the number of out-of-service end-users (Ncust ):

X
min( Ncust ) (1.5)

Maximization the capacity of served loads (Pbus ):

X
min( Pbus ) (1.6)

The objective functions formulations are the multi-objective. The secondary objective is to
reduce switching operations and system losses.

6
1.2.2 Distribution Losses

Distribution losses are unavoidable outcomes of the process of distributing energy between zone
substations and end-users. Losses are usually critical in evaluating planning and operating
strategies. The distribution utilities are concerned with reducing the losses in the distribution
system according to the standard level. Various circumstances will influence the level of technical
and operational losses. For instance network configuration, load characteristics, substations in
service, and power quality required. It is essential to control these circumstances with proper
incentives and therefore optimize the level of losses. Losses in distribution networks can be split
into technical losses and non-technical losses.
Technical losses can be split into variable losses and fixed losses.
Variable losses (load losses) are equivalent to the square of the current. It depends
on the power distributed across the network. They are referred to as copper losses that mainly
befall in lines, cables, and copper parts of transformers. Variable losses can be decreased by;

• increasing the cross-sectional area of lines and cables for a given load;

• reconfiguring the network, for example, by providing more direct or shorter lines to where
demand is situated;

• managing the demand to reduce the peaks on the distribution network;

• balancing the loads on three-phase networks;

• encouraging the end-users to improve their power factors; and

• locating the embedded generating units as close as possible to demand.

Fixed losses (no-load losses) occur principally in the transformer’s core and deliver
heat and noise. A method to decrease fixed losses is to switch off transformers operating at low
demand. That depends on the network configuration that lets the operator switch any loads to
other sources in the distribution network.
Nontechnical losses (commercial losses) are closely related to issues of measuring,
billing errors, unauthorized abstraction of electricity, and unread meters. Utilizing electronic
meters will help reduce losses since the accuracy is high. The correction and proper identification
of the unauthorized abstraction of electricity will also help.
Reducing losses may become an added value to the cost of capital. On the other hand,
it will help reduce the amount of electricity production needed to satisfy demand, which will
have more large-scale advantages [2].

7
Figure 1.4: Types and factors affecting distribution losses [2].

1.2.3 The Veitur A3-A5 Network

"The A3-A5 grid is a mixture of residential and industrial load, it includes the harbor and
industry by Sæbraut but also the residential area around Laugardalur."

• In red are the 11 kV cables.

8
Figure 1.5: The Veitur A3-A5 network blue highlighted (physical system).

1.3 Optimization Techniques

1.3.1 Literature Review

Distribution Network Reconfiguration for Power Loss Minimization

The distribution network reconfiguration issue has been analyzed during the last thirty years
with several strategies. Modern techniques as genetic algorithms (GAs) and constrained decision
problems (CDPs) have been implemented. However, the fact is that neither CDP nor GA
can guarantee that the optimal solution will be achieved. They can achieve high-quality sub-
optimal solutions, but the network with more than 100 switches is typically found on a realistic
distribution system. Loss minimization is the main objective [17].
The optimal operating state of the distribution network is regularly considered to be achieved
when line losses are minimized, out of any violations of line loading and voltage limits. Other
service quality standards can be further used, service connection or voltage stability. Loss mini-
mization in network reconfiguration is achieved when the proper topology structure is selected.
Heuristic, preferably than analytical methods, seems to be more beneficial for feeder reconfigu-
ration studies [18]. Most optimization problems are simply aimed at getting the most reliable

9
radial configuration between a massive number of combinations, to which efficient procedures to
verify or enforce radiality are required. Due to the magnitude of the problem and its non-linear
nature, using a blend of optimization and heuristic techniques is justified, mainly if the recon-
figuration technique is intended for real-time application [19]. Early studies on reconfiguration
were directed at the planning stage. Smart grids use reconfiguration during the operation of
distribution grids [20], [21].
One of the early academic work published regarding the issue of reconfiguration is the algo-
rithm of Merlin and Back. It modeled the distribution system with a spanning tree structure.
It is also possible to consider the algorithm of Civalar et al. that made use solely of heuristics.
Both algorithms focused on reducing line losses [22].
A normally closed switch connects line sections. A normally open switch on the tie-lines
connects two primary feeders [23]. In this case, the objective is to ensure the network has
minimum real power losses subject to the following constraints:

• The branch flow limit.

• To ensure the system remain radial.

• The allowable limits of voltage [24].

The techniques that can be used to solve reconfiguration are artificial intelligence and heuris-
tics.
Both SA and ACO are heuristic techniques. A heuristic technique is an approximate method
to find a solution with no guarantees that it is globally optimal. To be valid this technique needs
to form a proper function for the problem.
The meta-heuristic method is an iterative generation process that assists to locate a near-
optimal solution efficiently. The strategy mainly merges different concepts that will be further
used and explores the search space. This method can be divided into two distinct categories:
Single (unique) and population solutions [25].
SA is a part of the single (unique) category. When the aim is to optimize a large-scale
simulation problem, it becomes significantly costlier to simulate all the alternatives to estimate
the objective function for each option accurately. Therefore, strategies are required to solve
optimization problems that minimize the simulation time. One of the strategies used for this
problem is a SA algorithm. SA is a random search technique to solve discrete optimization
problems when specific objective values exist. The algorithm consists of initialization, cooling
schedule, perturbation, and acceptance probability to perform the search [26].

10
Figure 1.6: Optimization methodologies of distribution network reconfiguration [25].

In [27], artificial ants have memorizing skill. The pheromone value represents the retention of
artificial ants. The parameter ρ represents the pheromone evaporation rate. The more significant
the parameter ρ, the faster the algorithm converges to the optimal solution. In contrast, the
smaller the pheromone evaporation rate ρ, the higher the randomness and global searching
capability of the algorithm, but the convergence rate drops. In this instance, to improve the
ability of the algorithm to search it uses the fuzzy rule that decides the value of the parameter
ρ.

Reconfiguration Using SA

The authors in [28], applied SA. The case is a merge between global search capability by using
a mechanism adopted from SA and feasibility-preserving with several functions mainly focused
on checking the network graph connectivity. Random modification of the network configuration
is achieved with preserving solution feasibility.

11
Reconfiguration Using ACO

In [29], ACO was the main algorithm applied to distribution system reconfiguration. The op-
timization problem is modeled as an undirected graph. The amount of pheromone dropped on
edges is modeled with a pheromone-based crossover operator. It provides an upgrade to the
balance between exploration and exploitation by allowing the algorithm to utilize knowledge
gathering during the search process to perform crossover. This methodology in the search tries
to avoid extreme search without a real improvement in the quality of the solution. It is fair to
assume that better solutions can be located in the neighborhood of existing reasonable solutions
without high exploration costs.
It is possible to see that the majority of variations in the ACO applications is mainly focused
on implementing techniques to improve the search solution, execution times, and the quality of
the solution.

12
Figure 1.7: A brief description of the main benefits and weakness [25].

1.3.2 Thesis Objective

The primary motivation behind this thesis is the study of the distribution network managed
by Veitur. The idea is to apply distribution network reconfiguration and minimize power
losses with optimization methods.The novelty in this thesis is, on the one hand, the modeling
of these two distribution grids and the estimation of the power losses using Pandapower and
taking advantage of the integration with python the implementation of these two optimization
techniques directly in the same environment.
These three specific objectives are necessary to obtain the main objectives described above;

• to implement an algorithm that allows optimizing the configuration of the power distribu-
tion network;

• to understand the power flows and apply this knowledge in simulations in Pandapower;
and

• to understand different ways to resolve the reconfiguration problem and compare the two
methods.

13
The optimization methods that was applied to the reconfigured network were SA and ACO.
SA is an algorithm that appropriates an iterative move according to the variable temperature,
which imitates the metals’ annealing transaction. ACO is based on observing how ants can find
the best path following pheromone produced by other ants.

1.4 Simulated Annealing

1.4.1 Overview

Kirkpatrick et al [30] introduced SA as a heuristic method to solve optimization problems. The


annealing procedure defines the optimal molecular arrangements of metal particles where the
potential energy of the mass is minimized and associates with cooling the metals gradually
after subjected to high heat. The SA algorithm adopts an iterative movement according to the
variable temperature parameter that imitates the metals’ annealing transaction.
The optimization algorithm’s critical point is that it has a neighborhood of different points
inside the configuration space that can generate different results. It iteratively compares results
and saves them as a base solution until the next iteration chooses the best one.[31].
There are three aspects of SA that are comparable to annealing:

1. SA cost function corresponds to the energy in annealing. The central role of the cost
function consists of evaluating every solution. The cost of the most recognized solution
usually decreases through the SA process. An optimization problem aims to find a solution
with minimum cost.

2. SA regularly perturbs the current solution so that different regions in the solution space
are explored. In annealing, the move of molecules matches to small perturbations in the
current solution.

3. The temperature T corresponds to a control parameter in SA. T controls the probability


of admitting a new solution that is worse than the current solution. T begins at the
maximum temperature, producing the current solution changes almost randomly at the
opening. T then continuously decreases so that more significant suboptimal perturbations
are rejected. The algorithm terminates meanwhile T equals a user-specified value [32].

