Distributed Generation Allocation and Sizing: A

Comparison of Metaheuristics Techniques

Luis Lopez Jose Doria-Garcia
Dept. of Electrical & Electronics Engineering Dept. of Electrical & Electronics Engineering
Universidad del Norte Universidad del Norte
Barranquilla, Colombia Barranquilla, Colombia
[email protected] [email protected]

Carlos Pimienta Adriana Arango-Manrique

Dept. of Electrical & Electronics Engineering Dept. of Electrical & Electronics Engineering
Universidad del Norte Universidad del Norte
Barranquilla, Colombia Barranquilla, Colombia
[email protected] [email protected]

Abstract—This paper presents a comparison of metaheuristic Ploss Total active power losses
techniques to locate and sizing distributed generation in distribu- Vi Voltage at bus i
tion feeder. The objective function minimizes active power losses
in a radial distribution system including distributed generation. I. I NTRODUCTION
Tabu Search, Scatter Search and Ant Colony were the techniques
used to compare results in minimization losses and improvement Distributed Generation (DG) is about to introduce small-
of voltage profiles in critical nodes when distributed generation scale sources to produce electricity close to the end users.
was included in distribution feeder. The analysis compares the Distribution systems are those systems where a feeder supplies
performance of techniques in term of accuracy and execution the demand of the users. The main character in the distribution
times.
Index Terms—Distributed generation, metaheuristics, opti- lines models is that the ratio of resistance and reactance is
mization, radial distribution system. close to the unit (R/X = 1). Thus there is a considerable
amount of losses associated with the Joules effect present
N OMENCLATURE in the lines. These technical losses cause problems in the
lifespan of the cables and over-costs in the normal operation
Indices and sets
in distribution system [1].
i, j Index of buses
Distributed generation has always been present in the elec-
L Set of distribution lines
trical systems, however, recent changes in schemes, operation,
N Set of buses
and technology in the power systems have encouraged the
Parameters
nom growing development of strategies for the inclusion of DG
Ii,j Nominal current capacity of branch i,j
[2], [3]. The increase of use of DGs technology was asso-
Pg,min , Pg,max Active power limits generation
ciated with the decreasing costs of a generation because of
Pload , Qload Total active and reactive power demand for dis-
the increasing competitiveness of renewable energies in the
tribution system
electricity market [4]. DG can be located in different nodes
Qg,min , Qg,max Reactive power limits generation
of the distribution system, and the impact depends on the
Ri,j , Xi,j Resistance and reactance of branch i to j
characteristics inherent to the network [5].
Sbase Apparent power base
The problem of fixing the number of generators in dis-
Variables
2 tributed generation is not trivial mainly due to the high number
Ii,j Current from bus i to j
of variables to study and that the equations that govern the
Pg,i , Qg,i Active and reactive power generation in bus i by
power flow are of nonlinear origin, causing a non-convex prob-
generator g
lem. The research proposed the use of metaheuristic techniques
This paper was supported by Universidad del Norte and Instituto Colom- optimization such as Ant Colony, Tabu Search, and Scattered
biano para el Desarrollo de la Ciencia y la Tecnologı́a Francisco José de Search, that allows to include DG in a specific configuration
Caldas -Colciencias from a series of steps described in the methodology shows in
section II, evaluating which metaheuristic techniques have the

