You are on page 1of 15

1 3

Neural Computing and Applications



ISSN 0941-0643

Neural Comput & Applic
DOI 10.1007/s00521-012-0976-4
Hybrid neuro-swarm optimization
approach for design of distributed
generation power systems
T.Ganesan, P.Vasant & I.Elamvazuthi
1 3
Your article is protected by copyright and all
rights are held exclusively by Springer-Verlag
London Limited. This e-offprint is for personal
use only and shall not be self-archived in
electronic repositories. If you wish to self-
archive your work, please use the accepted
authors version for posting to your own
website or your institutions repository. You
may further deposit the accepted authors
version on a funders repository at a funders
request, provided it is not made publicly
available until 12 months after publication.
ORI GI NAL ARTI CLE
Hybrid neuro-swarm optimization approach for design
of distributed generation power systems
T. Ganesan

P. Vasant

I. Elamvazuthi
Received: 27 August 2011 / Accepted: 21 May 2012
Springer-Verlag London Limited 2012
Abstract The global energy sector faces major chal-
lenges in providing sufcient energy to the worlds ever-
increasing energy demand. Options to produce greener,
cost effective, and reliable source of alternative energy
need to be explored and exploited. One of the major
advances in the development of this sort of power source
was done by integrating (or hybridizing) multiple different
alternative energy sources (e.g., wind turbine generators,
photovoltaic cell panels, and fuel-red generators, equip-
ped with storage batteries) to form a distributed generation
(DG) power system. However, even with DG power sys-
tems, cost effectiveness, reliability, and pollutant emissions
are still major issues that need to be resolved. The model
development and optimization of the DG power system
were carried out successfully in the previous work using
particle swarm optimization (PSO). The goal was to min-
imize cost, maximize reliability, and minimize emissions
(multi-objective function) subject to the requirements of
the power balance and design constraints. In this work, the
optimization was performed further using Hopeld neural
networks (HNN), PSO, and HNN-PSO techniques. Com-
parative studies and analysis were then carried out on the
optimized results.
Keywords Particle swarm optimization (PSO)
Hopeld neural networks (HNN) Multi-objective
Optimization strategy Hybrid algorithms Alternative
energy Distributed generation (DG)
List of symbols
COST ($/year) Total cost
w, s, b Wind, solar, and battery
storage indices
I
i
, S
pi
, OM
pi
($/year) Initial cost, present worth
of salvage value, present
worth of operations, and
maintenance cost
N
p
(year) Lifespan of the project
C
g
Annual cost of
purchasing power from
the utility grid
a
w
, a
s
, a
b
($/m
2
) Initial cost of WTG, PV
panels, and storage
battery
A
w
, A
s
(m
2
) Swept area of WTG and
PV panels
S
w
, S
s
($/m
2
) Salvage value of WTG
and solar per square
meter
b, T, m Ination rate, interest
rate, and escalation rate
a
OMw
, a
OMs
, a
OMb
($/m
2
/year) Yearly operation and
maintenance cost for
wind, solar, and storage
batteries
N
p
, N
w
, N
s
, N
b
(year) Lifespan of project,
WTG, PV, and storage
batteries
g
s
, g
w
, g
b
Efciency of PV,
WTG, and storage
batteries
P
g, t
(kW) Purchased power from
the utility at hour t
psi ($/kW h) Grid power price
T. Ganesan P. Vasant (&) I. Elamvazuthi
University Technology Petronas, Tronoh, Malaysia
e-mail: pvasant@gmail.com
1 3
Neural Comput & Applic
DOI 10.1007/s00521-012-0976-4
Author's personal copy
EIR Energy index of
reliability
EENS (kW h/year) Expected energy not
served
E (kW h/year) Total power demand per
annum
k Ratio of purchased power
with respect to the hourly
insufcient power
PE (ton/year) Pollutant emission
X, u, U Coefcients
approximating the
generator emission
characteristic coefcients
P
bcap
(kW) Capacity of storage
batteries
P
bsoc
(kW) State of charge of storage
batteries
P
bmax
(kW) Maximum conversion
capacity
P
bmin
(kW) Minimum permissible
storage level
P
bcapmax
(kW) Allowed storage capacity
P
br
(kW) Rated battery capacity
P
b
(t) (kW) Discharge power from
the storage batteries
P
gmax
(kW) Maximum annual power
allowed to be purchased
from the utility grid
P
gmin
(kW) Minimum annual power
allowed to be bought
from the utility grid
T (h) Period under observation,
8,760 h (per year)
P
bsup
(t) (kW) Surplus power at hour t
P
d
(t) (kW) Load demand during
hour t
P
total
(t) (kW) Total power from WTG,
PV, and FFG
P
g
(kW) Power from the FFG
P
w
(kW) Power from the WTG
P
s
(kW) Power from the PV
R Ratio of maximum
permissible unmet power
P
dump
(kW) Dumped power
P
WTG
(kW) Output power from th
WTG
V, V
ci
, V
r
, V
co
(m/s) Wind speed, cut-in wind
speed, rated wind speed,
and cutoff wind speed
P
r
(kW) Rated WTG power
A
wmax
, A
wmin
(m
2
) Maximum and minimum
swept area of WTGs
A
smin
, A
smax
(m
2
) Minimum and maximum
swept area of PVs
1 Introduction
In recent times, the global energy sector faces two major
challenges in providing sufcient energy to the worlds
ever-increasing energy demand. First, there is a growing
need to produce greener and cleaner energy with respect to
stricter environmental regulations. Secondly, with the
diminishing fossil fuel reserves, a reliable and stable source
of alternative energy needs to be explored and exploited. In
seeking out these alternative power sources, it has been
identied that the capital investment as well as the main-
tenance costs are considerably high. Besides that, various
reliability issues have been addressed over the years. One
of the major advances, in developing a reliable and greener
[1] power source is by integrating or hybridizing multiple
different energy sources (e.g., wind turbine generators,
photovoltaic cell panels, and storage batteries) to form a
DG power system. These hybrid power generation systems
have been built and are now in stable operations [24].
However, even with DG power systems, cost effectiveness,
reliability [5, 6], and pollutant emissions are still major
issues that need to be tackled. Therefore, to address the
previously mentioned issues, PSO methods have been
applied to the problem by Wang et al. [7]. Other works on
the design and sizing of hybrid power systems with solar
and wind power sources include Chedid et al. [8] and
Chedid et al. [9].
In this work, optimization methods such as HNN and
PSO (stand-alone and hybridized form) were incorporated
into this problem. Comparison studies as well as result
analysis were then performed to identify the best optimi-
zation strategy that achieves all the objectives and obeys all
the power balance and design constraints.
The HNN was developed in 1982 by Hopeld [10, 11].
These neural nets observed to have applications in opti-
mization problems (for instance, see Lee, Sode-Yome et al.
[12] and Tank et al. [13]). One of the key features of the
HNN is that there is a decrease in the energy by a nite
amount whenever there is a change in the networks state.
This essential property conrms convergence of the output
whenever the network state is changed. The HNN uses
reinforced learning (Hebbian learning) to update the
weights in each recursion. In this work, the HNN was used
as an optimization algorithm.
Neural Comput & Applic
1 3
Author's personal copy
PSO is an optimization method developed based on the
movement and intelligence of swarms. PSO integrates the
concept of social interaction to problemsolving and decision
making. PSO was developed by James Kennedy and Russell
Eberhart [14] in 1995. Particle swarmis the systemmodel or
social structure of a basic creature which makes a group to
have some objectives such as food searching and predator
prey interactions. Hence, the governing principle is that it is
an important to take part with the most of the population in a
group that has the same activity. Recently, PSO has been
applied to various elds of power system including eco-
nomic dispatch problems as well as in optimization problems
in electric power systems (see [15]).
This paper is organized as follows: Sect. 2 introduces
the HNN, PSO and the HNN-PSO approaches; Sect. 3
presents the problem description for the optimization of the
DG system; and the analysis and computational results are
included in Sect. 4. The paper ends with concluding
remarks and recommendations for future research work.
2 Methodology
2.1 Hopeld neural network
The HNN consists of different components which are the
inputs, outputs, and weights (see Fig. 1). The other two
crucial subcomponents of the HNN are the Hebbian
learning mechanism and the energy function. The inte-
gration of all these features in the HNN makes it a good
optimization tool.
The outputs of the HNN are computed by the compo-
sition of the inputs and the associated weights such as the
following:
h
i
=