SA algorithm has two iterative loops: The cooling process corresponds to the annealing
process and Metropolis criterion. Metropolis criterion is implemented to search the neighbor-
hood of the candidate solution additionally to avoid being trapped to local extreme points.

14
To be more precise, it considers the minimization problem and defines an objective function
f (xi ) corresponding to the argument set of xi = {x1 , x2 , ..., xn } with n ∈ R. Therefore, if
f (xi+1 ) < f (xi ), then take xi+1 as a new candidate extreme point to check. Otherwise, define
w = exp[−{f (xi+1 ) − f (xi )/Tc } where Tc is the current temperature parameter and generate a
random number s, such that 0 < s < 1. Then, if the relation of w > s is true, it also accepts
xi+1 as a new candidate, or else rejects and goes back to the previous step and generates another
s [33], [34].

1.4.2 Simulated Annealing Algorithm

An annealing process has a tendency to relax the physical system to a state with minimum free
energy. The process starts with heating a solid to a melting point, followed by cooling down
until it crystallizes into a state of a perfect lattice.
In the algorithm, a parameter called temperature is defined, which is equivalent to cost. The
temperature in a physical system is relatively similar to the freedom with which the system’s
entities can move around. The solid’s ground state has been reached if the maximum temperature
is sufficiently high and the cooling has been achieved relatively slowly. It will be close to the
below energy (or cost) configuration. A set of movements is chosen with the system that can
change from one state of configuration to another. Movements to reduce the system cost are
called downhill and movements to increase the system cost are called uphill. For the algorithm,
all downhill and uphill movements are accepted where probability is defined as:
∆C
p = e− T (1.7)

where ∆C is the increase in cost and T is the temperature. Firstly, the Boltzmann factor that
is a probability measure, or frequency distribution of particles in a system over various possible
states is calculated (1.7), is based on the Boltzmann distribution in statistical mechanics, and r
a random number uniformly distributed in the interval [0,1) is then determined. Secondly, the
∆C
uphill movement is accepted if r < e− T ; unless the movement is rejected and the configuration
before this movement is the next step. It means that uphill movements do not have a moderate
probability and raise the cost more that T allows. Therefore, there are two scenarios for the al-
gorithm. Firstly, at a very high temperature all the movements are accepted in the configuration
space. Secondly, at very low temperature only the downhill movements are accepted and simu-
lated annealing behaves similarly to a greedy algorithm. The desired global optimum is obtained
through the probabilistic rule that avoids being trapped in a local minimum. The feature above
is the main difference between simulated annealing and the greedy search method. A reasonable

15
pseudo-code of simulated annealing toward each temperature could be represented [35]

repeat

move
evaluate ∆C
accept/update

until stop criterion = true

Physical system Optimization problem


State Solution (configuration)
Energy Cost
Ground state Optimal solution
Rapid quenching Local search
Careful annealing Simulated annealing

Table 1.1: Analogy between physical and simulated annealing.

This analogy is important to clarify some aspects. The configuration, or state, is a set of
allowed system configurations that means the switches will be the configuration space. The
cost function is the power losses of every feasible reconfiguration: Inside the implementation of
SA, two-parameter keys are temperature and cooling schedule; these are not analogous to the
distribution network.

16
Figure 1.8: Aspects of search pattern during annealing process [35].

In theory, this method converges asymptotically to the global optimum solution with prob-
ability one, if particular on the cost function, the movements, and the annealing schedule are
satisfied [35].
The block diagram algorithm that was implemented,

17
Figure 1.9: Flowchart simulated annealing algorithm [36]

1.5 Ant Colony Optimization

1.6 Overview
ACO is a population-based heuristic technique developed by Doringo et al [37]. The foundation
originated from observing ants and their particular behavior to find food. When ants are trying
to find food they start exploration and the moment they find it they go back to the colony.
Other ants can follow the path of the initial ant through pheromone. If most ants decide to
follow the same path, the pheromone will be more potent there. The main characteristic of this

18
technique is that ants prefer to go on with a more robust pheromone path. The variation in the
density of the pheromone level defines the selecting probability of each path.
Applying this technique requires different mathematical procedures, such as a model that
shows the number of pheromones in the path. During the problem’s structuring, the agents
(ants) will be responsible for distributing a pheromone model by a stochastic approach and
building the paths and their probabilities [31].
The search area is modeled as a graph and the agents are described as moving points on this
graph. These agents’ movements on the graph produced a simulation version of the pheromone
secreting model. Each ant provides a solution and one path among the solutions should satisfy
the termination criterion to finish all procedures. Moreover, ACO algorithm fits well with parallel
implemented solutions which gives the algorithm a responsive character.

Figure 1.10: Experimental setup for the double bridge experiment. Branches have different
lengths [38].

Goss et al [38] developed a model of the observed behavior: Assuming that at a given moment
in time m1 ants have used the first bridge and m2 the second one, the probability p1 for an ant
to choose the first bridge is:

(m1 + k)h
p1 = (1.8)
(m1 + k)h + (m2 + k)h
where parameters k and h are to be fitted to the experimental data, clearly, p2 = 1 − p1 .

1.6.1 ACO algorithm

ACO is inspired by biology. Ants mark promising paths that other ants of the colony should
follow, with pheromone deposit on the ground. Ant colony optimization uses a mechanism based
on this behavior to solve optimization problems. Stigmergy plays a vital role due to the impor-
tance of communication between the agents (artificial ants). It is possible to define stigmergy
as a type of communication in which "workers are stimulated by the performance they have

19
achieved." Two main features of stigmergy are: The environment is the medium and the insects
exchange information to modify their environment. The exchange of information is only possi-
ble with local access, that is insect in the immediate neighborhood. The optimization technique
reflects the ants’ behavior through several artificial ants built for the considered optimization
problem. They exchange information based on the quality of these solutions using a commu-
nication scheme based on stigmergy. An adequate model is needed to apply the ant colony
algorithm [39]. A discrete optimization problem ϕ = (S, f, Ω) described as:

• A finite set C = {c1 , c2 , ..., cN c } of solution components.

• A finite set L = {li,j |(ci , cj ) ∈ C̃} of possible connections among the elements of C̃, which
is a subset of C ×C (the Cartesian product of C and C), where li,j indicates the connection
between components ci and cj . The number of possible connections is denoted by |L| ≤ Nc2 .

• A finite set Ω ≡ Ω(C, L) of constraints assigned over the elements of C and L.

• A finite set X of states of the problem, defined in terms of all possible sequences s =
hci , cj , ..., ck , ...i over the elements of C. The number of components in the sequence denoted
by |s|. The maximum length of a sequence is bounded by a positive constant number
N < +∞.

• A set X̃ of all states that are feasible with respect to the constraints Ω(C, L), X̃ ⊆ X.

• A set S = {s|s ∈ X̃} of (candidate) solutions, S ⊆ X.

• Ji,j is the connection cost associated with each li,j ∈ L.

• Js (L) is a cost associated with the solution s, which can be calculated from all the costs
Ji,j of all the components belonging to the solution s.

• A neighborhood structure of components: Given two components, c1 and c2 , the component


c2 is the component c1 if both components are in C, and the component c2 can be reached
from the component c1 in one logical step. The set of all neighbor components of the
component c is denoted by Nc .

• Each connection li,j is weighted by τi,j [39]

The pheromone model of ACO can be explained with the model of a combinatorial problem.
Every solution component is correlated to a pheromone value. It is possible to allocate each
value to a variable. An artificial ant develops a solution by crossing the completely connected

20
construction graph Gc (V, E), where V is a set of vertices and E is a set of edges. This graph has
two ways of reaching a solution: Artificial ants move from vertex to vertex along the graph’s
edges. Ants deposit a determined quantity of pheromone on the components, on the vertices, or
on the edges that ants traverse. The quantity ∆τ of pheromone deposited is directly related to
the solution’s quality. The following ants using this knowledge as a indicator to explore search
area further.
A variation Ant Colony System (ACS) algorithm includes the subsequent steps:

1. Initialise. The problem parameters are encoded as real numbers. Before each execution,
an initial population (nest) of the colony is generated randomly within the feasible region.

2. Evaluate. An evaluation of the fitness of all ants is based on their objective function.

3. Add trail. The ants have chosen appropriate directions concerning the ants’ fitness and
have added trail quantity to them.

4. Send ants. The objective function influences the ants’ performance that will be weighted
as fitness value which guides control to the level of trail quantity added to the appropriate
direction the ants have chosen. Every ant selects the next node to move based on two
parameters: The visibility of the node and the trail intensity.

5. Evaporate. Lastly, the pheromone discharged by an ant evaporates and the starting point
is updated with the best tour found [39].