978-1-7281-0653-3/19/$31.00 © 2019 IEEE

better performance in the minimization of active power losses A. Distributed generation allocation
[6], [7]. An additional challenge that must be faced is to determine
This paper is organized as follows. Section II, Problem the nodes of the system in which DG will be allocated. The
formulation; Section III Case study; Section IV, Results and procedure performed to achieve this objective is described
discussion, and Section V Conclusions. below.
II. P ROBLEM F ORMULATION 1) Power losses: an analysis of total system losses was
carried out by placing a single generator in each node
This paper proposes a methodology to obtain the configu- and analyze the load flow while the DG capacity is
ration that minimizes active power losses by determining the varied.
placement and sizing of a set of DG. The objective function 2) Bus features: analysis of how each demand influences
is shown in Eqn. (2) and it is composed of the sum of the the node where it is directly connected and the surround-
technical losses of the network. The classic constraints of ing nodes.
the problem of optimization to integrate DG in distribution 3) Finding a good combination: selection of three nodes
systems correspond to operating characteristics of the network in the system that allow obtaining the best performance
and elements operational limits. The set of the constraint of minimizing losses when connecting DG.
the problem is presented in the Eqn.(3) to Eqn.(8). The Eqn.(3) The results of the active power losses analysis are recorded
represents the active power nodal balance, Eqn.(8) considers in Fig. 1. Fig. 1 shows how the injected power increases. The
the maximum and minimum limits of bus voltages, and finally total system losses decreases in every bus except for generator
Eqn.(5) displays the minimum and maximum active power on bus 634. In bus 634, injecting a power DG greater than 0.7
generation limits for each distributed generator connected to MW increases the losses. In Fig. 1 is also appreciated that
the distribution system. Objective function: locating a generator near the swing bus impacts have a lower
total system loss reduction than those located at the farthest
min Ploss (1)
nodes. The nodes with greater individual impact are nodes 671
and 692, followed by 675 and 680.
This paper aims to determine the best location to install
L
X three sources acting as distributed generation in the sys-
2
Ploss = Sbase Ii,j Ri,j (2) tem. From the above results, it could be concluded that a
i,j combination of generators with better individual performance
could be the solution. Therefore, a possible combination is
Subject to: Power Balance: set of DG connected to nodes 671, 692 and 675. However, a
N L
thorough analysis of the connected loads and the power flow
X
Pg,i +
X
2
Ii,j Ri,j = Pload (3) performance by introducing the three generators at the same
i i,j
time, leads that the allocation of the set of DG with the best
performance is in nodes 634, 671 and 675.
N L
B. Metaheuristic technique in DG Optimization problem
X X
2
Qg,i + Ii,j Xi,j = Qload (4)
i i,j Metaheuristics Techniques (MT) are approximate algo-
rithms applicable to different optimization problems only by
Generator Limits: making small changes of adaptation of problem characteristics.
Therefore, metaheuristics are general procedures that do not
Pg,min ≤ Pg,i ≤ Pg,max (5) depend on the issue. The MT provide excellent results ap-
proximating the optimum, in a short time and with reasonable
computational cost. These techniques carry out explorations
Qg,min ≤ Qg,i ≤ Qg,max (6) throughout the search space trying to find the best solution
in the shortest possible time, avoiding converge in local
Overheat Flow Limits:
optimums briefly [8].
Ii,j There are two main types of MT: trajectory-based and
nom ≤ 1
Ii,j
(7)
population-based. The first type uses a single point in the
space to be updated and the reference point to explore the
Voltage Limits: neighborhood. The second type works with a set of individuals
or solutions, and these will evolve or move according to the
0.95 ≤ |Vi | ≤ 1.05 (8) method of trajectory. There are algorithms trajectory-based
like The simulated cooling, or The Tabu Search, and algorithms
population-based like The genetic algorithms, The Ant Colony
Optimization and Scatter Search.
 300 Algorithm 1 Tabu Search
1: Stage Initialization
Bus 632
Bus 633 2: Generate population
250 Bus 634 3: Generate all the combinations for each population.
4: for Each iteration do
Bus 671
5: for For each combination i do
Bus 675
Active losses (kW)

200 6: Select a population.

Bus 680
7: Calculate the objective function
Bus 692
8: Save minimum value for each population
150 9: end for
10: for For each permutation i do
11: Get the solution found Lossi and compare with TL
100
Value
12: if Lossi < LossT L then
50 13: Lossglobal = Lossi
14: end if
15: end for
0 16: end for
0 0.5 1 1.5 2 2.5 3 17: return Value of TL
Active power DG (MW)

