Electrical Engineering

https://doi.org/10.1007/s00202-020-01106-3

ORIGINAL PAPER

Optimal distributed generation allocation in unbalanced radial

distribution networks via empirical discrete metaheuristic
and steepest descent method
Francisco Carlos Rodrigues Coelho1 · Ivo Chaves da Silva Junior2 · Bruno Henriques Dias2 · Wesley Peres1 ·
Vitor Hugo Ferreira3 · André Luiz Marques Marcato2

Received: 8 April 2020 / Accepted: 3 September 2020

© Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
Loss minimization and voltage improvement through distributed generation (DG) planning is a well-established problem.
However, a careful review of the literature shows that there is still room for the development of efficient algorithms for this
purpose. In special, hybridization between optimization techniques is suitable for this complex problem, as it allows taking
advantage of the positive features of different approaches. In this work, a novel empirical discrete metaheuristic (EDM) is
presented and merged with the steepest descent method to solve the DG allocation problem. The allocation is broken into two
subproblems: sitting the DGs and sizing them. The EDM deals with the first subproblem, while the second one is solved by the
steepest descent method in an interchangeable optimization structure. The EDM tackles some key limitations of metaheuristic
family methods. Relatively, it shows: low results variability in different executions; low initial conditions dependency; and
few number parameters to tune. All simulations are performed in a communication scheme using the softwares Matlab and
OpenDSS. The obtained results with IEEE 34-bus and IEEE 123-bus distribution test systems were compared to the literature
and other metaheuristics, attesting the quality of the proposed approach.

Keywords Empirical discrete metaheuristic (EDM) · Distributed generation allocation · Hybrid optimization · Unbalanced
radial distribution system · Loss minimization

φ
Abbreviations PDG j Real power injection into bus j by the DG unity,
in phase φ (kW)
φ φ
Optimization problem PG j , PD j Real power generation and demand in bus j, in
phase φ (kW)
φ φ
Fobj Objective function QG j , Q Dj Reactive power generation and demand in bus
PL Power losses (kW) j, in phase φ (kvar)
φ φ φ
Vj Voltage at bus j, in phase φ (V) f P jk , f Q jk Real and reactive power flow from bus k to bus
NB Set of network buses j, in phase φ (kW, kvar)
Ph Set of phases:φ ∈ {A, B, C} max
PDG Real power DG unity upper bound (kW)
DG j Distributed generation status at bus j (on/off) V lim Voltage limit, V min or V max (V)

B Francisco Carlos Rodrigues Coelho

[email protected]
Empirical discrete metaheuristic
1 Department of Electrical Engineering, Federal University of
São João del-Rei (UFSJ), São João del-Rei, MG, Brazil N Number of individuals/solutions
2 Department of Electrical Energy, Federal University of Juiz ITE Maximum number of iterations
de Fora (UFJF), Juiz de Fora, MG, Brazil G Number of groups
3 Department of Electrical Engineering, Fluminense Federal STAG Number of iterations to consider stagnation
University (UFF), Niterói, RJ, Brazil t
xn,g Individual n, belonging to group g, at iteration t

123
 Electrical Engineering

x∗ Best individual/solution losses and voltage deviations in radial distribution systems.

Xbests Matrix containing best individuals from each group In [9], the DG allocation problem is solved by the NSGA II,
xpr Provisory individual where the objectives considered are finding minimum costs
A1 Random solution vector at minimum risk. In [10], the Jaya algorithm is used to allo-
A2 Small random solution vector (to perform a local cate DG in IEEE 123 bus test system aiming to reduce real
search) losses. A heuristic optimization method is proposed in [11]
to simultaneously allocate DG and capacitor. The objective
is to minimize real losses and to improve the voltage profile.
1 Introduction In [12], a methodology based on Tabu search is employed
to solve the distribution system planning in the presence of
Distribution generation (DG) can be defined as small-scale DG. In [13], a multiobjective particle swarm optimization
energy generators located closer to the load and dispersed algorithm (PSO) sites renewable DGs and capacitors in an
throughout the power system. The impacts of DG units are unbalanced distribution network, also including the mini-
of great importance to the power sector, as they can affect mization of total power loss. Other metaheuristics, such as
not only the system operation but also economic and envi- ant lion optimization [14] [15], gray wolf optimizer [16],
ronmental costs. coyote optimization algorithm [17] and cuckoo search (CS)
The DG can represent a lower energy cost to society if [18] have also been used to obtain optimal DG configuration.
environmental costs are taken into account, especially con- Although less common, other approaches can be found
sidering the cost of renewable energy resources, such as solar in the specialized literature. In [19], a linearization process
and wind, which foster energy savings. Many governments is proposed to minimize costs due to the implantation of
have increased the share of DGs in the distribution system different DG technologies. After being linearized, the prob-
through the use of net metering or feed-in tariffs [1–3]. This lem is solved by a commercial solver, as well as in [20]. In
increase in the DG will certainly be challenging, but it can [21], a modified Monte Carlo method is presented, aiming
also be profitable, depending on the business model. The at minimizing the power losses of two hypothetical distri-
primary difficulties are associated with the complexity of bution systems with uniformly distributed loads. The main
operating such a distribution system with an increased dis- feature of the proposed plan is to avoid simulations that are
persed generation of different technologies [4]. Expected close to already-checked DG configurations, reducing the
benefits include expanding the system’s reliability, postpon- computational time. In [22], an analytical efficient method
ing investments, improving the voltage profile and reducing is proposed to find the best power injection by DGs and
active losses in the power system. its optimum power factor. To define the locations of the
Positive impacts can be achieved if the DG is adequately sources, all possible configurations are verified. As the size
allocated, requiring a study of the power system character- of the network grows and more DG units are placed, the
istics as well as the generation sources available. Proper computational burden makes it hard to solve the problem
planning depends on realistic simulations and the applica- within a reasonable computational time. In [23], analytical
tion of optimization techniques. expressions are derived from DG possible connections in
The DG allocation problem has been investigated for at unbalanced distribution networks. DGs are considered to be
least two decades [5], because of its challenging nonlin- only from renewable sources.
ear nature, composed by integer and continuous variables. In [24], a hybridization between Tabu search and Chu-
Among the approaches proposed to produce an efficient Beasley GA is presented for siting and sizing DG units.
solution to the DG allocation problem, are those based on While the Tabu search determines the location, the meta-
heuristics, metaheuristics, numerical, analytical and hybrid heuristic based on GA is used to size the DG units. This
methods, with a predominance of the heuristics and meta- work also considers the allocation of intermittent sources,
heuristics due to their ability to deal with multimodal such as wind turbines and photovoltaic (PV) systems. In this
optimization problems. case, a statistical method determines the operation point of
Genetic algorithm (GA) is employed in [6] aiming at the DGs. GA can also be found in hybrid approaches with
placing DGs on the distribution system considering yearly the interior point method [25] and local search [26], Monte
load growth, daily variations and customers’ load type to Carlo [27] and Hong’s two-point estimate method [28]. In
minimize investments, operational costs and power losses. [29], PSO is used to place distributed sources and an ana-
Similar work is presented in [7], where the DG unitis are lytical method based on derivatives determines the amount
placed simultaneously with capacitor banks. The sizing of of power injected, aiming to minimize power losses. In [30],
DG power plants is not considered in this work (leading to PSO is also employed in a hybrid version of a multiobjec-
a fixed amount of power injection). GA is also the chosen tive approach. In [31], an analytical method called zero bus
tool in [8], where its adaptative version (AGA) minimizes load flow is hybridized with symbolic organism search, a