The ACS algorithm is implemented as the main algorithm. The main difference for others
ACO algorithms is that a local pheromone update is introduced.
Concepts that are important to clarify: The ants navigate from nest to food source, which
means they are exploring the search space that is basically the possible configuration of switches.
To choose the next step probabilistically according to the pheromone model that is designed to
avoid being trapped in local optimal, if the pheromone has low evaporation it is a slow adaptation,
but if pheromone evaporation is high it means fast adaptation. To update pheromones is vital
for the feedback that an ant is going to receive.
The block diagram algorithm that was implemented:

21
Figure 1.11: Flowchart of an ant colony system algorithm [40].

22
1.7 Power Flow Analysis and Power Loss Estimation

1.7.1 Power Flow

Power flow is the name given to a network solution that shows currents, voltages, and real and
reactive power flow at every bus in the system which is implemented in Pandapower. The distri-
bution network has linear dynamics while the lines and transformers’ parameters are constants.
Nevertheless, the relationship between the voltage and current at every bus is nonlinear. The
same holds for the relationship between the real and reactive power consumption at a bus or
the generated real power and listed voltage magnitude at a generator bus [41]. Therefore the
solution of the nonlinear equations is involved in the calculation to get a power flow. Even power
flows are vital for the operation and planning of the network.
For a network with n buses, the power flow can be written with the following n equations:

Y11 V˙1 + Y12 V˙2 + · · · Y1n V˙n = I˙1

Y21 V˙1 + Y22 V˙2 + · · · Y2n V˙n = I˙2


(1.9)
···

Yn1 V˙1 + Yn2 V˙2 + · · · Ynn V˙n = I˙n


The matrix form is:    

Y Y12 · · · Y1n V˙1 I˙
 11     1
 Y21 Y22 · · · Y2n   V˙2   I˙2 
    
 .. .. ..   ..  =  ..  (1.10)
    
 . . .  .   . 
    
Yn1 Yn2 · · · Ynn V˙n I˙n
or
[Y ][V ] = I (1.11)

where I is the bus current injection vector, V is the bus voltage vector, Y is the bus admittance
matrix. Its diagonal element Yij is the self admittance of bus i, which equals the sum of all branch
admittances connecting to bus i. The off-diagonal buses i and j, this term is zero. Obviously,
the bus admittance matrix is a sparse matrix. In addition, the current can be represented by
bus voltage and power, that is

Ŝi SˆGi − SˆDi (PGi − PDi ) − j(QGi − QDi )


I˙i = = = (1.12)
V̂i V̂i V̂i
where,
S: The complex power injection vector.
PGi : The real power output of the generator connecting to bus i.

23
QGi : The reactive power output of the generator connecting to bus i.
PDi : The real power load connecting to bus i.
QDi : The reactive power load connecting to bus i.
Substituting equation (1.12) into equation (1.9), results in
(PGi − PDi ) − j(QGi − QDi )
= Yi1 V̇i + Yi2 V˙2 + · · · + Yin V˙n , i = 1, 2, . . . , n (1.13)
V̂i
The load demands are know variables. It defines the following bus power injection as

Pi = PGi − PDi (1.14)

Qi = QGi − QDi (1.15)

The general form of the power flow is obtained by substituting the above two equation in equation
(1.13)
n
Pi − jQi X
= Yij V˙j , i = 1, 2, . . . , n (1.16)
V̂i j=1
or
n
X
Pi + jQi = V̇i Yˆij Vˆj , i = 1, 2, . . . , n (1.17)
j=1

If equation (1.17) is divided into real and imaginary parts, two equations can be provided
for each bus with four variables real power P , reactive power Q, voltage V , and angle θ. To
solve the power flow equations, two of these variables must be known for each bus. Additionally,
there are three bus types:

• PV bus: The bus real power P and the magnitude of voltage V are known, and the bus
reactive power Q and the angle of voltage θ are unknown. Usually, the bus connected to
the generator is a PV bus and in a typical distribution networks there are no generators,
just an external grid.

• PQ bus: The bus real power P and reactive power Q are known, and the magnitude and
the angle of voltage (V , θ) are unknown. Usually, the bus connected to the load is a PQ
bus.

• Slack bus: This bus could be called swing or reference bus. It is necessary to have a bus
with a zero voltage angle as a reference for the other angle calculation. The variables
of voltage magnitude and angle (V, θ) are known, but the real power P and reactive
power Q are unknown variables. Usually, there is only one slack bus in the power flow
calculation [41].

24
1.7.2 Pandapower Power Flow Method

Pandapower is a tool for modeling, analysing, and optimizing electric power systems. Pan-
dapower was used for the modeling process of this thesis to execute the power flow and calculate
the power losses. Pandapower is Python-based, with free availability and adjustable expansion
with other open-source libraries. It has no real barriers for new users to perform a basic power
system analysis. It was originally designed for analysis of symmetric distribution systems. It is
evolving and is nowadays well-suited for the analysis of transmission systems.
The Pandapower power flow is solved using the Newton-Raphson method that has advantages
such as fast convergence as long as initial guess is close to solution and large region of convergence.
The Newton-Raphson power flow is used to determine the voltage magnitude and angle at each
bus in the power system that satisfies power balance. The internal power flow parameters for the
calculation, such as slack, PQ, and PV nodes or per unit conversion are automatically actualized
by Pandapower. The reference problem for the radial distribution grid is managed through the
external grid so that relative voltage angle shifts of transformers have no impact on the power
flow result. That leads to a faster and more robust convergence in a radial distribution grid.
Also, it is possible to choose an implementation backward or forward sweep, fast decoupled, and
Gauss-Seidel algorithms through an interface PYPOWER [42].

Figure 1.12: Pandapower uses PYPOWER to solve the power flow problem.

After the power flow is executed, the power loss estimation of the distribution system is
calculated mainly in the power losses on the lines, so the lines are modeled in Pandapower with
the pi-equivalent circuit:

25
Figure 1.13: Line modeled with the π − equivalent circuit.

Π-equivalent is utilized to accurately describe the line because it adequately contemplates


the shunt capacitance and the effect of the uniformly distributed parameters [43].
Pandapower also includes elements such as buses, switches, and loads in the power flow
calculation.
The electric bus model is,

26
Figure 1.14: Pandapower electric bus modeled.

Buses represent the nodes of the network. They are defined by a nominal voltage, which is
the reference voltage for the per-unit system [42].
The electric load model is,

Figure 1.15: Pandapower electric load modeled.

Loads are used to model electric consumption. The active power and reactive power define
them. The load model includes a scaling factor that allows scaling the load [42].
The switch model is,

27
Figure 1.16: Pandapower electric switch modeled.

The switch element allows the modeling of ideal switches. A switch element connects a bus
with an element that can be a second bus, a line, or a transformer. A closed bus-bus switch
galvanically connects two buses without a voltage drop. In-network calculation tools without
an explicit switch model, bus-bus switches can only be modeled as small impedance between
buses. It can point to undesired voltage drops and convergence problems in the power flow. The
Pandapower switch model avoids this problem by internally fusing buses connected by closes
bus-bus switches. Branches connected to a bus through an open switch are often modeled by
disabling the branch element. It means that the information about the switch position is lost
and the open-loop current of the branch element is neglected [42].
From the lines power flow (net.res_line), has the pl_mw that is active power losses of the
line [MW] define as:

p_f rom_mw = Re(v f rom × i∗f rom )

p_to_mw = Re(v to × i∗to )

pl_mw = p_f rom_mw + p_to_mw

The calculation for the power loss of the system by Pandapower is defined and based on
equation 1.1.

n
X
T otal P ower Loss = pl_mw (1.18)
i

where n is the total number of branches of the system and pl_mw is the active power loss
for the line.

28
Chapter 2

Methodology

2.1 Base Models

2.1.1 The CIGRE Network

Initial Conditions

The MV distribution network (CIGRE network) has the following characteristics:

Switches 8 elements
Load 18 elements
External grid 1 element
Line 15 elements
Transformer 2 elements
Bus 15 elements
Bus geodata 15 elements

Table 2.1: Characterization of the CIGRE network.

29
Figure 2.1: Medium voltage distribution network.

30
Vm [p.u.] Va [degree] P [MW] Q [MVAR]
1 1.03 0 -45.0457 -16.3414
2 0.9919 -6.5567 19.8390 4.6371
3 0.9681 -7.6041 0 0
4 0.9309 -9.3309 0.5017 0.2088
5 0.9290 -9.4349 0.4316 0.1081
6 0.9278 -9.5066 0.7275 0.1823
7 0.9263 -9.5917 0.5480 0.1373
8 0.9251 -9.5895 0.0765 0.0474
9 0.9254 -9.5789 0.5868 0.1470
10 0.9244 -9.6191 0.5737 0.3555
11 0.9231 -9.6854 0.5433 0.1612
12 0.9229 -9.6965 0.3298 0.0826
13 1.0001 -5.4871 20.01 4.6933
14 0.9953 -5.5382 0.034 0.0210
15 0.9925 -5.5679 0.5400 0.2577

Table 2.2: Initial power flow of the CIGRE network.

31
2.1.2 The Veitur Network

Initial Conditions

The Veitur A3-A5 network (the Veitur Network) belongs to the distribution network of the
utility company in Greater Reykjavík. It is characterized by:

Switches 188 elements


Load 87 elements
External grid 2 elements
Line 194 elements
Bus 188 elements

Table 2.3: Characterization the Veitur network.

Figure 2.2: Initial configuration of the Veitur network.