j
w
il
x
j
(1)
where x
j
is the input column vector, h
i
is the output column
vector, and w
ij
are the weights. Since this is a recurrent
network, hence the outputs are fed back as the inputs:
x
T
j
_ _
m1
= w
ij
y
T
i
_ _
m
(2)
where m is the number of iterations.
Due to its recursive nature, the inputs and outputs
change with respect to the number of iterations. Thus, the
modied Hebbian learning (a form of reinforced learning)
is used to alter the weights based on the network outputs
and the inputs:
w
ij
= x
T
j
y
i
_ _
h
(3)
where h is the learning rate coefcient.
The energy function is computed as a sum for weights
and neurons for all i and j for each of the iterations:
E =

ij
k w
ij
x
j
y
T
i
_ _
(4)
where k is a scalar constant.
Since the energy of the network reduces nitely as the
network states change, hence as the number of iterations
increases to its maxima, the differential of the energy
between the states approximates to 0:
As m max; DE 0
DE = E
m1
E
m
(5)
where m is the number of iterations.
At this point a convergence criterion is set, whereby if
the differential energy DE is lesser than some value then
the program is halted and solution (outputs of the network,
h
i
) is printed out. Otherwise, the iterations continue until
this criterion is satised. The differential energy continu-
ously minimizes as the neurons states change. Some
threshold value is required to avoid the energy differential
to minimize ad innitum. Hence, in this work, a minimal
differential energy criterion is set (dE\1). Since at these
states the difference between the energy states are minimal,
the network ceases from performing further optimization.
The working algorithm of the HNN is as the following:
Step 1: Set x
i
as the inputs
Step 2: Initialize neural network weights, w
ij
Step 3: Compute neural network output, y
j
Step 4: Compute an energy state of the neural network, E
ij
Step 5: If the energy difference dE
ij
is greater than 1, go
to step 2 and
Update neural network output, y
j
= y
/
j
Update the weights by Hebbian learning, w
ij
= w
/
ij
Step 6: If energy difference is less than 1, proceed to step 7
Step 7: Initialize DG system coefcients and compute
design parameters
Step 8: If any constraints are not satised, go to step 2,
otherwise proceed to step 9.
Fig. 1 Hopeld recurrent neural network
Neural Comput & Applic
1 3
Author's personal copy
Step 9: Compute tness values of the design parameters.
If tness criterion is satised then halt program and print
solutions, otherwise, go to step 2.
The stopping criterion used in this work is identical to
the tness criterion. If the tness criterion is not met, then
the program proceeds iteratively, otherwise it halts and
prints the solutions. The tness criterion of the HNN is met
if the network output converges to some constant value
(which means that no further optimization occurs in the
objective function), no constraints are broken, and all the
decision variables are nonnegative. If these conditions are
met, then it is considered that the solutions are at its ttest,
and thus, the program comes to a halt.
2.2 Particle swarm optimization
The PSO algorithm was initially developed in 1995 by
Kennedy and Eberhart [14]. This technique originates from
two different ideas. The rst idea was based on the obser-
vation of swarming or ocking habits of certain types of
animals (for instance: birds, bees, and ants). The second
concept was mainly related to the study of evolutionary
computation. The PSO algorithm works by searching the
search space for candidate solutions and evaluating them to
some tness function with respect to the associated criterion.
The candidate solutions are analogous to particles in
motion (swarming) through the tness landscape in search
for the optimal solution. Initially, the PSOalgorithmchooses
some candidate solutions (candidate solutions can be ran-
domly chosen or be set with some a priori knowledge). Then,
each particles position and velocity (candidate solutions)
are evaluated against the tness function. If the tness
function is not satised, then update the individual and social
component with some update rule. Next, the velocity and the
position of the particles are updated. This procedure is
repeated iteratively until all candidate solutions satisfy the
tness function and thus converge into a x position. The
owchart for the PSO algorithm is given in Fig. 2.
It is important to note that the velocity and position
updating rule is crucial in terms of the optimization capa-
bilities of the PSOalgorithm. The velocity of each particle in
motion (swarming) is updated using the following equation:
v
i
(n 1) = wv
i
(n) c
1
r
1
[^ x
i
(n) x
i
(n)[
c
2
r
2
[g(n) x
i
(n)[ (6)
where each particle is identied by the index i, v
i
(n) is the
particle velocity, and x
i
(n) is the particle position with
respect to iteration (n). The parameters w, c
1
, c
2
, r
1
, and r