123
Electrical Engineering

recently proposed metaheuristic. In [32], sine cosine algo- need to be tuned to apply the proposed metaheuristic: two
rithm is merged with chaos map theory. of them are the number of individuals and the number of
Due to the number of works presented in the literature, iterations. Those two parameters are common to all popu-
some reviewing articles have been published concerning lation metaheuristics. This is an important advantage of the
the DG allocation problem. In [33], a review considering EDM compared to other metaheuristics because the tuning
the mathematical model and the importance of combinato- process can be laborious and lead to varied performances.
rial optimization techniques is highlighted. As a conclusion, This paper evaluated the optimization approach by analyz-
the authors point out the development of hybrid optimiza- ing the proposed EDM in terms of efficiency, convergence
tion techniques, including the advantages of heuristics and and robustness.
analytical techniques, also mitigating their limitations by The remainder of this paper is organized as follows: Sect. 3
the hybridization process. In [34], the main advances in shows the distribution generation allocation problem that will
the area are chronologically presented. The authors high- be optimized. In Sect. 4, the proposed algorithm (EDM)
light that the major part of published papers makes use of is detailed. Section 5 presents the aspects related to the
heuristic optimization techniques as they are adequate to computer implementation of the system and the proposed
multimodal problems. Also, the authors conclude that there method. Section 6 presents the promising results obtained
is a vast opportunity to investigate and to improve heuris- using the proposed method, especially considering a large-
tic techniques. In [35], the authors suggest the development scale system, together with the discussions. Section 7 draws
of algorithms capable of tracking the global optimum of the the conclusions of the paper.
problem and compare different algorithms regarding accu-
racy, robustness, convergence and computational effort. In
[36], there is a conflict of choice among analytical and heuris- 2 Optimization problem
tic techniques. So, the authors highlight hybridization as the
most promising way of solving the DG siting problem. The The DG allocation is an optimization problem that can be
same conclusion about hybridization can be found in the formulated in different ways, depending on the kind of tech-
review survey [37]. nology under evaluation, the system constraints and desired
The main objective of the present paper is to propose an objectives. This paper analyzes the allocation of multiple real
efficient optimization approach to allocate DG units in three- power generation in three-phase unbalanced radial networks,
phase unbalanced networks, following the paths pointed out aiming to minimize the power losses and keeping the voltage
as future research in the bibliographical reviews. To han- levels at acceptable levels [41]. The DG technology consid-
dle the three-phase unbalanced power flow, the open-source ered are PV systems modeled in OpenDSS without voltage
software OpenDSS is employed together with MATLAB. controls [42].
The allocation problem is divided into two parts: (a) the sit- Given the operational conditions of the distribution sys-
ing and (b) the sizing of DG units. This division allows to tem (load and generation profiles), OpenDSS returns the total
make use of two different dedicated algorithms, taking the real power losses PL in the network. The three-phase opti-
best of each one, following the promising ways pointed by mization problem formulation is described by Eqs. (1) to (6),
aforementioned state-of-art papers. This results in an algo- where Eq. (1) represents the objective function Fobj . To keep
rithm with a greater capacity to find good solutions. In this the voltages at desired levels, a penalty factor λ, fixed in 107 ,
paper, the DG unit locations are chosen by using the pro- is added to the objective function. This λ value was empiri-
posed empirical discrete metaheuristic (EDM) and the units cally obtained, following the strategy presented in [43].
size is defined by the steepest descent method (SD). ⎛ ⎞
The development of the EDM is based on the prac-   φ 2
tical experience with the DG allocation problem. As the Min Fobj  ⎝ PL + λ V j − V lim ⎠ (1)
genetic algorithm [38], PSO [39] and bat algorithm [40], j∈N B φ∈Ph
the proposed technique is also based on a population of indi-
viduals that represent possible solutions to the problem. The Subject to
EDM scans the search space through interaction among its φ φ φ
 φ
individuals combined with random solutions to explore the DG j · PDG j + PG j − PD j + f P jk  0, φ ∈ Ph, j ∈ N B
k∈N B
allocation possibilities. This process reduces the likelihood
(2)
of stagnating in local optima.
φ φ
 φ
The main contribution is a new method with an origi- QG j − Q Dj + f Q jk  0, φ ∈ Ph, j ∈ N B (3)
nal search mechanism, which provides high-quality solutions k∈N B
to the DG allocation problem and a low number of adjust- 0, without DG
DG j  j ∈ N B, (4)
ment parameters required by the EDM. Only four parameters 1, with DG