32
Vm [p.u.] Va [degree] P [MW] Q [MVAR]
0 1.000000 0.000000 -12.269944 -2.512995
1 1.000000 0.000000 -5.233295 -1.072495
2 0.998645 -0.030258 0.062720 0.012736
3 0.996950 -0.117416 0.031360 0.006368
4 0.996950 -0.117416 0.031360 0.006368
.. ... ... ... ...
183 0.993798 -0.145382 0.000000 0.000000
184 0.993777 -0.145902 0.000000 0.000000
185 0.993613 -0.147397 0.000000 0.000000
186 0.996950 -0.117416 0.000000 0.000000
187 0.996950 -0.117417 0.000000 0.000000

Table 2.4: Initial power flow of the Veitur network.

2.2 Distribution System Reconfiguration


The optimal reconfiguration problem can be used as a planning tool within power system op-
erations. The aim being a topology reconfiguration under certain conditions provided by the
distribution system. Distribution network reconfiguration (DNRC) presents nonlinearity and
discrete actions due to switches that modify a distribution network structurally [11].
A distribution system is comprised of feeders (electrical distribution circuits) that go radially
from a substation to an end-user. They are designed to serve all loads of the substations operation
area. Several tie switches and sectionalizing switches should consider a configurable system
because it depends on the possible topologies and maneuverability. Topology changes will define
redistribution of the loads in substations and feeder parts, affecting voltage profile and losses.
These changes are nonlinear. Consequently, the problem is reduced to find a radial topology
of a distribution network that ends in minimum losses as a nonlinear and combinatorial nature
problem [44].
Distribution networks are classified into subsystems of radial feeders, containing several nor-
mally closed switches and the number of normally open switches. According to graph theory,
a distribution network can be described with a G(N, B) that contains nodes N and a set of
branches B. Every node represents either a source node (supply transformer) or a sink node
(end-user load point), whereas a branch describes a feeder section that can either be loaded
(switch closed) or unloaded (switch open). The network is radial so that the feeder section form

33
a set of trees where each sink node is supplied from precisely one source node [11].
The mathematical model of DNRC is defined as:
XNL
minf = kl Rl Il2 l ∈ N L (2.1)
l=1
such that

kl |Il | ≤ Ilmax , l ∈ N L (2.2)

Vimin ≤ Vi ≤ Vimax , i ∈ N (2.3)

gi (I, k) = 0 (2.4)

gi (V, k) = 0 (2.5)

ϕ(k) = 0 (2.6)

where
Il : The current in branch l.
Rl : The resistance of branch l.
Vi : The node voltage at node i.
kl : The topological status of the branches – kl = 1 if the branch l is closed and kl = 0 if the
branch l is open.
N : The set of nodes.
N L : The set of branches.
In the above model, equation (2.2) stands for the branch current constraints. Equation (2.3)
stands for the node voltage constraints. Equation (2.4) represents Kirchhoff’s first law (KCL)
and equation (2.5) represents Kirchhoff’s second law (KVL). Equation (2.6) stands for topo-
logical constraints that ensure radial structure of each candidate topology. It consists of two
structural constraints:

• Feasibility: All nodes in the network must be connected by some branches, there is no
isolated node.

• Radiality: The number of branches in the network must be smaller than the number of
nodes by one unit (kl ∗ N L = N − 1).

Therefore, the final network configuration must be radial and all loads must remain connected.

2.3 The Reconfiguration Problem


The reconfiguration itself is a combination of steps such as encode, selection of a feasible switches
set, and a solution for the combinatorial problem. These steps will be described for SA and ACO.

34
2.3.1 Encoding Using Graph Theory

Encoding in the reconfiguration process is to get all the distribution network information so that
optimization techniques can get the entries necessary to provide a feasible solution.
The concept of connectivity can be defined based on graph theory as: "A graph is k-connected
if any two of its vertices can be joined by k independent paths" [45].
The network can be represented by either a reference bus or a loop. Reference bus is defined
as nodal voltages and nodal currents. Loop variables are defined as loop voltages and loop
currents.

2.3.2 Graph

In order to describe the geometrical structure of a network, it is adequate to replace the network
components with single line segments. A graph describes the network in a way that presents
the geometrical interconnection of the elements. A branch inside a graph is the line segment,
and the line segments’ terminals are the nodes. A subgraph is the subset of the analogous graph
elements, and a path is no more than two elements connected to each node. A graph is connected
if and only if there is a way between every two nodes. If each element of the connected graph
indicates a direction, it is then oriented [46].

35
Figure 2.3: Power systems representations. (a) Single line diagram; (b) positive sequence network
diagram; (c) oriented connected graph [46].

A representation of a power system and the corresponding oriented graph are shown in Figure
2.3.

2.3.3 Tree

The concept of no closed path defines a tree. Branches are elements in the tree, creating a subset
of the connected graph elements. The number of branches b required to form a tree is

b=n−1 (2.7)

where n is the number of nodes in the graph.


The links are those not involved in the tree and form a subgraph, not connected, called the
cotree. The complement of the tree is the cotree. The number of links l of a connected graph
with e elements is

l =e−b (2.8)

36
From equation 2.7 it follows that

l =e−n+1 (2.9)

It is called a loop if a link is added to the tree, as the resulting graph includes one closed
path. The path adds a subsequent link that forms one more additional loop and so on.

Figure 2.4: Tree and cotree of the oriented Figure 2.5: Basic loops the oriented connected
connected graph [46]. graph [46].

In figure 2.4 the tree and the corresponding cotree of the graph is given in Figure 2.3c. In
Figure 2.5 the basic loops of the graph is given in figure 2.3.

2.3.4 Incidence Matrix

An element-node incidence matrix explains the incidence of elements to nodes in a connected


graph. The elements of the matrix are as follows:

aij = 1 If the ith element is incident to and oriented away from the jth node.

aij = −1 if the ith element is incident to and oriented toward from the jth node.

aij = 0 If the ith element is not incident to the jth node.

The dimension of the matrix is e × n, where e is the number of elements and n is the number
of nodes in the graph.
The reference node could be whatever node that belongs to a connected graph. All the
variables of the other nodes are measured based on the reference. The matrix A obtained by
deleting the column corresponding to the reference node is the element-bus incidences matrix
A. This matrix dimension is e × (n − 1), and the rank is n − 1 = b, where b is the number of
branches in the graph.

37
Figure 2.6: The element-node incidence matrix for the graph shown in Fig. 2.4 [46].

2.3.5 Network Connectivity and Radiality Check

Loop and incidence matrix are constructive for checking and keeping two constraints of the
reconfiguration issue; radiality and connectivity.
The loop concept is applied in ACO. It is because it is used to generate only the feasible
topologies and avoid isolating the exterior and interior nodes following a set of rules generated
using graph theory. These rules are [47]:

• Rule 1: The mth member of the feasible reconfiguration must belong to loop vector Lm .

• Rule 2: Only one member from a common branch vector can be selected to form an
individual.

• Rule 3: All the common branch vectors of any prohibited group vector cannot simultane-
ously form an individual.

38
Figure 2.7: IEEE 33-bus system divided by loops and described in the rule 1 2.3.5 [47].

Figure 2.8: Table with the common branch vectors described in rule 2 2.3.5 [47].

Figure 2.9: Table with the prohibited group vectors and islanded principal interior nodes
2.3.5 [47].

These rules are mainly focused on keeping connectivity inside the radial topology. These
rules and their application in the solution were explained in 1.6.1.
The incidence matrix plays a vital role in checking the two constraints; radiality and connec-
tivity [48]. Radiality is checked through the determinant of the matrix A with dimension based

39
on the network elements and branches and its reference node. The network topology is radial if
the determinant is one, and the network is not radial if it is zero [49]. The connectivity is checked
through the matrix’s elements. There exists a relation if the element incident in the node is an
indicator of connection. The details were described in the section Simulated Annealing 1.4.2.

2.4 SA Implementation
The optimal, or near-optimal, solution can be found by understanding the SA algorithm [50].
Step 1: Set i = 0 and define adequately high temperature, cooling schedule, and initial
network reconfiguration.
Step 2: Initialize feasible states of switches xi and calculate the associated interruption cost
E(xi ). Feasible states can be found of a current configuration.
Step 3: Generate new states of switches xi+1 . If the new configuration satisfies the two
system constraints, a new interruption cost E(xi+1) is calculated.
Step 4: Perform an acceptance test for the new solution in step 4. If ∆E = E(xi+1 )−E(xi ) <
0, the new interruption cost is accepted and skip to step 6, otherwise go to step 5.
Step 5: Generate a uniform random number in the range [0, 1] and calculate the probability
of acceptance p = exp(−∆E/kT ). If r < p, the interruption cost obtained in step 5 is accepted
and proceed to step 5; if not, return to step 2.
Step 6: Decrease the temperature in the next iteration by setting Ti+1 = αT where α is the
cooling schedule and 0 < α < 1.
Step 7: Terminate the calculation process if T Tmin where Tmin is minimum specified tem-
perature; otherwise k = k + 1 and repeat steps 1-6 [50].

2.5 ACO Implementation


The first step in the development of a solution is encoding using graph theory described in
section 2.3.5. The second step is selecting an edge when ant k is in the switch position i to all
other switch positions in the loop vector is evaluated using the probability:

α )(η β )

(τi,j i,j
 P α β
(τi,j )(ηi,j )
i, j ∈ Lm
Pi,j = (2.10)
0 otherwise

where Lm is the loop vector, τi,j is the amount of pheromone on edge i − j. The parameters α
and β manage the corresponding importance of pheromone versus the heuristic value (α, β > 0).