2
are usually dened by the user. These parameters are
typically constrained by the following closed intervals:
w[0;1:2[; c
1
[0;2[; c
2
[0;2[; r
1
[0;1[; r
2
[0;1[ (7)
The term wv
i
(n) in Eq. 6 is the inertial term which keeps
the particles moving in the same direction as its original
direction. The inertial coefcient w serves as a dampener or
an accelerator during the particles motion. The term
c
1
r
1
[^ x
i
(n)x
i
(n)[ also known as the cognitive component
functions as memory. Hence, the particle tends to return to
the location in the search space where the particle had a
very high tness value. The term c
2
r
2
[g(n)x
i
(n)[ known
as the social component serves to move the particle to the
locations where the swarm has moved in the previous
iterations.
In the rst loop in Fig. 2, the constant n
0
is introduced.
The idea behind this loop is to prime the particles in motion
Stop
YES
NO
Is fitness criterion
satisfied?
Evaluate fitness of the
design parameters.
YES
Initialize power system
coefficients and design
parameters
Is n > n
o
+T ?
T = T +1
n = n +1
NO
Compute individual and
social influence
Compute position x
i
(n+1)
and velocity v
i
(n+1) at next
iteration
START
Initialize no of
particles, i
Initialize algorithm parameters w,
c
1
,c
2
, r
1
, r
2
,n
o
Set initial position
x
i
(n) and velocity v
i
(n)
Fig. 2 Flowchart for PSO algorithm
Neural Comput & Applic
1 3
Author's personal copy
prior to applying them into the DG system. The larger the
value of n
0
, the longer the particles would be in motion
prior their application and evaluation. Thus, as the sec-
ondary loop (T = T ? 1) increases, the primary loop
(n = n ? 1) also increases. Since, the primary loop keeps
the particles in motion incrementally as the program is
executed, thus the particles gradually converge to some
constant position. This primary loop strengthens the algo-
rithm in terms of its convergent properties.
After the computation of the particle velocity, the par-
ticle position is then calculated as follows:
x
i
(n 1) = x
i
(n) v
i
(n 1) (8)
The iterations are then continued until all candidate
solutions are at their ttest positions in the tness
landscape and some stopping criterion which is set by the
user is met. For more comprehensive texts on PSO
methods, refer to [16, 17] and [18]. The working
algorithm of the PSO in this work is as follows:
Step 1: Initialize no of particles, i and the algorithm
parameters w, c
1
, c
2
, r
1
, r
2
, n
o
Step 2: Set initial position x
i
(n) and velocity v
i
(n)
Step 3: Compute individual and social inuence
Step 4: Compute position x
i
(n ? 1) and velocity
v
i
(n ? 1) at next iteration
Step 5: If the swarm evolution time is n [n
o
? T,
update position x
i
and velocity v
i
and go to Step 3, else
proceed to step 6, where n
o
is some constant, n is the
swarm iteration, and T is the overall program iteration
Step 6: Initialize DG system coefcients and design
parameters
Step 7: Evaluate tness of the design parameters
Step 8: If tness criterion is satised, halt and print
solutions, else go to step 3.
As for the PSO algorithm, the tness criterion used is
as follows:
If during the iteration process, the position of all the
particles converges to some constant value, no further
optimization occurs in the objective function, no
constraints are broken, and all the decision variables
are nonnegative then, it can be considered that the
tness criterion is met. Then, the solutions are at its
ttest, and thus, the program comes to a halt.
2.3 Hybrid neuro: particle swarm optimization
In this work, the NN and the PSO algorithm was hybridized
and used as an alternative optimization tool in line with the
stand-alone NN and the PSO algorithms. The working
algorithm and the ow chart (see Fig. 3) of the hybrid
HNN-PSO technique are as follows:
Step 1: Set x
i
as the inputs and initialize neural network
weights, w
ij
Step 2: Compute network output, y
j
and energy differ-
ence, dE
Step 3: Check differential energy threshold: if (dE[1),
continue network recursion (Step 2)
else, proceed to step 4.
Step 4: Initialize no of particles, i and the algorithm
parameters w, c
1
, c
2
, r
1
, r
2
, n
o
Step 5: Set initial position x
i
(n) and velocity v
i
(n)
Step 6: Compute individual and social inuence
Step 7: Compute position x
i
(n ? 1) and velocity
v
i
(n ? 1) at next iteration
NO
T =T+1
Is n > no+T?
Compute position xi(n+1) and
velocity vi(n+1)
YES
Set initial conditions for the
PSO segment
Is dE >1?
NO
n =n+1
Initialize power system
coefficients and design
parameters
YES
Evaluate fitness of the design
parameters.
YES
Stop
Is fitness criterion
satisfied?
START
Set xi as the inputs and initialize
neural network weights, wij
Compute network output, yj &
energy difference, dE
NO
Update yj = xi
wij= wij
m = m +1
Fig. 3 Flowchart for hybrid HNN-PSO algorithm
Neural Comput & Applic
1 3
Author's personal copy
Step 8: If the swarm evolution time n [n