123
 Electrical Engineering

max
0 ≤ PDG j ≤ PDG (5) an exhaustive enumeration in large-scale high-dimension
⎧ systems, especially when considering the allocation of mul-
⎪ min , if V φ is lower thanV min tiple DG units in large distribution systems. Additionally,
⎨ V j
max φ
V lim  V , if V j is greater than V max , φ ∈ Ph, j ∈ N B metaheuristics approaches are suitable for addressing multi-
⎪
⎩ φ
V j , otherwise modal problems because those techniques have mechanisms
(6) to escape from local optima, thus better exploring the search
space. As optimization algorithms are highly dependent on
Equations (2) and (3) represent the active and reactive balance the initial solutions [44], classical optimization methods may
in the system buses, respectively. Equation (4) is the DG get stuck in an optimum local region.
status in each bus, while Eq. (5) represents the generation A flowchart of the proposed metaheuristic is depicted in
limits of the DG units. Equation (6) presents the constraints to Fig. 2. In the highlighted steps, OpenDSS is called by the
keep voltages within proper values. In summary, the decision MATLAB (Fig. 1) to calculate the unbalanced three-phase
variables are the DG status DG j , handled by EDM, and the power flow to evaluate the fitness function Fobj , shown in
φ
amount of real power dispatched by phase PDG j , which is Eq. (1). The area delimited in the flowchart represents the
determined by the SD. local search.
Initially, all the parameters of the EDM are defined, and an
initial random population is created. This population is eval-
3 Proposed methodology uated regarding the objective function Fobj . After this step,
N individuals are distributed in G groups. The distribution
The general approach proposed to deal with the DG alloca- is done once, such that each group receives individuals with
tion problem is presented in Fig. 1. The optimization tools good fitness and some not so well evaluated, which creates
are implemented in MATLAB, while a Three-phase Power a diversified set of solutions in each group. Then, the best
Flow (TPF) is carried out by OpenDSS. individual in each group is directed to the Xbests matrix and
The optimization problem has two decision variables the best individual x∗ is identified. Concerning the DG allo-
types: cation problem, all EDM individuals are vectors with binary
values corresponding to the integer decision variables in the
• DG j stands for the DG connection status (0 or 1) at bus j. optimization problem, as shown in Eq. (7).
The proposed metaheuristic (EDM) deals with this deci- After those first steps, the algorithm begins the itera-
sion variable through a binary vector. Follows an example tive process. The step of generating new solutions occurs
of EDM proposed solution with 3 DGs in a hypothetical according to Eq. (8):
10 bus network:
t+1
xn,g  xn,g
t
⊕ Xbests (g) ⊕ A1 (8)
(7)
New solutions are created based on the combination of one
In this example, there is DG connection in buses 4, 5 and individual with the best from its group Xbests (g) and with a
10. Thus, DG 4  DG 5  DG 10  1, and DG 1  random vector A1 containing a possible DG configuration.
DG 2  DG 3  DG 6  DG 7  DG 8  DG 9  0. The process of merging solutions is exemplified by Eq. (9).
Since EDM is population-based and divided into groups,
t
x n,g is the solution n, belonging to group g, at iteration t
(EDM’s search mechanism will be further explained).
• PDG j stands for the injected power by the existing DG at
bus j. This decision variable is handled by the SD method,
executed for every EDM’s binary vector proposal. It is
valid to point out that this injection is meant to be balanced,
even when there are differences in bus j phase voltages.
The OpenDSS software guarantees this for a wider range
than we are restricting, which is from 0.93 to 1.05 pu (based
on the Brazilian regulatory limits for these test systems
voltage levels).

(9)
The reason for using metaheuristics in the solution of the
integer part of the problem is due to the difficulty of using

123
Electrical Engineering

Fig. 1 Proposed DG allocation

scheme

Fig. 2 Flowchart of the EDM

t+1 , two operations are made with the

After obtaining xn,g DG unit at that point. Thus, higher numbers are forced to
resulting solution. The maximum value that each position be 1. This is shown in Eq. (10), called “O peration1”. This
(bus) receives must be 1, representing the connection of a hypothetical EDM’s proposal of allocation indicates DGs at

123
 Electrical Engineering

buses 3, 4, 7, 8, 12, 14, 15, 17 and 19, yet the maximum

number of generation units is four.

(10)

Then, the resulting vector can have a number of connec- Back to the general algorithm, the last step is the update
tions higher than expected. As presented in the example of of the best individuals to repeat the search process while the
Eq. (9), only four DG units can be placed. To solve this prob- convergence criterion has not been achieved.
lem, the “1” buses are randomly erased to “0” until they As stated, each EDM solution has information about
reach the desired number of DG connections. This is called where are the DGs. The DG power injection PDG j is defined
“Operation 2”. In the above example, the resulting individual by the SD method, which is embedded in the EDM algorithm
corresponds to DG units allocated in buses 4, 7, 15, and 17. presented in Fig. 2. This method is an iterative optimization
The next step of the algorithm is to evaluate new individuals. algorithm based on partial derivatives of the objective func-
Thus, OpenDSS is called by MATLAB (Fig. 1), and the SD tion regarding the decision variables [45].
defines the size of each DG unit.