40
Figure 2.10: Edge selection of the ant in loop vector Lm [47].

The third step, the heuristic information ηij or desirability of edge i − j is usually defined
1
by dij
; dij is the distance between node i and j. One of an algorithm’s main characteristics is
that the greedy heuristic eases to get acceptable solutions in the early stages of the exploration
process. For the reconfiguration problem, the heuristic information is redefined to send each
line’s end voltage for the best radial network. Sending end voltage will decrease it as its distance
from the source progresses. The matrix Dm is the desirability matrix that describes the selection
switch’s desirability switch j from switch i. Dmij is represented as

 V − V , i 6= j
i j
Dmi,j = (2.11)
 λV , i = j
i,j

where λ is a constant multiplier.


The fourth step is the pheromone update ∆τi,j . It is applied when an ant moves from switch
i to the switch position j. The pheromone is updated using

τi,j = (1 − ρ)τi,j + ∆τi,j (2.12)

where τi,j is the amount of pheromone on a given edge i − j, ρ is the rate of pheromone
evaporation and ∆τi,j is the amount of pheromone deposited, typically given by

1
k

Pk
, if ant k travels on edge i-j
∆τi,j = (2.13)
 0, otherwise
where Pk is real power loss of the radial configuration obtained by the k th ant’s travel.

41
The fifth step is the global update when ants communicate the best food location with other
ants. To implement that behavior extra pheromone is summed for the edge from a specific switch
in the loop vector to the switch of the corresponding loop vector of global best solution. For the
loop vector Lm

τi,j = τi,j + σ∆τi,j (2.14)

where i ∈ Lm and j is the switch position of the global best solution in Lm AND σ is a
constant multiplier.

42
Chapter 3

Results

3.1 Computational Time


SA and ACO belong to the metaheuristic algorithms. SA provides a unique feasible reconfigu-
ration every iteration instead of a population of feasible solutions in ACO.
The Veitur network has 194 lines and seven switches must be opened to fulfill the radiality
constraint. The CIGRE network has 15 lines and three switches must be opened to fulfill the
radiality constraints.
Taking into account the parameters of every algorithm, the processing time with Intel(R)
Core(TM) i5-8250U CPU - 1.60 GHz 1.80 GHz as a processor, and the conditions of every
network, the computational times are:

• Average computational time for the CIGRE network reconfiguration solving with SA al-
gorithm is 2.5 h.

• Average computational time for the CIGRE network reconfiguration solving with ACO is
0.17 h.

• Average computational time for the Veitur network reconfiguration solving with SA algo-
rithm is 6 h.

• Average computational time for the Veitur network reconfiguration solving with ACO is
2.5 h.

43
3.2 Parameter Values
The execution of the SA and ACO algorithms depends on the variety of the values specified to
a set of parameters that direct the iterative process.

3.2.1 Parameter Values SA

Tuning up the SA algorithm requires designating several parameters:

• Temperature: Taking as premise the initial temperature should be chosen high enough
that the initial acceptance allows the configuration with higher power losses. In [50] the
temperature adapted was 5000, then this parameter is set for the CIGRE and Veitur A3-A5
networks respectively. For Veitur A3-A5 the values that were set applied hit-trial method
5000, 20000, 30000, 40000, 50000, 60000. the best result was obtained with the temperature
of 60000. For CIGRE network the values that were set applied hit-trial method 5000, 4000,
3500, 3000, and 2500. The best result was obtained with the temperature of 2500.

• The cooling rate: It is set between 0.85 and 0.99. It is adjusted to a higher value in order
to avoid staying at a local optimal. The value that was set in all simulation for both
networks was 0.95.

3.2.2 Parameter Values ACO

It was set as reference for the parameters values in [29]. The start point for the values are: 1,
0.5, 0.1, and 2.5 for α,β,ρ,and σ respectively. Tuning up the SA algorithm requires designating
several parameters:

• α: the distribution of pheromone trail is highly converged, then α will influence ACO
to intensify towards past ants, otherwise ACO will perform diversification. For CIGRE
network the value for α is 1 . For Veitur A3-A5 network the value for α is 1. That means
is keeping same value as the start point for both netwrok.

• β: the distribution of heuristic information is highly converged, then β will influence ACO
to intensify towards heuristic information, otherwise ACO will perform diversification. For
CIGRE network the value for β is 3. For Veitur A3-A5 network the value for β is 2. That
means for both networks was selected larger values for the heuristic influence with the
purpose to have a faster results.

44
• ρ:Evaporation rate determines how important the new information gathered by ants in
newer iterations.For CIGRE network the value for ρ is 0.3. For Veitur A3-A5 network the
value for ρ is 0.3. That means for both networks was selected larger values for allowing
to every branch in every loop that has certain amount of pheromone every iteration and
allow to choose the indicate branch for opening it.

• σ: This parameter allows to identify what is the best branch for opening and is based on
ants communication of bets food location. For both networks were same value as in the
reference point.

3.3 Changes in Power Losses

3.3.1 SA Results - the CIGRE Network

The effectiveness of the SA algorithm is tested on the balanced CIGRE distribution network.
The CIGRE distribution network has the following specifications:

• Initial configuration: 12 - 13 - 14.

• Real power losses: 0.2337 [MW].

• Minimum node voltage (pu): 0.9229.

The proposed SA is implemented. The parameters are adjusted for their optimal values by
hit and trial and the final values are:

Temperature 2,500
Cooldown factor 0.95
Number max iterations 5,000

Table 3.1: Parameters of the proposed SA (the CIGRE network).

The CIGRE distribution network has the following specifications for the optimal configura-
tion:

• Initial configuration: 9 - 7 - 4.

• Real power losses: 0.1510 [MW].

• Minimum node voltage (pu): 0.9621.

45
Figure 3.1: Optimal configuration - CIGRE Network by SA

The optimal set switches that must be opened for the CIGRE reconfiguration of the dis-
tribution network obtained by the SA method is 9-7-4, which are the lines between buses 3-8,
buses 9-10, and buses 5-6. The real power loss of the ideal reconfiguration improved by 35%,
from 0.2337 M W to 0.1510 M W , compared to the initial case. The minimum voltage was raised
by 3.9% from 0.9229pu to 0.9620pu. The average voltage of 15 buses improved by 1.8% from
0.9542pu to 0.9728pu.

46
3.3.2 ACO Results - the CIGRE Network

The effectiveness of the ACO algorithm is tested on the balanced CIGRE distribution network.
The CIGRE distribution network has the following specifications:

• Initial configuration: 12 - 13 - 14.

• Real power losses: 0.2337 [MW].

• Minimum node voltage (pu): 0.9229.

The proposed ACO is implemented. The parameters are adjusted for their optimal values
by hit and trial and the final values are:

ρ 0.3
α 1
β 3
σ 2.5
Number max iterations 30
Number ants 10

Table 3.2: Parameters of the proposed ACO (the CIGRE network).

The CIGRE distribution network has the following specifications for the optimal configura-
tion:

• Initial configuration: 7 - 4 - 9.

• Real power losses: 0.1510 [MW].

• Minimum node voltage (pu): 0.9621.

In this case both methods have the same results for the optimal configuration.

47
Optimal Configuration Power Losees [MW]
1 7-4-9 0.1511
2 8-5-9 0.1517
3 8 - 12 - 9 0.1520
4 7 - 12 -9 0.1546
5 7-5-9 0.1569
6 6-4-9 0.1579
7 13 - 5 - 9 0.1582
8 6-3-9 0.1592
9 13 - 12 - 9 0.1601
10 7 - 3 - 9 0.1808
11 7 - 12 - 14 0.2294
12 7 - 5 - 14 0.2299
13 6 - 4 - 14 0.2311

Table 3.3: Feasible individuals provided by ACO - the CIGRE network.

The ACO simulations generated 13 different feasible results where power losses were under
the power losses of the initial configuration. The best one is the same optimal configuration
provided by SA. In addition, these 13 feasible configurations are 86% of total simulations, so
14% of the simulations are not feasible as they do not fulfill any of the constraints.

3.3.3 SA Results - The Veitur Network

Reaching the Target Optimum

The effectiveness of the SA algorithm is tested on the balanced Veitur distribution network.
The Veitur distribution network has the following specifications:

• Initial configuration: 12 - 24 - 64 - 150 - 99 - 160 - 127 - 110 - 55 - 57.

• Real power losses: 0.0749 [MW].

• Minimum node voltage (pu): 0.9916.

The proposed SA is implemented. The parameters are adjusted for their optimal values by
hit and trial and the final values are:

48
Temperature 60,000
Cooldown factor 0.95
Number max iterations 30,000

Table 3.4: Parameters of the proposed SA (the CIGRE network).

The CIGRE distribution network has the following specifications for the optimal configura-
tion:

• Initial configuration: 176 - 160 - 148 - 127 - 96 - 52 - 16.

• Real power losses: 0.0801 [MW].

• Minimum node voltage (pu): 0.9907.