Fig. 1. Total active power losses in different DG bus allocation. 4) Combination method: To combine all the solutions of
the reference set (usually between couples) seeking to
improve them.
C. Tabu Search 5) Updating reference set: To update the set with solutions
Tabu Search (TS) optimization is a metaheuristic procedure with better features [13].
based on concepts of artificial intelligence and adaptive mem- Algorithm 2 shows a description of the methodology im-
ory [9]. TS handles random and deterministic trajectory move- plemented for SS.
ments unlike other methods such as Ant Colony optimization
and Particle Swarm optimization, which only control random Algorithm 2 Scatter Search
movements [10]. 1: Stage Initialization
TS traverses the whole neighborhood of possible solutions 2: n = Initial population size
due to the deterministic movements. This behavior could 3: P =∅
cause stagnation at a local minimum because next answer 4: b = Reference set size
could be not as right as the previous.TS uses memories tools, 5: while |P | ≤ n do
classified as follows: short-term memory, which stores recent 6: S = DiverseGeneration()
events, and long-term memory, where are keeped the most 7: if S ∈
/ P then
frequent data. Therefore, TS penalizes new movements as 8: P ∪S
Tabu movement [9]. In [10]–[12], authors applied TS for the 9: end if
optimization of distributed generation systems. Algorithm 1 10: end while
shows a description of the methodology implemented for TS. 11: Bluid RefSet with the best b/2 solutions of P, and the
most diverse b/2 solutions with respect to those already
D. Scatter Search in RefSet
12: Improving method
Scatter Search (SS) is an algorithm population-based; it
13: while New Solution do
performs a local search while avoiding getting stuck in a local
minimum. SS seeks to preserve a reference set b conformed by 14: Combination method
b/2 high-quality solutions and b/2 diverse solutions. Combina- 15: Improving method
16: end while
tions are carried out on the reference set looking for improving
17: return Best solution
each solution. The strategy involves five main stages:
1) Initial population: To generate a set P with diverse and
high-quality solutions.
E. Ant Colony
2) Creating reference set: To select the b/2 best quality
solutions and the b/2 most diverse solutions of the P set. The theory of the optimization for the colony of ants (Ant
3) Improving method: To improve the reference set before Colony Optimization, ACO), was introduced by Marco Dorigo
starting the iterative process. in the early 1990s as a tool for the solution of complex
optimization problems [8]. ACO belongs to the class of meta- Algorithm 3 Ant Colony Optimization
heuristic methods, which are approximate algorithms to obtain 1: Initialization stage
satisfactory answers as complex problems in a reasonable 2: A set amount of ants
amount of computing time. 3: General solution
4: Pheromone matrix t = t0
5: for Each iteration do
F F F
6: for Each ant i do
7: Select a route randomly
b 8: Calculate the objective function
9: Update the pheromone matrix
10: end for
a
11: for Each ant i do
12: Obtain the best solution Lossi
13: if Lossi < Lossglobal then
N N N
14: Lossglobal = Lossi
15: end if
1 2 3 16: end for
17: Evaporate pheromone t
Fig. 2. Ant Colony Optimization diagram. 18: end for
19: return Best route obtained
The ACO algorithms are iterative programs that consider
that each ant creates a solution to the problem through a
graphic (Fig. 2). Each edge represents the possible paths that
the insect can take and has the following type of information that allows evaluating the injection of distributed generation
associated with it: in the system. The changes made to the system in question
• Heuristic information: it measures the heuristic prefer- are listed below:
ence that ants have to move from one node to another.
The path traveled from one node to another is an edge. 1) The 650 node was added as the feeder node of the
This information is not modified during the execution of distribution system.
the algorithm. 2) Line 650-632 was added with the characteristics de-
• Information on the traces of artificial pheromones: the scribed in the original test system.
choice learned in the movement from one node to an- 3) Seven generators were connected in each of the three-
other is measured, which attempts to emulate the real phase buses with the initial generation of zero MW.
pheromone dropped by natural ants. This information is 4) Constant impedance loads were modified to constant
adjusted while the algorithm is running and depends on power loads on nodes 646 and 652.
the solutions obtained by the insects [14].
III. C ASE S TUDY 650 Feeder Bus
1.0 pu 0.0 4160 V
IEEE 13 Node Test Distribution ?
1.0 pu -120
1.0 pu 120
swing System
C
A
B