o
? T, update
position x
i
and velocity v
i
and go to Step 2, else, proceed
to step 6, where n
o
is some constant, n is the swarm
iteration, m is the network recursion, and T is the overall
program iterations
Step 9: Initialize DG system coefcients and design
parameters
Step 10: Evaluate tness of the design parameters
Step 11: If tness criterion is satised, halt and print
solutions, else go to step 3
As for the HNN-PSO algorithm, since the PSO segment
comes after the HNN segment, thus the tness criterion is
identical to the pure PSO algorithm. That is if during the
execution of the program, the position of all the particles
converges to some constant value, no further optimization
occurs in the objective function, no constraints are broken,
and all the decision variables are nonnegative then, it can
be considered that the tness criterion is met. Then, the
solutions are at its ttest, and thus, the program comes to a
halt.
3 Description of problem statement
The goal of this work is to optimize the design parameters
of a hybrid DG power system with alternative energy
power sources (solar and wind power) with respect to
power balance as well as design constraints as done pre-
viously in [7]. The problem in this work is multi-objective,
thus the design parameters would have to be optimized
such that it minimizes the cost, maximizes the reliability,
and minimizes pollutant emissions of the power system.
The conguration the grid-connected hybrid DG system is
as in Fig. 4.
The hybrid DG system consists of wind turbine gener-
ators (WTGs), photovoltaic cell panels (PVs), storage
batteries (SBs), and the fuel-red generators (FFGs). The
usage of each of these power sources inuences the reli-
ability, cost, and the environment criterions differently.
One of the cheaper fuel types that can be used for an
FFG would be coal. Coal is reliable, abundantly available,
and a relatively cheap fossil fuel source. The only major
drawback with fuel sources like coal is that they have a
high rate of pollutant emission (PE). Similar issues are
currently faced by other fossil fuel alternatives, for
instance: diesel, petrol, and natural gas (NG). Also take
note that other fossil fuels are not as cheap as coal, and
thus, cost effectiveness is an issue. The oxidation of these
fuel types produces alarming levels of pollutant gases such
as NO
x
, SO
x
, carbon monoxide, and carbon dioxide. Thus,
with the increasingly stricter enactment of environmental
regulation, sole dependency on FFGs for power is clearly
an unfeasible option. This is the main factor that motivates
the development of hybrid DG systems which reduces the
global dependency on fossil fuel.
One of the cleanest and cheapest power sources
(despite the initial cost) known is wind power since it has
no pollutant emissions, and wind power is available with
no purchase cost. However, the magnitude of wind power
is heavily dependent on weather conditions. Thus, the
location where the wind turbine is placed (on-shore or off-
shore locations) is a critical factor. Due to varying
weather conditions, the reliability of wind power is
intermittent, and this makes sole dependency on wind
power unfeasible. Unfortunately, solar power also suffers
similar issues with wind power. Sunlight (insolation) is
the main source of energy for PV cells. Like wind, this
power source is very dependent on weather conditions
although relatively cheap (besides the initial cost) and
clean (with zero emissions). This makes solar power an
unreliable energy source due to uctuations in weather
conditions.
Since wind and solar power are highly unreliable energy
source, thus including storage batteries into the DG system
is highly desirable. The energy storage mechanism can thus
lter-out the uctuations and give a consistent amount of
power supply with respect to time [19]. The storage bat-
teries then can be considered to behave like a regulator that
balances the supply and demand variability.
Each power source caters differently for reliability, cost
effectiveness, and pollutant emissions. Therefore, the
development of the hybrid DG system seems to be an
attractive option in catering for all three criterions
simultaneously.
This is a nonlinear problem that involves 9 constraints
and 67 decision variables. The problem statement is for-
mulated as follows:
Min ? COST ($/year)
Min ? Pollutant emissions (PE) (ton/year)
Max ? Energy index of reliability (EIR)
subject to power balance and design constraints.
WTG
PV
Other
Renewable
Energy
Sources
AC/DC
DC/DC
AC(DC)
/DC
Storage
Batteries
DC/AC
Dump
Load
Utility
Grid
Fig. 4 Schematic of the grid-connected hybrid DG system
Neural Comput & Applic
1 3
Author's personal copy
For the denitions of the variables, please refer to the
nomenclature above. The objective functions (refer to [3, 4,
8, 9]), for the overall cost, COST ($/year) is as the
following:
COST =