Figure 3.2: Optimal configuration - the Veitur network using SA

The optimal set switches that must be opened for the Veitur reconfiguration of distribution
network obtained by the SA method is 176-160-148-127-96-52-16, which are the lines between
buses 76-77, buses 168-5, buses 6-162, buses 65-66, buses 50-137, buses 33-113, and buses 21- 22

49
respectively. The real power loss of the ideal reconfiguration deteriorated by 9.2%, from 0.0749
M W to 0.0801 M W , compared to the initial reconfiguration. The minimum voltage was lowered
by 0.12% from 0.9916pu to 0.9927pu. The average voltage of 188 buses improved by 0.06% from
0.9961 to 0.9955.

3.3.4 ACO Results - the Veitur Network

The effectiveness of the ACO algorithm is tested on the balanced Veitur distribution network.
The Veitur distribution network has the following specifications:

• Initial configuration: 12 - 24 - 64 - 150 - 99 - 160 - 127 - 110 - 55 - 57.

• Real power losses: 0.0749 [MW].

• Minimum node voltage (pu): 0.9916.

The proposed ACO is implemented. The parameters are adjusted for their optimal values
by hit and trial and the final values are:

ρ 0.3
α 1
β 2
σ 2.5
Number max iterations 1,000
Number ants 50

Table 3.5: Parameters of the proposed ACO (the Veitur network).

The Veitur distribution network has the following specifications for the optimal configuration:

• Initial configuration: 18 - 191 - 97 - 155 - 175 - 127 - 55.

• Real power losses: 0.0730 [MW].

• Minimum node voltage (pu): 0.9907.

50
Figure 3.3: Optimal configuration - the Veitur Network by ACO.

The optimal set switches that must be opened for the Veitur reconfiguration of distribution
network obtained by the ACO method is 18-191-97-155-175-127-55, which are the lines between
buses 23-24, buses 3-186, buses 137-51, buses 69-165, buses 75-76, buses 65-66, and 114-31
respectively. The real power loss of the ideal reconfiguration improved by 2.5%, from 0.0749
M W to 0.0730 M W , compared to the initial case. The minimum voltage was lowered by 0.09%
from 0.9916 to 0.9907pu. The average voltage of 188 buses improved by 0.06% from 0.9961 to
0.9955.

51
Optimal Configuration Power Losees [MW]
1 18 -191 - 97 - 155 - 175 - 127 - 55 0.07307
2 18 - 148 - 96 - 147 - 176 - 126 - 56 0.07330
3 18 - 148 - 97 - 146 - 176 - 125 - 56 0.07330
4 18 - 149 - 96 - 147 - 176 - 125 - 56 0.07330
5 18 - 191 - 96 - 147 - 176 - 126 - 57 0.07337
6 18 - 192 - 96 - 147 - 176 - 125 -57 0.07337
7 18 - 148 - 97 - 145 - 176 - 126 - 54 0.07341
8 18 - 191 - 97 - 146 - 176 - 125 - 55 0.07347
9 18 - 193 - 97 - 147 - 176 - 127 - 54 0.07349
10 18 - 191 - 96 - 146 - 177 - 126 - 56 0.07361

Table 3.6: Feasible individuals provided by ACO - the Veitur network.

Table 3.6 shows the top ten results with ACO of different feasible individuals with power losses
that are under the power losses of the initial configuration. In addition,the feasible configurations
are 61% of total simulation, so 39% of the simulations are not feasible as they do not fulfill any
of the constraints.

3.4 Changes in Voltage Profile


One of the constraints involved in the optimization problem is to keep the bus’s voltage in a
range between 0.95 to 1.05.
The CIGRE network voltage profile (figure 3.4) has two peaks close to 1 p.u exactly in bus
one and twelve because they are the buses connected to the external grid. Then, the proposed
reconfiguration has bus’s voltages between 0.95 p.u to 1.05 p.u, which is one of the constraints
of the optimization problem.
The Veitur network voltage profile is represented with a "dumbbell plot" that provides com-
parison of two values. The voltage before reconfiguration (navy circle) and voltage after recon-
figuration (gold circle). The main aim is that most voltages after reconfiguration are on the left
side or close to 1 p.u.

52
Figure 3.4: The CIGRE network voltage profile. Distribution network 15-bus before reconfigu-
ration and after reconfiguration by SA and ACO.

53
Figure 3.5: The Veitur network voltage profile. Distribution network 188-bus before reconfigu-
ration and after reconfiguration by SA.

54
Figure 3.6: Veitur network voltage profile. Distribution network 188-bus before reconfiguration
and after reconfiguration by ACO.

3.5 Plotting the Veitur Network


One of the objectives in terms of the Veitur network was elaborating a diagram with Pandapower.
The diagrams of the network are below:

• The Veitur network initial configuration. Veitur network with open switches. The
dash-lines represent the open switches.

• The Veitur network initial configuration without switches. Veitur network without
open switches.

55
3.6 Discussion

3.6.1 Pandapower, Power Flow and Python

The power flow is a mandatory entry for the implementation of SA and ACO. The role of
Pandapower is essential as it provides the execution of the power flow of every reconfiguration
candidate. It is executed faster and it is an advantage that Pandapower is a python-based tool.
All this data can be easily used for the rest of the implementation of the optimization algorithms.
Most academic examples or works related are generally executed with software with any cost
associated or specialized tools. For that reason, Python and Pandapower are novel because all
the implementations are in an open-source programming language and framework that is easy
to scale to other similar problems.

3.6.2 Simulated Annealing Versus Ant Colony Optimization

Both programs ran successfully and were implemented in Python. SA has less structure flexibility
than ACO. For that reason the assessment of every possible candidate by SA makes it slower
than ACO, but SA has the advantage that it can reach the global optimum. SA required more
pre-treatment of the information of branches and voltages because all information comes from
the original structure.
ACO implementation has a structural advantage because the network is processed as a graph
that is a strategy with a main purpose to divide the networks into loops with the objective
to assure the optimization constraints through the rules implemented to keep radiality and
assurance of the network’s connection. The generation of possible reconfigurations is another
advantage. After choosing the better candidate, every iteration, the pheromone modifies the
probability of a successful new reconfiguration process faster than SA. Nevertheless, it is not
always possible because the result could be a local minimum even no under the initial power
losses.

3.6.3 Reaching the Target Optimum

The implementation of SA has an iteration system that compares the losses power of the distri-
bution network for each individual. The first critical point is the number of feasible individuals
that need to compare that increase as the number of nodes of each network. The two networks
that have been assessed have a different sizes and a different number of feasible individuals.
SA execution works appropriately for finding an optimal configuration in the CIGRE network.

56
Nevertheless, the Veitur A3-A5 network had not a positive result from the SA method. During
the execution, the SA method was not possible to get a feasible individual. It was running the
algorithm with a considerable amount of iterations. However, any of the execution was success-
ful. Some of them provided a feasible individual with power losses more significant than the
power losses of the original configuration, and other executions were not finished. As primary
evidence during this implementation was that is not a direct way to estimate the number of iter-
ations connected with the size of the network. In addition, this size is critical for the successful
application of SA.
The parameters and proper tuning of this for different contexts are vital. It is possible to
say that for the Veitur network it was not possible to get a proper tuning of the temperature
and cooling schedule. These parameters are crucial for getting a feasible result.
These methodologies need to be reconsidered or changed when it comes to real distribution
networks where the number of buses and branches are considerably higher than 100. ACO
provided an optimum feasible individual in the case of Veitur. This thesis could be a good
beginning if there is interest to continue researching the topic because it has already established a
framework to understand and pre-process the network and methodology to implement algorithms
based on population.

3.6.4 Future Work in Optimization of Distribution Network

It is possible to divide future work into three parts. First, because both the networks, that
were the objective of the study, do not have any previous academic reconfiguration results, it is
possible to consider this work as an initial point for further work on these networks and future
validations of the work and results provided so far.
Second, the Veitur network is currently operating and providing service. This project can
be an initial point to improve the estimation of the power losses because the company today
considers the power losses under the comparison of energy-in versus energy-out. The reconfigu-
ration proposed is the initial step for future applications. It is turning this reconfiguration as a
base step in load balancing or restoration service in real-time application.
Third, the industry demands better planning and operating of future and current infrastruc-
ture. This kind of solution can explain the main elements inside the networks and the areas that
need improvement. This work could improve base knowledge for these networks and motivation
to add more information.

57
Chapter 4

Conclusion

Reconfiguration of distribution networks has three key points. Firstly, encoding of pre-treatment
is achieved during the implementation of two different encoding techniques; SA and ACO. The
process of encoding in SA ensures the radiality and the connectivity check of the network. They
guarantee the determinant of the matrix for every possible reconfiguration. The encoding process
is adequate, but it increases time compared to ACO. The ACO encoding is implemented with
rules that are applied to the loops in the grid. It is possible to say that encoding with a rule
framework is straightforward, easy to implement, and fast to execute.
Secondly, the heuristic factor has a significant impact that is reflected in the final result,
which is the best reconfiguration with minimum power losses. After encoding and evaluating
power losses of every possible reconfiguration with SA, the algorithm starts a comparison of each
movement that could be uphill or downhill. Due to the nature of the method, this process can
spend more computing resources than usual for every move. ACO has a heuristic that includes
a probability accumulated to the best reconfiguration individual in every iteration and has a
lighter consumption of computing resources that result in a good fit.
Thirdly, finding a global optimum reconfiguration that fulfills all constraints is a process that
depends on proper encoding and code. From the perspective of this project, ACO has better
adjustment as it has better execution times and provides results that fulfill all requirements.
All this implementation, to find the optimal reconfiguration, is the first step for future im-
provements in distribution grids, and it provides better quality of power and reliability to the
end customer. This project can be used as an early step for restoration services issues and a
better understanding of grids.