(11)

Regarding the local search procedure, if the best solution

x∗ does not change during a previously defined number of The main searching feature of EDM is within Eq. (8),
where all individuals xn,gt can be deepely transformed in a
iterations, a small variation is applied to it, updating the indi-
vidual if a better solution is found. Equation (12) shows how single step of the algorithm. This possibility provides extraor-
the local search creates a new individual. In this example, the dinary diversification in the population in a single iteration.
resulting individual has DGs in buses 3, 6, 8, 11 and 17 and From the perspective and experience of the authors, this is
has passed through O peration2, randomly deleting connec- crucial to sitting DG successfully. The results in the next
tion number 8. It must be highlighted that the new provisory section support this claim. Also, the employment of the SD
individual xpr is always close to the original one as this is a method appears to be really effective in sizing DGs, since this
local search procedure. problem presents a quadratic aspect [46]. Thus, the presented

(12)

123
Electrical Engineering

hybridization of EDM and SD seeks the positive character- Table 1 EDM chosen parameters
istics of both when facing DG allocation problem. Parameter Value Parameters description

3.1 Dealing with Constrains N 20 Number of individuals

G 4 Number of groups
The EDM itself has a mechanism to deal with a special kind of IT E 20 Maximum number of iterations
constraint, related to limited resources. In O peration1 and ST AG 6 Stagnation parameter
O peration2, the available number of DG units is taken into
account. This feature could be generalized in other real opti-
mization problems, such as transmission planning, whereas
Table 2 SD and Golden Section Search chosen parameters
the number of reinforcement lines is limited. In fact, real opti-
mization problems commonly present some kind of resource Parameter Value Parameters description
limitation, thus, the proposed mechanism of “cutting” parts I tmax 6 Maximum number of iterations
from a solution due to limited resources can be extended to δ 10e-4 Infinitesimal disturbance
other problems.
∈ 5 Tolerance to golden section converge
Concerning the SD, the search mechanism just stops at the
αmax 10e5 α Upper limit (golden section)
current point when it reaches some constraint. Thus, if the
αmin 0 α Lower limit (golden section)
power of a DG increases and reaches the maximum size (like
3 MW) that value is returned from the SD as the solution.
The same principle applies if the voltages reach a limit. As
the DG size increases close to an upper limit, there is a trend
in voltage rising, perhaps beyond the voltage limit on that So, the value of 1.5 MW has been chosen as the initial value
bus. So, a penalty factor λ is used to deal with the constraint. as it represents the mean generation between the DG power
limits.

4 Implementation

4.1 Test systems 4.2 EDM and SD parameters

The main simulations were conducted in the IEEE 34 bus The choice of EDM and SD parameters is also based on
and IEEE 123-bus system. Those systems have been chosen empirical tests, by evaluating the objective function sensitiv-
due to their high complexity and the high number of com- ity regarding a certain range of parameter values. Thus, a set
binations to be evaluated for siting the DG units. The first of different values for each parameter is tested, and those pro-
one has nominal voltages of 24.9 and 4.16 kV, depending on viding good objective function results, in a reasonable time,
the region of the circuit. The total attached load is approxi- are chosen. It can be noticed that the number of 20 individu-
mately 1.6 MVA with a lagging power factor at 0.98. In base als and 20 iterations are enough to achieve significant results
case (without DG), real losses are 221.8 kW. The 123-bus when compared to the technical literature, as can be seen in
feeder operates at 4.16 kV and has a power demand corre- the following section.
sponding to 3.4 MW and 1.2 MVAr, with base case losses The population is initially divided into four groups while
about 96 kW. Both systems have unbalanced loads in three the stagnation parameter value was set to six, that is, after
different models: constant power, constant current and con- six iterations without an upgrade in x∗ , a local search is per-
stant impedance. The 123-bus system has also four voltage formed. Table 1 shows the parameter used in the proposed
regulators and capacitor banks, as well as aerial and under- EDM.
ground distribution lines. The complete systems data can be The division of the initial population into four groups aims
found in [47]. to diversify the search without letting it be excessively ran-
All the tests have been conducted considering the alloca- dom. It was noted that for 20 individuals, a higher number
tion of three DGs in the systems. The results consider only of groups makes the solution worse, considering the conver-
the active power injection, limited to 3 MW for each DG. For gence of the problem. Also, the six iterations of the stagnation
example, those DG units can be photovoltaic panels or fuel parameter chosen are justifiable to avoid frequent interrup-
cells [41]. OpenDSS presents models for that generation, as tion of the primary search mechanism, thus letting the local
also for other kinds of technologies, thus allowing realistic search occur when necessary. The values used in the param-
simulations. To calculate the active power generation by each eters of the SD and the golden section search are presented
DG unit, the steepest gradient method needs an initial value. in Table 2 [45].

123
 Electrical Engineering

(a) (b)
Fig. 3 IEEE 34-bus voltage profile by distance from substation: (a) base case; (b) with DG

(a) (b)
Fig. 4 IEEE 123-bus voltage profile by distance from substation: (a) base case; (b) with DG

5 Results and Discussions Tables 3 and 4 exhibit the best configuration found by the
proposed methodology, considering three DGs inserted in
The proposed optimization approach has been used to allo- the two distribution networks. The summarized information
cate three DG units, injecting real power only into the IEEE is the chosen buses, power dispatched, losses, PLR % and
34-bus and IEEE 123-bus systems, which aims to minimize the optimization approach. As can be seen in both tables,
the active losses and to improve the voltage profile. Thus, the real power losses were severely reduced by proper DG
the results are presented concerning the power loss reduction allocation. For the 123-bus test system, there is a comparison
(PLR %), and the voltage profile. Further, the proposed meta- with works found in the specialized literature, showing that
heuristic is compared to other ones concerning efficiency and the proposed method is competitive, achieving better results
convergence behavior as well as robustness. than those published so far.

123
Electrical Engineering

Table 3 Results of 3DGs

allocation in IEEE 34 node test Chosen bus (Power - kW) Loss (kW) PLR Optimization
approach

Proposed 820 (159.3*) 844 (805.1) 890 (464.5) 18.45 91.7% EDM + Steepest
Methodology Descent
* Power injected only on bus 820 “Phase A”. This bus is single-phase

Table 4 Results of 3DGs allocation in IEEE 123 node test: Proposed methodology and literature
Chosen bus (kW) Loss (kW) PLR Optimization approach

Proposed Methodology 44 (1.037) 67 (1.827) 112 (0.145*) 19.87 79.5% EDM + Steepest Descent

[10] 44 (1.310) 64 (0.580) 86 (1.110) 22.8 76.5% Jaya Algorithm

[41] 47 (0.540) 67 (1.080) 72 (1.320) 29.5 69.1% Modified PSO
[48] 28 (0.200) 47 (0.880) 67 (2.380) 22.1 77.0% Sensitivity analysis + Quadratic curve
fitting
* Power injected only on bus 112 “Phase A”. This bus is single-phase