There are several computational techniques and algorithms

A
B
C

650 Power-Flow Phasors

Results
powergui

to address the problem of the analysis of distribution systems. 650-632

Universidad del Norte
Barranquilla - Colombia
+

Authors:
aA
bB
cC

L650-632
Jose Doria Garcia

In 1992, works were published with a complete set of data P= [8500 33000 58500] W
Ql=[5000 19000 34000] var
Qc=[0 0 0] var
Luis Lopez Diaz
Carlos Pimienta Martinez
Mails:
A
B
C

G632 632 632 Yg PQ

{jdoriad, llopeza, cepimienta}@uninorte.edu.co
A A

from radial distribution systems whose purpose was to have a B

C
B
C 634 Yg PQ
P= [160000 120000 120000] W
C
A
B

Ql=[110000 90000 90000] var

Qc=[0 0 0] var

set of methods in which to verify and validate the operation aA aA + Aa A A a Aa A

+

B bB + B bB + Bb B B b Bb B
646_Z
P=[230000] W C cC C cC Cc C C Yg Ygc Cc C
Ql=[132000] var 645-646 632-645 632-633
Qc=0 var
646 L645-646 645 L632-645 L632-633 633 XFM-1 L633-634 634
4.16 kV /480 V
A
B
C

of solutions [15].
A
B
C
+

645 PQ G633 G634

P=[170000] W
Ql=[125000] var 632-671
Qc=0 var

IEEE 13 test feeder has been used as a basis for carrying

aA
bB
cC

L632-671

P= [8500 33000 58500] W

out different researches in distribution networks. In [16],

Ql=[5000 19000 34000] var
Qc=[0 0 0] var
671 Yg PQ
A
B
C

authors developed an algorithm of estimation of the steady-

A
B
C

G671 671
A A
B B
C C

state of distribution system based on current and impact of 671 D PQ

P= [385000 385000 385000] W
Ql=[220000 220000 220000] var
692 D I
P= [0.001 0.001 170000] W
675 Yg PQ
P= [485000 68000 290000] W
C

Qc=[0 0 0] var
A
B

C
A
B

Ql=[0 0 151000] var Ql=[190000 60000 212000] var

Qc=[0 0 0] var Qc=[0 0 0] var
611 I 692 675

meter location. In [17] proposes distributed generation models P=[170000] W

+

Ql=[80000] var aA A aA + A a Aa A + Aa A
Qc=0 var bB bB B b Bb B Bb B
??? cC + C cC C c Cc C Cc C
L684-611 684-611 L671-684 Breaker L671-692 692-675 L692-675
684 671-684
A

611
B
C

A
B
C

for power flow studies. In [18] performs a research about

A
B
C
+

611 Z G675 675 Yg Z

G692
+

P=0 W
P= [0 0 0] W
Ql=0 var 684-652 Ql=[0 0 0] var
Qc=100e3 var 671-680
Qc=[200 200 200]*1e3 var

photovoltaic of renewable energy laboratory in Universidad

aA
bB
cC

L684-652
aA
bB
cC

L671-680

del Norte with the objective of its inclusion in the microgrid

A

652
G680
A
+

there built. 652 Z

P=[128000] W
Ql=[86000] var
Qc=0 var
B
C
A
B
C

680

A. Modified Radial distribution test feeder (IEEE-13)

Fig. 3. Modified IEEE-13 Implemented in MATLAB®
MATLAB has a version of IEEE-13 nodes test feeder imple-
mented in Simulink. The excersice include some modification
 IV. R ESULTS AND DISCUSSION 1.1
It is not necessary that n iterations for the TS method Taboo Search
represent the same effort neither for the SS nor the ACO. Scatter Search
Therefore, to make a reasonable comparison, it is proposed to Ant Colony
execute the MT considering the number of times the Objective 1.05