58
Bibliography

[1] L. Willis, “1 - introduction to transmission and distribution (t&d) networks: T&d in-
frastructure, reliability and engineering, regulation and planning”, in Electricity Trans-
mission, Distribution and Storage Systems, ser. Woodhead Publishing Series in Energy,
Z. Melhem, Ed., Woodhead Publishing, 2013, pp. 3–38, isbn: 978-1-84569-784-6. doi:
https : / / doi . org / 10 . 1533 / 9780857097378 . 1 . 3. [Online]. Available: http : / / www .
sciencedirect.com/science/article/pii/B9781845697846500019.

[2] A. A. Sallam and O. P. Malik, “Distribution system structure”, in Electric Distribution


Systems. IEEE, 2019, pp. 9–22, isbn: 9781119509301. doi: 10.1002/9781119509332.ch2.
[Online]. Available: https://ieeexplore.ieee.org/document/8538819.

[3] R. Pegado, Z. Ñaupari, Y. Molina, and C. Castillo, “Radial distribution network recon-
figuration for power losses reduction based on improved selective bpso”, Electric Power
Systems Research, vol. 169, pp. 206–213, 2019, issn: 0378-7796. doi: https://doi.org/
10.1016/j.epsr.2018.12.030. [Online]. Available: http://www.sciencedirect.com/
science/article/pii/S0378779618304279.

[4] Y. Fan, F. Ming-Tian, and Z. Zu-Ping, “China mv distribution network benchmark for
network integrated of renewable and distributed energy resources”, in CICED 2010 Pro-
ceedings, Sep. 2010, pp. 1–7.

[5] J. Glover, T. Overbye, and M. Sarma, Power System Analysis and Design. Cengage Learn-
ing, 2016, isbn: 9781305886957. [Online]. Available: https://books.google.is/books?
id=KHYcCgAAQBAJ.

[6] A. Bansal, A. Kumar, and N. Kumar, “Loss optimization of ieee 12 bus radial distribution
system integration with wind weibull distribution function using pso technique”, in 2016
IEEE 7th Power India International Conference (PIICON), 2016, pp. 1–6. doi: 10.1109/
POWERI.2016.8077324.

59
[7] K. Prakash, A. Lallu, F. R. Islam, and K. A. Mamun, “Review of power system distribution
network architecture”, in 2016 3rd Asia-Pacific World Congress on Computer Science and
Engineering (APWC on CSE), 2016, pp. 124–130. doi: 10.1109/APWC-on-CSE.2016.030.

[8] P. Lata and S. Vadhera, “Reliability improvement of radial distribution system by optimal
placement and sizing of energy storage system using tlbo”, Journal of Energy Storage,
vol. 30, p. 101 492, 2020, issn: 2352-152X. doi: https://doi.org/10.1016/j.est.2020.
101492. [Online]. Available: http://www.sciencedirect.com/science/article/pii/
S2352152X1930619X.

[9] M. Sepehry, M. Heidari-Kapourchali, A. Banajiger, and V. Aravinthan, “A new algorithm


for reliability evaluation of radial distribution networks”, in 2014 North American Power
Symposium (NAPS), Sep. 2014, pp. 1–6. doi: 10.1109/NAPS.2014.6965483.

[10] J. Schlabbach and K. Rofalski, Power System Engineering: Planning, Design, and Op-
eration of Power Systems and Equipment. Wiley, 2008, isbn: 9783527622801. [Online].
Available: https://books.google.is/books?id=8kRBtmQnvdYC.

[11] J. Zhu, “Optimal reconfiguration of electrical distribution network”, in Optimization of


Power System Operation. 2015, pp. 483–528. doi: 10.1002/9781118887004.ch12.

[12] A. Abdelaziz, F. Mohammed, S. Mekhamer, and M. Badr, “Distribution systems reconfig-


uration using a modified particle swarm optimization algorithm”, Electric Power Systems
Research, vol. 79, no. 11, pp. 1521–1530, 2009, issn: 0378-7796. doi: https://doi.org/
10.1016/j.epsr.2009.05.004. [Online]. Available: http://www.sciencedirect.com/
science/article/pii/S0378779609001394.

[13] T. T. Nguyen, T. T. Nguyen, N. A. Nguyen, and T. L. Duong, “A novel method based


on coyote algorithm for simultaneous network reconfiguration and distribution generation
placement”, Ain Shams Engineering Journal, 2020, issn: 2090-4479. doi: https://doi.
org/10.1016/j.asej.2020.06.005. [Online]. Available: http://www.sciencedirect.
com/science/article/pii/S209044792030126X.

[14] M. R. Vuluvala and L. M. Saini, “Load balancing of electrical power distribution system:
An overview”, in 2018 International Conference on Power, Instrumentation, Control and
Computing (PICC), 2018, pp. 1–5. doi: 10.1109/PICC.2018.8384780.

[15] P. R. Babu, R. Shenoy, N. Ramya, Soujanya, and S. Shetty, “Implementation of aco tech-
nique for load balancing through reconfiguration in electrical distribution system”, in 2014

60
Annual International Conference on Emerging Research Areas: Magnetics, Machines and
Drives (AICERA/iCMMD), 2014, pp. 1–5. doi: 10.1109/AICERA.2014.6908233.

[16] A. E. Abu-Elanien, M. Salama, and K. B. Shaban, “Modern network reconfiguration


techniques for service restoration in distribution systems: A step to a smarter grid”,
Alexandria Engineering Journal, vol. 57, no. 4, pp. 3959–3967, 2018, issn: 1110-0168.
doi: https : / / doi . org / 10 . 1016 / j . aej . 2018 . 03 . 011. [Online]. Available: https :
//www.sciencedirect.com/science/article/pii/S1110016818301893.

[17] H. Schmidt, N. Ida, N. Kagan, and J. Guaraldo, “Fast reconfiguration of distribution


systems considering loss minimization”, IEEE Transactions on Power Systems, vol. 20,
no. 3, pp. 1311–1319, 2005. doi: 10.1109/TPWRS.2005.846180.

[18] G. Peponis, M. Papadopoulos, and N. Hatziargyriou, “Distribution network reconfiguration


to minimize resistive line losses”, IEEE Transactions on Power Delivery, vol. 10, no. 3,
pp. 1338–1342, 1995. doi: 10.1109/61.400914.

[19] E. Ramos, A. Exposito, J. Santos, and F. Iborra, “Path-based distribution network mod-
eling: Application to reconfiguration for loss reduction”, IEEE Transactions on Power
Systems, vol. 20, no. 2, pp. 556–564, May 2005, issn: 1558-0679. doi: 10.1109/TPWRS.
2005.846212.

[20] F. Llorens-Iborra, J. Riquelme-Santos, and E. Romero-Ramos, “Mixed-integer linear pro-


gramming model for solving reconfiguration problems in large-scale distribution systems”,
Electric Power Systems Research, vol. 88, pp. 137–145, 2012, issn: 0378-7796. doi: https:
/ / doi . org / 10 . 1016 / j . epsr . 2012 . 01 . 014. [Online]. Available: https : / / www .
sciencedirect.com/science/article/pii/S037877961200034X.

[21] M. Baran and F. Wu, “Network reconfiguration in distribution systems for loss reduction
and load balancing”, IEEE Transactions on Power Delivery, vol. 4, no. 2, pp. 1401–1407,
Apr. 1989, issn: 1937-4208. doi: 10.1109/61.25627.

[22] R. Sarfi, M. Salama, and A. Chikhani, “A survey of the state of the art in distribution
system reconfiguration for system loss reduction”, Electric Power Systems Research, vol. 31,
no. 1, pp. 61–70, 1994, issn: 0378-7796. doi: https://doi.org/10.1016/0378-7796(94)
90029-9. [Online]. Available: https://www.sciencedirect.com/science/article/pii/
0378779694900299.

61
[23] M. W. Siti, D. V. Nicolae, A. A. Jimoh, and A. Ukil, “Reconfiguration and load balancing
in the lv and mv distribution networks for optimal performance”, IEEE Transactions on
Power Delivery, vol. 22, no. 4, pp. 2534–2540, Oct. 2007, issn: 1937-4208. doi: 10.1109/
TPWRD.2007.905581.

[24] H. Karimianfard and H. Haghighat, “An initial-point strategy for optimizing distribution
system reconfiguration”, Electric Power Systems Research, vol. 176, p. 105 943, 2019, issn:
0378-7796. doi: https://doi.org/10.1016/j.epsr.2019.105943. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S0378779619302627.