Figure 3 shows how the voltage profile was improved in and the proposed EDM were utilized to site three DG units
the IEEE 34-bus after DG unit connections at buses 820, 844 in the IEEE 34-bus and IEEE 123-bus systems, aiming to
and 890. The voltages by phase are presented according to the minimize the real power loss. For a fair comparison, all
distance from the substation. In the base case, there are many these methodologies used the same initial population con-
buses with voltages lower than the 0.95 pu limit, mostly from taining 20 individuals/solutions and the same number of 20
“Phase A”. It can be seen in Fig. 3b that after DG allocation, iterations. Table 5 presents all the parameters used in the
all voltages in all phases are between pre-established limits. simulations (EDM parameters are the same as offered in
Following the same voltage analysis that was done to the Table 1). All the values in Table 5 were chosen according
IEEE 34-bus, Fig. 4 depicts the voltage profile in the IEEE to the recommendations of the authors of each metaheuristic
123-bus for the base case and the case with the connection of and through experience with the DG allocation problem. The
DGs at buses 44, 67 and 112. Again, all the missing quality SOSA parameters are exactly the same presented in [49].
voltages were brought back to the recommended limits. In It is important to highlight the simplicity of the EDM
the presence of DG, the lower voltage is about 0.97 pu. related to the number of parameters when compared to other
metaheuristics. The parameters adjustment is one point of
high influence in the convergence of those techniques. The
5.1 Efficiency analysis higher the number of parameters, the more difficult the tuning
of a metaheuristic for a given problem. So, four parameters
This paper has also evaluated the efficiency of the pro- presented by the EDM represent a significant advantage of
posed optimization methodology in finding good quality this technique over others used here.
solutions by carrying out tests comparing the EDM to other The metaheuristics have been executed 100 times, starting
metaheuristics algorithms. It is important to emphasize that from the same set of initial solutions, for each test system.
only the metaheuristics are being compared, i.e., SD does Using a PC with a speed of 2.8 GHz and 4 GB of RAM,
the DG sizing for all metaheuristic algorithms used in this a single simulation takes 9 and 15 min to converge for the
paper, thus, similar to EDM in Fig. 1. The only exception IEEE 34-bus and IEEE 123-bus systems, respectively. For the
is the recently proposed symphony orchestra search algo- IEEE 34-bus, the final results of real power losses obtained
rithm (SOSA) [49], which is a multi-level music inspired by each metaheuristic are shown in a boxplot depicted in
metaheuristic technique. This algorithm has presented high Fig. 5. Some statistical data like median, mean and standard
performance solving large-scale, nonlinear and non-convex deviation are presented in Table 6.
optimization problems. The reason for not using SD sizing The same scheme of repetitions, starting from the same
strategy on this one is that its search mechanism has a particu- set of initial solutions, was carried out for the IEEE 123-bus
lar way to deal with integer and continuous decision variables system. The sets of final losses obtained by each metaheuris-
simultaneously. So, by comparing the proposed methodology tic are shown in the boxplot in Fig. 6. The median, mean and
with the SOSA, it can be pointed out how suitable are hybrid standard deviation from these sets can be seen in Table 7.
approaches for DG allocation problem. Moreover, the Wilcoxon signed rank test was performed
Bat algorithm (BA) [40], cuckoo search [50], firefly algo- in order to evaluate the differences previously presented.
rithm (FA) [51], particle swarm optimization (PSO) [39],

123
 Electrical Engineering

Table 5 Metaheuristics parameters

Metaheuristic Parameter Value Parameters description

BA n 20 Number of individuals
ite 20 Number of iterations
α 0.9 Volume decrease constant
γ 0.8 Emission rate increase
r0 1 Maximum rate of pulse
emission
f min 0 Minimum frequency
f max 0.8 Maximum frequency
CS n 20 Number of individuals
ite 20 Number of iterations
α 0.02 Step size
λ 1 Lévy distribution exponent
Pa 0,2 Abandon nest probability Fig. 5 Box plot—Metaheuristics comparison—IEEE 34-bus
FA n 20 Number of individuals
ite 20 Number of iterations Table 6 Statics data from DG allocation—IEEE 34-bus: Fobj
α 1 Random step weight
Metaheuristic Median (kW) Mean (kW) Standard deviation
γ 0.4 Light absorption coefficient (kW)
β0 0.1 Maximum attractiveness
δ EDM 18.52 18.62 0.43
0.98 Random step reduction rate
n BA 19.10 19.78 1.44
PSO 20 Number of individuals
ite CS 18.85 19.21 0.95
20 Number of iterations
c1 FA 19.45 19.90 1.37
0.3 Weight of local information
PSO 18.57 18.96 0.96
c2 0.7 Weight of global information
SOSA 19.25 19.71 1.09
θ 1 Velocity inertia
SOSA GI MPN 4 Number of groups of the
inhomogeneous music
players
PMN 10 Player memory number
PMS 250 Player memory size
PMCR 0.85 Player memory considering
rate
P A Rmin 0.2 Minimum pitch adjusting rate
P A Rmax 2 Maximum pitch adjusting rate
BWmin 0.4 Minimum bandwidth
BWmaxn 0.9 Maximum bandwidth
NIII 400 Number of iterations for the
second stage
NII 600 Number of iterations for the
third stage
NI 800 Number of iterations for the
fourth stage
Fig. 6 Box plot—Metaheuristics comparison—IEEE 123-bus

The results shown in this section reveal that the EDM

As shown in Table 8, the P-values corresponding to each is efficient to site DG units in unbalanced distribution net-
comparison of methodologies are negligible, meaning that works, finding better solutions more often and presenting
the null hypothesis (both data have equivalent mean) can be lower dispersion when compared to the other five well-
rejected as all P-values are far from the usual 5% significance succeed metaheuristics. For both test systems, the proposed
α-value [52]. metaheuristic obtained a set of solutions less dispersed and