i=w;s;b
(I
i
S OM
P
i
)
N
p
C
g
(9)
For the WTG:
I
w
= a
w
A
w
(10)
S
P
w
= S
w
A
w
1 b
1 c
_ _
N
p
(11)
OM
P
w
= a
OM
w
A
w

N
p
i=1
1 m
1 c
_ _
i
(12)
For the PV:
I
s
= a
s
A
s
(13)
S
P
s
= S
s
A
s
1 b
1 c
_ _
N
p
(14)
OM
P
s
= a
OM
s
A
s

N
p
i=1
1 m
1 c
_ _
i
(15)
For the storage battery:
I
b
= a
b
P
bcap

X
b
i=1
1 m
1 b
_ _
(i1)=N
b
(16)
OM
P
b
= a
OM
b
P
bcap

N
p
i=1
1 m
1 c
_ _
i
(17)
The annual cost for purchasing power from the grid is
calculated as follows:
C
g
=

T
t=1
P
g;t
u (18)
The objective functions for the reliability (refer to [3, 4,
8, 9]) are as follows:
EIR = 1
EENS
E
(19)
EENS =

T
t=1
(P
bmin
P
bsoc
(t) P
sup
(t)) U(t) (20)
P
total
(t) = P
w
(t) P
s
(t) P
g
(t) (21)
P
g
(t) = j (P
d
(t) P
w
(t) P
s
(t) P
b
(t)) (22)
The objective function for the pollutant emissions which
was quadratically approximated (see [20, 21]) is as follows:
PE = X U

T
t=1
P
g;t
(t)
_ _
C

T
t=1
P
g;t
(t)
_ _
_ _
2
(23)
(A) Power balance constraints:
P
b
(t) P
w
(t) P
s
(t) P
g
(t) _(1 R)P
d
(t) (24)
P
b
(t) P
w
(t) P
s
(t) P
g
(t) P
dump
(t) _P
d
(t) (25)
The WTG output power is calculated as follows:
P
WTG
=
0 iff V\V
ci
a V
3
b P iff V
ci
_V\V
P
r
iff V
r
_V _V
co
0 iff V [V
co
_