[25] O. Badran, S. Mekhilef, H. Mokhlis, and W. Dahalan, “Optimal reconfiguration of distribu-


tion system connected with distributed generations: A review of different methodologies”,
Renewable and Sustainable Energy Reviews, vol. 73, pp. 854–867, 2017, issn: 1364-0321.
doi: https://doi.org/10.1016/j.rser.2017.02.010. [Online]. Available: https:
//www.sciencedirect.com/science/article/pii/S1364032117302101.

[26] M. H. Alrefaei and A. H. Diabat, “A simulated annealing technique for multi-objective sim-
ulation optimization”, Applied Mathematics and Computation, vol. 215, no. 8, pp. 3029–
3035, 2009, issn: 0096-3003. doi: https : / / doi . org / 10 . 1016 / j . amc . 2009 . 09 .
051. [Online]. Available: https : / / www . sciencedirect . com / science / article / pii /
S0096300309008807.

[27] D. Guan and A. Fan, “Power distribution network reconfiguration based on fuzzy control
theory and ant colony algorithm”, in Proceedings of 2011 International Conference on Elec-
tronic Mechanical Engineering and Information Technology, vol. 3, Aug. 2011, pp. 1230–
1232. doi: 10.1109/EMEIT.2011.6023272.

[28] S. Koziel, A. L. Rojas, and S. Moskwa, “Power loss reduction through distribution net-
work reconfiguration using feasibility-preserving simulated annealing”, in 2018 19th Inter-
national Scientific Conference on Electric Power Engineering (EPE), May 2018, pp. 1–5.
doi: 10.1109/EPE.2018.8396016.

[29] A. Ahuja, A. Pahwa, B. K. Panigrahi, and S. Das, “Pheromone-based crossover operator


applied to distribution system reconfiguration”, IEEE Transactions on Power Systems,
vol. 28, no. 4, pp. 4144–4151, Nov. 2013, issn: 1558-0679. doi: 10.1109/TPWRS.2013.
2263257.

[30] S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, “Optimization by simulated annealing”,


Science, vol. 220, no. 4598, pp. 671–680, 1983, issn: 0036-8075. doi: 10.1126/science.

62
220.4598.671. eprint: https://science.sciencemag.org/content/220/4598/671.
full.pdf. [Online]. Available: https://science.sciencemag.org/content/220/4598/
671.

[31] Y. Eren, İ. B. Küçükdemiral, and İ. Üstoğlu, “Chapter 2 - introduction to optimiza-


tion”, in Optimization in Renewable Energy Systems, O. Erdinç, Ed., Boston: Butterworth-
Heinemann, 2017, pp. 27–74, isbn: 978-0-08-101041-9. doi: https://doi.org/10.1016/
B978-0-08-101041-9.00002-8. [Online]. Available: http://www.sciencedirect.com/
science/article/pii/B9780081010419000028.

[32] C.-Y. ( Huang, C.-Y. Lai, and K.-T. ( Cheng, “Chapter 4 - fundamentals of algorithms”,
in Electronic Design Automation, L.-T. Wang, Y.-W. Chang, and K.-T. ( Cheng, Eds.,
Boston: Morgan Kaufmann, 2009, pp. 173–234, isbn: 978-0-12-374364-0. doi: https :
/ / doi . org / 10 . 1016 / B978 - 0 - 12 - 374364 - 0 . 50011 - 4. [Online]. Available: https :
//www.sciencedirect.com/science/article/pii/B9780123743640500114.

[33] C.-R. Hwang, Simulated annealing: Theory and applications, 1988.

[34] N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller, “Equation


of state calculations by fast computing machines”, The journal of chemical physics, vol. 21,
no. 6, pp. 1087–1092, 1953.

[35] H.-D. Chiang and R. Jean-Jumeau, “Optimal network reconfigurations in distribution sys-
tems. i. a new formulation and a solution methodology”, IEEE Transactions on Power De-
livery, vol. 5, no. 4, pp. 1902–1909, Oct. 1990, issn: 1937-4208. doi: 10.1109/61.103687.

[36] A. Skoonpong and S. Sirisumrannukul, “Network reconfiguration for reliability worth en-
hancement in distribution systems by simulated annealing”, in 2008 5th International Con-
ference on Electrical Engineering/Electronics, Computer, Telecommunications and Infor-
mation Technology, vol. 2, May 2008, pp. 937–940. doi: 10.1109/ECTICON.2008.4600585.

[37] M. Dorigo, D. de Recherches Du Fnrs Marco Dorigo, T. Stützle, and T. Stützle, Ant Colony
Optimization, ser. Bradford book. BRADFORD BOOK, 2004, isbn: 9780262042192. [On-
line]. Available: https://books.google.is/books?id=%5C_aefcpY8GiEC.

[38] S. Goss, S. Aron, J.-L. Deneubourg, and J. M. Pasteels, “Self-organized shortcuts in the
argentine ant”, Naturwissenschaften, vol. 76, no. 12, pp. 579–581, 1989.

[39] M. Dorigo, M. Birattari, and T. Stutzle, “Ant colony optimization”, IEEE Computational
Intelligence Magazine, vol. 1, no. 4, pp. 28–39, Nov. 2006, issn: 1556-6048. doi: 10.1109/
MCI.2006.329691.

63
[40] E. Carpaneto and G. Chicco, “Distribution system minimum loss reconfiguration in the
hyper-cube ant colony optimization framework”, Electric Power Systems Research, vol. 78,
no. 12, pp. 2037–2045, 2008, Special Issue Papers Presented at the 6th World Energy
System Conference, issn: 0378-7796. doi: https://doi.org/10.1016/j.epsr.2008.
06.009. [Online]. Available: https://www.sciencedirect.com/science/article/pii/
S0378779608001909.

[41] J. Zhu, “Power flow analysis”, in Optimization of Power System Operation. IEEE, 2015,
pp. 13–50. doi: 10.1002/9781118887004.ch2. [Online]. Available: https://ieeexplore.
ieee.org/document/7046736.

[42] L. Thurner, A. Scheidler, F. Schäfer, J.-H. Menke, J. Dollichon, F. Meier, S. Meinecke, and
M. Braun, “Pandapower—an open-source python tool for convenient modeling, analysis,
and optimization of electric power systems”, IEEE Transactions on Power Systems, vol. 33,
no. 6, pp. 6510–6521, Nov. 2018, issn: 1558-0679. doi: 10.1109/TPWRS.2018.2829021.

[43] N. Kang and Y. Liao, “Equivalent pi circuit for zero-sequence double circuit transmission
lines”, in 2012 IEEE Power and Energy Society General Meeting, Jul. 2012, pp. 1–6. doi:
10.1109/PESGM.2012.6344945.

[44] N. C. Koutsoukis, P. S. Georgilakis, N. D. Hatziargyriou, O. Mohammed, A. Elsayed, K.


An, K. Hur, N. H. Luong, P. A. N. Bosman, M. O. W. Grond, H. La Poutré, H. Chiang, J.
Cui, T. Xu, E. N. Asada, J. B. A. London, F. O. Saraiva, W. Sun, K. Venayagamoorthy, Q.
Zhou, S. Wang, Y. Zhang, D. Gusain, J. Rueda, J. C. Boemer, P. Palensky, D. L. Alvarez,
and S. Rivera, “Distribution system”, in Applications of Modern Heuristic Optimization
Methods in Power and Energy Systems. 2020, pp. 381–611. doi: 10.1002/9781119602286.
ch5.

[45] R. Diestel, Graph Theory (Graduate Texts in Mathematics). Springer, Aug. 2005, isbn:
3540261826. [Online]. Available: http://www.amazon.ca/exec/obidos/redirect?tag=
citeulike04-20%7B%5C&%7Dpath=ASIN/3540261826.

[46] G. Stagg and A. El-Abiad, Computer Methods in Power System Analysis, ser. McGraw-
Hill series in electronic systems. McGraw-Hill, 1968. [Online]. Available: https://books.
google.is/books?id=q7U8AAAAIAAJ.

[47] A. Swarnkar, N. Gupta, and K. R. Niazi, “Efficient reconfiguration of distribution systems


using ant colony optimization adapted by graph theory”, in 2011 IEEE Power and Energy
Society General Meeting, 2011, pp. 1–8. doi: 10.1109/PES.2011.6039006.

64
[48] B. Stojanović and T. Rajić, “Novel approach to reconfiguration power loss reduction prob-
lem by simulated annealing technique”, International Transactions on Electrical Energy
Systems, vol. 27, no. 12, e2464, 2017, e2464 ITEES-17-0236.R1. doi: https://doi.org/
10.1002/etep.2464. eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/
etep.2464. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/
etep.2464.

[49] A. Shaheen, A. Elsayed, R. A. El-Sehiemy, and A. Y. Abdelaziz, “Equilibrium opti-


mization algorithm for network reconfiguration and distributed generation allocation in
power systems”, Applied Soft Computing, vol. 98, p. 106 867, 2021, issn: 1568-4946. doi:
https://doi.org/10.1016/j.asoc.2020.106867. [Online]. Available: https://www.
sciencedirect.com/science/article/pii/S156849462030805X.

[50] B. Stojanović and T. Rajić, “Novel approach to reconfiguration power loss reduction prob-
lem by simulated annealing technique”, International Transactions on Electrical Energy
Systems, vol. 27, no. 12, e2464, 2017.

65

You might also like