123
Electrical Engineering

Table 7 Statics data from DG allocation—IEEE 123-bus: Fobj

Metaheuristic Median (kW) Mean (kW) Standard deviation
(kW)

EDM 20.31 20.33 0.32

BA 22.01 22.01 1.18
CS 20.96 21.02 0.62
FA 21.09 21.17 0.63
PSO 21.43 21.52 1.15
SOSA 21.31 21.34 0.27

Table 8 Wilcoxon signed rank test: P-value

Metaheuristic IEEE 34-bus IEEE 123-bus

EDM—BA 1.95e−12 8.74e−17

EDM—CS 1.07e−09 2.29e−13
EDM—FA 2.23e−14 4.48e−16 Fig. 7 Metaheuristics convergence track—IEEE 34-bus
EDM—PSO 1.49e−09 2.28e−15
EDM—SOSA 2.07e−15 6.12e−18

significantly closer to minimum power loss reached so far in

the technical literature.
At this point, it is noteworthy that the computational effort
and complexity of implementing two different algorithms is
worthy since good quality results were achieved. Also impor-
tant to mention that this problem is run offline, thus not in
real time.

5.2 Convergence analysis

A convergence analysis is essential to determine how an algo-

rithm evolves throughout the iterative process. As one of
the most significant advantages that metaheuristics have over
classical methods is avoiding getting stuck in local optima. Fig. 8 Metaheuristics convergence track—IEEE 123-bus
Tracking the best solution during the iterative process allows
an examination of this feature.
Figures 7 and Fig. 8 show the convergence track of the groups, containing ten musicians (40 individuals), and each
compared metaheuristics in one of the 100 repetitive simu- individual keeping 250 different melodies (solutions) in their
lations performed in the efficiency analysis. memory. As the number of initial solutions is higher, more
The convergence of EDM regarding the IEEE 34-bus of the search space is explored at the algorithm initialization.
test system occurs on the first iterations, showing the meta- Additionally, the number of iterations is not the same, as can
heuristic capacity to explore the search space quickly. After be seen in Table 5. If, on one hand, SOSA employs a larger
allocating DGs in the IEEE 123-bus system, the proposed number of individuals and iterations, on the other hand, it
algorithm found a good solution quickly. However, as the does not have to compute the objective function evaluations
search space was considerably larger, the method kept evolv- that SD requires in the other five metaheuristics. Aiming at
ing during the final iterations. a fair comparison, the total number of the objective function
The SOSA method has some characteristics to be men- calculations are essentially the same. While SOSA performs
tioned. The first one is the starting value obtained by this 82,000 objective function calculations, the other metaheuris-
algorithm, which is lower in comparison with the others. tics keep that value close to 80,500 (Due to the nature of
The reason for this is the values of the parameters used SD, its convergence may take different computational efforts
in SOSA, mainly the initial population, composed by four depending on the starting point). The SOSA track iterations

123
 Electrical Engineering

Fig. 9 EDM robustness regarding the initial population—IEEE 34-bus Fig. 10 EDM robustness regarding the initial population—IEEE 123-
bus

6 Conclusions
are normalized in 20 to be presented together with the other
five algorithms.
This paper presented a hybrid optimization algorithm to solve
the siting and sizing of DG units. The proposed approach
consists of separating the siting and sizing problems. The
first part is the DG siting, which is solved by a metaheuristic
5.3 Robustness analysis developed in this paper named empirical discrete metaheuris-
tic (EDM). The second part is the sizing of DG units, solved
The final analysis performed to evaluate the proposed meta- by the steepest descent method (SD). The EDM has only
heuristic is its capacity to produce good quality solutions with four parameters to be adjusted, thus making it simple to be
relative independence from the initial population (the initial applied to different problems when compared to other meta-
set of solutions). To undertake the robustness analysis, the heuristics. The simulations results confirm the potential of
EDM was executed 100 times starting from the same set of the presented approach, as it outperforms published results,
initial solutions (Fixed Initial Population—FIP) and another and other metaheuristic algorithms.
100 times starting from different random sets (Variable Ini- As future research, there is always space for improvements
tial Population—VIP). The obtained results for the IEEE 34 in metaheuristics algorithms through sensibility analysis;
node test are shown in a boxplot depicted in Fig. 9. thus, the potential of EDM can still be exploited. Different
It appears that the two sets are equivalent. However, to allocation scenarios, such as the number of DGs and different
find out more accurately, a Wilcoxon rank sum test was levels of load, may also contribute to the continuation of this
performed. This is a nonparametric hypothesis test for the research as well as the consideration multiobjective formula-
equality of means in two samples [52]. The returned P- tions. Also, the proposed metaheuristic has only been applied
value was 0.92. Thus, the null hypothesis was far from being to the DG allocation problem. Other optimization problems
rejected, using the 5% significance α-value. This test result can be addressed.
denotes robustness by EDM regarding the initial conditions.
In summary, there is a 92% probability that the similarities
between the two samples did not occur by chance and were
due to the robustness of the search mechanism present in the
References
algorithm. 1. Fanjun M, Chowdhury BH (2012) Economics of grid-tied
The same type of test was performed when attempting customer-owned photovoltaic power generation. In: 2012 IEEE
to allocate DGs in the IEEE 123 node test. The results are Power and Energy Society General Meeting. IEEE, pp 1–5
presented in Fig. 10. Another Wilcoxon rank sum hypothesis 2. Del Carpio-Huayllas TE, Ramos DS, Vasquez-Arnez RL (2012)
Feed-in and net metering tariffs: An assessment for their applica-
test was carried out, providing a 0.44 P-value. Thus, the null tion on microgrid systems. In: 2012 6th IEEE/PES Transmission
hypothesis cannot be rejected, assigning a certain level of and Distribution: Latin America Conference and Exposition (T&D-
robustness to EDM regarding the initial population. LA). IEEE, pp 1–6