Voltage [p.u.]
Function (EOF: Evaluating Objective Function) was evaluated.
As an example, it can be observed in the Fig. 4: for the TS,
1200 EOF corresponds to 345 iterations, while for the SS only
20 iterations are enough to perform 1100 EOF. Finally, for 1
the ACO 100 iterations imply 1000 EOF. Figure 4 shows the
evolution of the minimization for TS, SS, and ACO.
The Fig. 4 shows that the evolution of the OF for the case
of the TS does not seem to stabilize, but with each iteration, it 0.95
advances towards a new minimum. On the other hand, the SS
5 10 15 20 25 30
and ACO methods were reached in the first ten (10) iterations.
Finally, Fig. 5 shows the voltage deviation of every MT. Bus ID
4
Taboo Search
Tabu Search
Minimal active power losses (kW)

Scatter Search
50 Scatter Search Ant Colony
Ant Colony 3

Sbus [p.u]
40
2

30 1

20 0
0 5 10 15 20 634 671 675
Iteration Bus ID

Fig. 5. Voltage profile and power generation for metaheuristic solutions

Fig. 4. Comparison of Metaheuristic Optimization

Table I shows the final solution obtained for each of the

optimization methods. Table I shows the capacity in active TABLE I
power (kW) of distributed generation connected in nodes 634, R ESULTS COMPARISON 13 N ODE IEEE T EST F EEDER
671 and 675; and an analytical approach for minimizing active TS (kW/%) SS (kW/%) AC (kW/%) MG
power losses in IEEE-13 node test Feeder is in the column DGs Power
label like MG. 634 Node 440 (92%) 432 (92%) 492 (92%) 475
A box diagram for each one of the techniques used, dis- 671 Node 920 (92%) 1000 (100%) 800 (80%) 1000
played the average value and the dispersion of the results 675 Node 810 (81%) 752 (75%) 813 (81%) 1000
under multiple iterations (40 times with the same initial
Losses 29.2 (9.5%) 28.84 (7.9%) 29.3 (9.7%) 26.7
conditions) of the algorithms. In Fig. 6 is evident that SS
EOF 1200 1100 1000 -
have on average the highest accuracy to obtain the minimum
of the objective function. TS presents the highest dispersion
of the programmed methods. AC has on average the worst
performance work in calculating the minimization of the
objective function. Memory 2.42 GB. System Type x64-based PC). TS presented
Fig. 7 shows the execution elapsed times of each of the the highest average execution elapsed time to determine the
algorithms in minutes. For the same number of performances, minimization of the function. SS showed the shortest perfor-
the elapsed time was evaluated in the same machine (Pro- mance passed time of the three programmed algorithms. AC
cessor: Intel(R) Core(TM) i5-4210U CPU @ 1.70GHz, 2401 obtained the lowest dispersion in execution elapsed times of
Mhz, 2 Core(s), 4 Logical Processor(s). Available Physical the algorithm.
 34 R EFERENCES
[1] Pavlos S. Georgilakis and Nikos D. Hatziargyriou. Optimal distributed
generation placement in power distribution networks: Models, methods,
33
Active power loss (kW)

and future research. IEEE Transactions on Power Systems, 2013.

[2] T. Q. D. Khoa, P. t. t. Binh, and H. b. Tran. Optimizing location and
sizing of distributed generation in distribution systems. In 2006 IEEE
32 PES Power Systems Conference and Exposition, pages 725–732, Oct
2006.
[3] M. H. Elkazaz, A. Hoballah, and A. M. Azmy. Optimizing distributed
31 generation operation for residential application based on automated
systems. In 2015 4th International Conference on Electric Power and
Energy Conversion Systems (EPECS), pages 1–6, Nov 2015.
30 [4] Zeineb Abdmouleh, Adel Gastli, Lazhar Ben-Brahim, Mohamed
Haouari, and Nasser Ahmed Al-Emadi. Review of optimization
techniques applied for the integration of distributed generation from
29 renewable energy sources. Renewable Energy, 113:266–280, 2017.
[5] Juan Andrés Martı́n Garcı́a and Antonio José Gil Mena. Optimal dis-
Taboo Search Scatter Search Ant Colony tributed generation location and size using a modified teaching-learning
based optimization algorithm. International Journal of Electrical Power
and Energy Systems, 2013.
Fig. 6. Comparison of minimal active power losses (kW). [6] John Silberholz, Bruce Golden, Swati Gupta, and Xingyin Wang.
Computational comparison of metaheuristics. In International Series
in Operations Research and Management Science. 2010.
[7] Nagaballi Srinivas and Vijay Kale. Real power loss reduction in radial
distribution system by integrating dgs using mtlbo algorithm. 06 2018.
22 [8] Christian Blum. Metaheuristics in Combinatorial Optimization:
Overview and Conceptual Comparison Metaheuristics in Combinatorial
Optimization. Technical Report 3, 2003.
20 [9] Fred Glover and Eric Taillard. A user’s guide to tabu search. Annals of
Elapsed time