_
(26)
where
a = P
r
=(V
3
r
V
3
ci
) (27)
a = V
3
ci
=(V
3
r
V
3
ci
) (28)
P
w
= P
WTG
A
w
g
w
(29)
The PV output power is calculated as follows:
P
s
= H A
s
g
s
(30)
(B) Design constraints:
A
wmin
_A
w
_A
wmax
(31)
A
smin
_A
s
_A
smax
(32)
P
bmin
_P
bsoc
_P
bcap
(33)
0 _P
bcap
_P
bcapmax
(34)
P
b
_P
bmax
(35)
P
gmin
_

T
t=1
P
g;t
_P
gmax
(36)
0 _j _1 (37)
The input parameters considered in this work is as in
Table 1.
The data used (obtained from [7]) for the hourly input of
the insolation, wind speed patterns, and the hourly load
demand in this simulation program are as in Figs. 5, 6, and
7, respectively.
4 Experimental results
The algorithms used in this work were programmed using
the C?? programming language on a personal computer
with an Intel dual core processor running at 2 GHz. The
objective functions, cost ($ year), EIR, and the emissions
(ton/year) were optimized using the HNN, PSO, and the
hybrid HNN-PSO approaches. The results were then
compared against each other as well with the results
Neural Comput & Applic
1 3
Author's personal copy
obtained in [7]. In the analysis of the results, design 1 and
design 2 were the computational results obtained from [7]
while HNN, PSO, and the HNN-PSO are the results of the
algorithms used in this work.
The comparisons of the values of the objective functions
are as in Table 2. The comparisons of the design parame-
ters for each of the methods are provided in Table 3.
Figures 8, 9, and 10 provide the performance of the
methods employed with respect to their effect to the
objective functions [cost ($/year), EIR, and emissions (ton/
year)].
In Fig. 8, it can be observed that design 2 produces the
highest cost while the most minimum cost was achieved by
the HNN method. The second highest cost was obtained by
the design proposed by the PSO method, followed by the
HNN-PSO, and design 1 method. The design by the HNN
method in this case is the most effective cost minimizer.
In Fig. 9, it is observed that the design produced by the
HNN method achieves the highest EIR while the lowest
EIR was produced by design 1. The second highest EIR
was produced by the design proposed by the PSO method,
followed by the HNN-PSO, and the design 2 methods,
respectively. Therefore, it can be observed that the design
obtained using the HNN method produces maximum
reliability.
The lowest pollutant emission rate (PE) was achieved
using the design produced by the PSO method while the
highest is given by design 2 which can be seen in Fig. 10.
The second lowest PE was produced by the design pro-
posed by the HNN-PSO method followed by the HNN and
the design 1 method, respectively. Therefore, it can be
observed that the PSO method was the greenest design
option since it minimizes the PE effectively. In the scatter
plot in Fig. 11, the simultaneous optimization of all three
objectives with respect to the methods utilized in this work
could be observed.
It was also inferred that although the design option
achieved by the HNN method was cost effective and highly
reliable; it produces more emissions (PE) with respect to
the designs produced by the PSO and the HNN-PSO
method. As for the design proposed by the PSO method, it
can be seen that although it provides the lowest emissions
among all the other designs, it was not as cost effective as
the design proposed by design 1, HNN, and the HNN-PSO
method. Besides, the PSO method was not as reliable
compared with the design proposed by the HNN method
but more reliable than the design proposed by all the other
methods. The HNN-PSO method seems to be the middle
point in terms of performance between the HNN and the
Table 1 Input parameters for the hybrid DG system
System parameters Values
g
s
, g
w
, g
b
50 %, 16 %, 82 %
b, T, m 9 %, 12 %, 12 %
N
p
, N
w
, N
s
, N
b
(year) 20, 20, 22, 10
a
w
, a
s
, a
b
($/m
2
) 100, 450, 100
S
w
, S
s
($/m
2
) 10, 45
a
OMw
, a
OMs
, a
OMb
($/m
2
/year) 2.5, 4.3, 10
V
ci
, V
r
, V
co
(m/s) 2.5, 12.5, 20
P
r
(kW) 4
A
wmax
, A
wmin
(m
2
) 10,000, 400
A
smin
, A
smax
(m
2
) 30, 8,000
P
bmax
(kW) 3
P
bmin
(kW) 3
P
br
(kW) 8
P
bcapmax
(kW) 40
psi ($/kW h) 0.12
Fig. 5 Hourly insolation prole
Fig. 6 Hourly wind speed prole
Fig. 7 Hourly power demand prole
Neural Comput & Applic
1 3
Author's personal copy
PSO method. The design proposed by the HNN-PSO
method was more cost effective as compared to the design
by the PSO and design 2 method; however, it was more
costly than the design by the HNN and the design 1
methods. In terms of emissions, the design proposed by the
HNN-PSO method has lower PE as compared to all the
methods except the PSO method. Thus, it can be said that
the HNN-PSO method was the second greener alternative
design after the design by the PSO method. The design by
the HNN-PSO method was more reliable than the design 1
and design 2 but less reliable than the design proposed by