123
Electrical Engineering

3. Ramírez FJ, Honrubia-Escribano A, Gómez-Lázaro E, Pham DT 19. Rueda-Medina AC, Franco JF, Rider MJ et al (2013) A mixed-
(2017) Combining feed-in tariffs and net-metering schemes to bal- integer linear programming approach for optimal type, size and
ance development in adoption of photovoltaic energy: comparative allocation of distributed generation in radial distribution systems.
economic assessment and policy implications for European coun- Electr Power Syst Res 97:133–143. https://doi.org/10.1016/j.epsr.
tries. Energy Policy 102:440–452. https://doi.org/10.1016/j.enpol. 2012.12.009
2016.12.040 20. Ugranlı F (2019) Analysis of renewable generation’s integration
4. Razavi S-E, Rahimi E, Javadi MS et al (2019) Impact of distributed using multi-objective fashion for multistage distribution network
generation on protection and voltage regulation of distribution sys- expansion planning. Int J Electr Power Energy Syst 106:301–310.
tems: a review. Renew Sustain Energy Rev 105:157–167. https:// https://doi.org/10.1016/j.ijepes.2018.10.002
doi.org/10.1016/j.rser.2019.01.050 21. Martinez JA, Guerra G (2014) A Parallel Monte Carlo Method
5. Rau NS, Wan Yih-Heui (1994) Optimum location of resources in for Optimum Allocation of Distributed Generation. IEEE Trans
distributed planning. IEEE Trans Power Syst 9:2014–2020. https:// Power Syst 29:2926–2933. https://doi.org/10.1109/TPWRS.2014.
doi.org/10.1109/59.331463 2317285
6. Rahmani-andebili M (2016) Distributed Generation Placement 22. Mahmoud K, Yorino N, Ahmed A (2016) Optimal Distributed
Planning Modeling Feeder’s Failure Rate and Customer’s Load Generation Allocation in Distribution Systems for Loss Minimiza-
Type. IEEE Trans Ind Electron 63:1598–1606. https://doi.org/10. tion. IEEE Trans Power Syst 31:960–969. https://doi.org/10.1109/
1109/TIE.2015.2498902 TPWRS.2015.2418333
7. Rahmani-andebili M (2016) Simultaneous placement of DG and 23. Zhang C, Li J, Zhang YJ, Xu Z (2018) Optimal Location Plan-
capacitor in distribution network. Electr Power Syst Res 131:1–10. ning of Renewable Distributed Generation Units in Distribution
https://doi.org/10.1016/j.epsr.2015.09.014 Networks: an Analytical Approach. IEEE Trans Power Syst
8. Ganguly S, Samajpati D (2017) Distributed generation allocation 33:2742–2753. https://doi.org/10.1109/TPWRS.2017.2749410
with on-load tap changer on radial distribution networks using 24. Pereira BR, Martins da Costa GRM, Contreras J, Mantovani JRS
adaptive genetic algorithm. Appl Soft Comput 59:45–67. https:// (2016) Optimal Distributed Generation and Reactive Power Alloca-
doi.org/10.1016/j.asoc.2017.05.041 tion in Electrical Distribution Systems. IEEE Trans Sustain Energy
9. Zhang S, Cheng H, Li K et al (2018) Multi-objective distributed 7:975–984. https://doi.org/10.1109/TSTE.2015.2512819
generation planning in distribution network considering correla- 25. Zhang S, Cheng H, Wang D et al (2018) Distributed genera-
tions among uncertainties. Appl Energy 226:743–755. https://doi. tion planning in active distribution network considering demand
org/10.1016/j.apenergy.2018.06.049 side management and network reconfiguration. Appl Energy
10. Kumawat M, Gupta N, Jain N, Bansal RC (2017) Optimally 228:1921–1936. https://doi.org/10.1016/j.apenergy.2018.07.054
Allocation of Distributed Generators in Three-Phase Unbalanced 26. Almabsout EA, El-Sehiemy RA, An ONU, Bayat O (2020) A
Distribution Network. Energy Procedia 142:749–754. https://doi. Hybrid Local Search-Genetic Algorithm for Simultaneous Place-
org/10.1016/j.egypro.2017.12.122 ment of DG Units and Shunt Capacitors in Radial Distribution
11. Bayat A, Bagheri A (2019) Optimal active and reactive power allo- Systems. IEEE Access 8:54465–54481. https://doi.org/10.1109/
cation in distribution networks using a novel heuristic approach. ACCESS.2020.2981406
Appl Energy 233–234:71–85. https://doi.org/10.1016/j.apenergy.2 27. Pinto RS, Unsihuay-Vila C, Fernandes TSP (2019) Multi-objective
018.10.030 and multi-period distribution expansion planning considering reli-
12. Hincapie IRA, Gallego RRA, Mantovani JRS (2019) A decompo- ability, distributed generation and self-healing. IET Gener Transm
sition approach for integrated planning of primary and secondary Distrib 13:219–228. https://doi.org/10.1049/iet-gtd.2018.5037
distribution networks considering distributed generation. Int J 28. Barati F, Jadid S, Zangeneh A (2019) Private investor-based dis-
Electr Power Energy Syst 106:146–157. https://doi.org/10.1016/ tributed generation expansion planning considering uncertainties
j.ijepes.2018.09.040 of renewable generations. Energy 173:1078–1091. https://doi.org/
13. Gangwar P, Singh SN, Chakrabarti S (2019) Multi-objective 10.1016/j.energy.2019.02.086
planning model for multi-phase distribution system under uncer- 29. Kansal S, Kumar V, Tyagi B (2016) Hybrid approach for optimal
tainty considering reconfiguration. IET Renew Power Gener placement of multiple DGs of multiple types in distribution net-
13:2070–2083. https://doi.org/10.1049/iet-rpg.2019.0135 works. Int J Electr Power Energy Syst 75:226–235. https://doi.org/
14. Ali ES, Abd Elazim SM, Abdelaziz AY (2016) Optimal allocation 10.1016/j.ijepes.2015.09.002
and sizing of renewable distributed generation using ant lion opti- 30. Gao Y, Hu X, Yang W et al (2017) Multi-Objective Bilevel Coordi-
mization algorithm. Electr Eng 100:99–109. https://doi.org/10.10 nated Planning of Distributed Generation and Distribution Network
07/s00202-016-0477-z Frame Based on Multiscenario Technique Considering Timing
15. Moghaddam MJH, Kalam A, Shi J et al (2020) A New Model for Characteristics. IEEE Trans Sustain Energy 8:1415–1429. https://
Reconfiguration and Distributed Generation Allocation in Distri- doi.org/10.1109/TSTE.2017.2680462
bution Network Considering Power Quality Indices and Network 31. Das B, Mukherjee V, Das D (2019) Optimum DG placement for
Losses. IEEE Syst J. https://doi.org/10.1109/jsyst.2019.2963036 known power injection from utility/substation by a novel zero bus
16. Sanjay R, Jayabarathi T, Raghunathan T et al (2017) Optimal load flow approach. Energy 175:228–249. https://doi.org/10.1016/
Allocation of Distributed Generation Using Hybrid Grey Wolf j.energy.2019.03.034
Optimizer. IEEE Access 5:14807–14818. https://doi.org/10.1109/ 32. Selim A, Kamel S, Jurado F (2020) Efficient optimization tech-
ACCESS.2017.2726586 nique for multiple DG allocation in distribution networks. Appl
17. Chang GW, Cong Chinh N (2020) Coyote Optimization Algorithm- Soft Comput 86:105938. https://doi.org/10.1016/j.asoc.2019.1059
Based Approach for Strategic Planning of Photovoltaic Distributed 38
Generation. IEEE Access 8:36180–36190. https://doi.org/10.1109/ 33. Viral R, Khatod DK (2012) Optimal planning of distributed gen-
ACCESS.2020.2975107 eration systems in distribution system: a review. Renew Sustain
18. Moravej Z, Akhlaghi A (2013) A novel approach based on cuckoo Energy Rev 16:5146–5165. https://doi.org/10.1016/j.rser.2012.05.
search for DG allocation in distribution network. Int J Electr Power 020
Energy Syst 44:672–679. https://doi.org/10.1016/j.ijepes.2012.08. 34. Georgilakis PS, Hatziargyriou ND (2013) Optimal Distributed
009 Generation Placement in Power Distribution Networks: mod-