operations research, 41(1):1–28, 1993.

18 [10] Hiroyuki Mori and Yoshinori Iimura. Application of parallel tabu search
to distribution network expansion planning with distributed generation.
In Power Tech Conference Proceedings, 2003 IEEE Bologna, volume 1,
16 pages 6–pp. IEEE, 2003.
[11] Koichi Nara, Yasuhiro Hayashi, Kazushige Ikeda, and Tomoo Ashizawa.
Application of tabu search to optimal placement of distributed gen-
14 erators. In 2001 IEEE Power Engineering Society Winter Meeting.
Conference Proceedings (Cat. No. 01CH37194), volume 2, pages 918–
12 923. IEEE, 2001.
[12] M Gandomkar, M Vakilian, and M Ehsan. A genetic–based tabu search
Taboo Search Scatter Search Ant Colony algorithm for optimal dg allocation in distribution networks. Electric
Power Components and Systems, 33(12):1351–1362, 2005.
[13] R.A.G. Rendón, E.M.T. Ocampo, and A.H.E. Zuluaga. Técnicas
Fig. 7. Comparison of performance (min). Heurı́sticas y Metaheurı́sticas. Colección trabajos de investigación.
Universidad Tecnológica de Pereira. Vicerrectorı́a de Investigaciones,
Innovación y Extensión. Ingenierı́as Eléctrica, Electrónica, Fı́sica y
Ciencias de la Computación, 2015.
V. C ONCLUSIONS [14] Carlos Arturo Robles Algarı́n. Optimización por colonia de hormigas:
aplicaciones y tendencias. 2018.
In this study was executed the implementation of TS-SS-AC [15] William H Kersting. Radial distribution test feeders. IEEE Transactions
algorithms to reduce power losses including DGs in the distri- on Power Systems, 6(3):975–985, 1991.
bution network. In this article, the reduction of power losses [16] Caisheng Wang and M. Hashem Nehrir. Analytical approaches for
optimal placement of distributed generation sources in power systems.
considers the active power of DGs and the optimal location of IEEE Transactions on Power Systems, 2004.
the DGs. The effectiveness of the proposed algorithm to solve [17] SM Moghaddas-Tafreshi and Elahe Mashhour. Distributed generation
the DG power losses has been demonstrated with the study modeling for power flow studies and a three-phase unbalanced power
flow solution for radial distribution systems considering distributed
example. generation. Electric Power Systems Research, 79(4):680–686, 2009.
The proposed method was successfully applied to the IEEE [18] Paula Andrea De La Hoz Gómez, Andrea Carolina Silva Sandoval,
13 node test feeders for three different algorithms and cases et al. Microred: caso de estudio fotovoltaico del laboratorio de energı́as
renovables universidad del norte. 2017.
of DGs sizes and types of DGs. With proposed Algorithm
coding, this study showed that the setting parameters are
independent of the dimensions and types of DGs, and depend
on specifications of distribution networks.
A comparison was made between SS, TS, and AC. SS was
much better than TS and AC in terms of solution accuracy and
convergence process. The obtained solution has succeeded in
reducing distribution power losses up to 28.84kW. Thus, SS
proposed method is efficient for solving a distribution network
and DG allocation problem.