the HNN and the PSO method.
In Fig. 11, the optimum value labeled as OPTIMA at
(0, 1, 0) which is the point with the most minimum cost,
maximum reliability (EIR), and minimum emissions (PE).
The method that provides the closest design option to the
optimum point is the HNN method while the furthest from
the optimum point is the PSO and the HNN-PSO method.
Thus, in this work, it can be said the HNN method pro-
duces the best design option in terms of the overall triple-
objective optimization. The HNN method would then rank
rst in terms of providing the most optimal solution, fol-
lowed by the HNN-PSO, and the PSO algorithm,
respectively.
Table 2 Comparison of value of objective functions
Design 1 Design 2 HNN PSO HNN-PSO
Cost ($/year) 5,323 6,802 4,928.88 5,340.95 5,326.2
EIR 0.9394 0.9505 0.999988 0.998138 0.975175
Emissions (ton/year) 12.4609 65.3812 8.09501 4.91565 6.15164
Computational time (ms) 122 67 127
Table 3 Comparison of the optimized design parameters for each method
Design 1 Design 2 HNN PSO HNN-PSO
A
w
(m
2
) 420 640 412.928 432.406 434.777
A
s
(m
2
) 50 40 25.2482 43.2406 43.4777
P
bcap
(kW h) 16 16 19.9243 16.2571 15.3477
j 0.2 0.58 0.16466 0.12225 0.136145
Fig. 8 Cost ($/year) against the methods
Fig. 9 EIR against the methods
Fig. 10 Emissions (ton/year) against the methods
Fig. 11 Cost versus EIR versus PE
Neural Comput & Applic
1 3
Author's personal copy
Figure 12 shows the network energy differential versus
the number of network recursions (m) in the HNN segment
of the algorithms used. The position of the particles versus
the number of iterations (t) is shown in Fig. 13.
Figure 13 shows the progression of the network energy
difference with respect to the number of network recur-
sions (m). The maximum and minimum value of the net-
work energy difference is 4544.65 and 2.03786,
respectively. When m is about 12, the network energy
difference stabilizes and converges to its minimum value.
Figure 13 shows the position of the 5 particles with respect
to the number of iterations (n). It can be observed that
oscillations occur in the initial iterations. As the iterations
progress, the positions of all the particles stabilize and
converge to a xed value at about 241 iterations onwards.
The list of parameters initialized in the HNN and the
PSO algorithm prior to execution is as in Tables 4 and 5,
respectively.
5 Discussion on ndings and results
The HNN, PSO, and HNN-PSO techniques were observed
to perform stable computations without divergent issues.
From the results, it can be seen that the HNN approach
performs a better job in nding the optimal solution as
compared to the pure PSO and the HNN-PSO approaches.
In terms of ease of implementation (which is directly
linked with the algorithmic complexity), the PSO method
would rank rst, followed by the HNN, and the HNN-PSO,
respectively. Tables 6 and 7 show the advantages and the
disadvantages of the PSO and the HNN algorithm,
respectively. Table 8 shows the advantages and the dis-
advantages of the HNN-PSO algorithm.
From Tables 6, 7, and 8, overall, it can be stated that
PSO method nevertheless outperforms all of the other
techniques in terms of reliability (EIR) maximization,
emissions (PE) minimization, as well as efciency in terms
of computational time. This may be attributed to the
capabilities of the PSO approach in handling nonlinearities
since the emission approximate used in this problem is
nonlinear in nature. The HNN method outperforms all of
the other techniques in terms of cost minimization and
reliability (EIR) maximization. In terms of reliability (EIR)
maximization, the HNN-PSO method performs better than
design 1 and design 2.
An increase in the complexity of the algorithm to
include more optimization as well as constraint handling
features may improve all three approaches. Hence, this
strategy may produce a solution closer to the global opti-
mal. Since the execution of these sorts of programs
(algorithmically complex) would be computationally
costly, thus a more powerful computer would be required.
Another method to improve the three methods utilized in
this work would be to further hybridize it with other
methods like simulator annealing, genetic programming, or
Fig. 12 The network energy differential versus the number of
network recursions (m)
Fig. 13 The position of the particles versus the number iterations (t)
Table 4 List of parameters for the HNN algorithm
Parameters Values
Initial weight for the network 1.5
Initial network inputs
(x
1
, x
2
, x
3
, x
4
, x
5
)
(0.111, 0.026, 0.004, 0.0009,
0.0003)
Number of particles 5
Rate of learning coefcient 0.0001
k, scalar constant in the energy
function
0.5
Table 5 List of Parameters for the PSO algorithm
Parameters Values
Initial parameter (c
1
, c
2
, r
1
, r
2
, w) (0.05, 0.05, 0.35, 0.45, 0.4)
Number of particles 5
Initial social inuence (s
1