123
 Electrical Engineering

els, Methods, and Future Research. IEEE Trans Power Syst 45. Luenberger DG, Yinyu Y (2008) Linear and Nonlinear Program-
28:3420–3428. https://doi.org/10.1109/TPWRS.2012.2237043 ming, 3rd edn. Springer, New York
35. Rezaee Jordehi A (2016) Allocation of distributed generation units 46. Anwar A, Pota HR (2011) Loss reduction of power distribution
in electric power systems: a review. Renew Sustain Energy Rev network using optimum size and location of distributed generation.
56:893–905. https://doi.org/10.1016/j.rser.2015.11.086 Universities Power Engineering Conference (AUPEC), 2011 21st
36. Pesaran HAM, Huy PD, Ramachandaramurthy VK (2017) A Australasian. IEEE, Brisbane, pp 1–6
review of the optimal allocation of distributed generation: objec- 47. IEEE PES Distribution System Analysis Subcommittee’s, Distribu-
tives, constraints, methods, and algorithms. Renew Sustain Energy tion Test Feeder Working Group Distribution Test Feeders. http://
Rev 75:293–312. https://doi.org/10.1016/j.rser.2016.10.071 ewh.ieee.org/soc/pes/dsacom/testfeeders/index.html. Accessed 3
37. Singh B, Sharma J (2017) A review on distributed generation plan- Apr 2019
ning. Renew Sustain Energy Rev 76:529–544. https://doi.org/10.1 48. Anwar A, Pota HR (2012) Optimum capacity allocation of DG
016/j.rser.2017.03.034 units based on unbalanced three-phase optimal power flow. In: 2012
38. Holland JH (1975) Adaptation in natural and artificial systems: IEEE Power and Energy Society General Meeting. IEEE, pp 1–8
An introductory analysis with applications to biology, control, and 49. Kiani-Moghaddam M, Shivaie M (2017) An Innovative
artificial intelligence., 1st ed. London, England Multi-Stage Multi-Dimensional Multiple-Inhomogeneous Melody
39. Kennedy J, Eberhart R (1995) Particle swarm optimization. In: Search Algorithm: Symphony Orchestra Search Algorithm
Proceedings of ICNN’95 - International Conference on Neural Net- (SOSA). pp 1–40
works. IEEE, pp 1942–1948 50. Yang X-S, Suash Deb (2009) Cuckoo Search via Lévy Flights. In:
40. Yang X-S (2010) A New Metaheuristic Bat-Inspired Algorithm. 2009 World Congress on Nature & Biologically Inspired Comput-
In: Studies in Computational Intelligence. pp 65–74 ing (NaBIC). IEEE, pp 210–214
41. Dahal S, Salehfar H (2016) Impact of distributed generators in the 51. Yang X-S (2009) Firefly Algorithms for Multimodal Optimization.
power loss and voltage profile of three phase unbalanced distribu- In: Proceedings of the 5th International Conference on Stochastic
tion network. Int J Electr Power Energy Syst 77:256–262. https:// Algorithms: Foundations and Applications. pp 169–178
doi.org/10.1016/j.ijepes.2015.11.038 52. Montgomery DC, Runger GC (2014) Applied Statistics and Prob-
42. Dugan RC, Institute EPR (2016) OpenDSS Manual. Train Mater ability for Engineers. Sixth Edit. John Wiley & Sons Inc, Danvers
1–184
43. Abido MA (2002) Optimal power flow using particle swarm opti-
mization. Int J Electr Power Energy Syst 24:563–571. https://doi.
Publisher’s Note Springer Nature remains neutral with regard to juris-
org/10.1016/S0142-0615(01)00067-9
dictional claims in published maps and institutional affiliations.
44. Bazaraa M, Sherali H, Shetty C (2006) Nonlinear Programming
Theory and Algorithms

123