, s
2
, s
3
, s
4
,
s
5
)
(1.1, 1.05, 1.033, 1.025,
1.02)
Initial personal inuence
(p
1
, p
2
, p
3
, p
4
, p
5
)
(3, 4, 5, 6, 7)
Neural Comput & Applic
1 3
Author's personal copy
other classes of evolutionary algorithms. Hybridizing these
methods with algorithms like different evolution (DE)
search may provide it with a more efcient system for
handling constraints and thus pave the way for a solution
closer to the global optimum.
The HNN method can be considered robust and appli-
cable for a wide range of optimization problems. With
modications to the energy function of the HNNand some of
the numerical parameters (see Table 4), the HNN method
could be customized for solving a broader class of nonlinear
Table 6 Advantages and the disadvantages of the PSO algorithm
Advantages Disadvantages
1. The PSO method takes lesser computational time to produce the
solution as compared to the HNN and the HNN-PSO approach
1. The PSO method unfortunately is a much poorer cost minimizer as
compared to the other techniques except the design 2 method
mentioned in [7]
2. The pure PSO method is more easily implemented as compared to
the HNN and the HNN-PSO approach since it is has low algorithmic
complexity
2. Due its poor qualities as a cost minimizer, the PSO method does not
perform as well as the HNN-PSO or the HNN methods in terms of
overall optimization
3. The PSO method nevertheless outperforms all of the other
techniques in terms of reliability (EIR) maximization, emissions (PE)
minimization as well as efciency in terms of computational time.
This may be attributed to the capabilities of the PSO approach in
handling nonlinearities since the emission approximate used in this
problem is nonlinear in nature
Table 7 Advantages and the disadvantages of the HNN algorithm
Advantages Disadvantages
1. The HNN method outperforms all of the other techniques
mentioned in this work in terms of overall optimization (see
Fig. 9)
1. The HNN method takes longer computational time to produce the
solution as compared to the pure PSO approach. The HNN technique has
a high algorithmic complexity compared to the PSO method and thus the
computation takes more time
2. The HNN method outperforms all of the other techniques in
terms of cost minimization and reliability (EIR) maximization
2. The HNN method is less easily implemented as compared to the PSO
approach since it is higher in terms of algorithmic complexity
3. In terms of emissions (PE) minimization, the HNN performs
better than design 1 and design 2 [7]
3. The HNN was also observed to be very poor in terms minimization of
emissions as compared to the PSO and the HNN-PSO method. As
mentioned previously, the emission approximate is a nonlinear function,
thus this may be a factor that compromises its performance unlike the
PSO and the HNN-PSO method
Table 8 Advantages and the disadvantages of the HNN-PSO algorithm
Advantages Disadvantages
1. The HNN-PSO method outperforms the PSO and design 2
[7] methods in terms of cost minimization
1. The HNN-PSO method takes longer computational time than all three
methods to produce the solution. This can be attributed to algorithmic
complexity. Since the HNN-PSO technique is a hybrid of two methods, hence
it has the highest algorithmic complexity compared to the PSO and the HNN
method. This makes the computational time high
2. In terms of reliability (EIR) maximization, the HNN-PSO
method performs better than design 1 and design 2
2. The HNN-PSO method is less easily implemented as compared to the HNN
and PSO approach since it is higher with respect to algorithmic complexity
3. As for emissions (PE), the HNN-PSO technique is second
only to the PSO method in terms of minimization
3. The HNN-PSO was also observed to have mediocre performance in terms
minimization of cost, since it outperforms the PSO and design 2 [7] but
underperforms the HNN and design 1 methods
4. The HNN-PSO was also observed to have mediocre performance for the
maximization of reliability (EIR), since it outperforms the PSO and design 2
[7] but underperforms the HNN and design 1 methods
Neural Comput & Applic
1 3
Author's personal copy
problems. This robustness can mainly be attributed to the
convergence capabilities and the recursive nature of the
HNN network to handle multivariate and nonlinearities in a
particular problem. The PSO method on the other hand can
be said to be very suitable for handling nonlinearities.
Therefore, in this work, since the level of nonlinearity is
considerably low, the PSO and the HNN-PSO method are
less robust as compared to the HNN method.
The level of nonlinearity is considerably low where the
only nonlinear scenario in the problem is the quadratic
approximation of the emissions (PE) based on the load. The
recursive nature of the HNN gives it the capacity to handle
this issue where the degrees of nonlinearities are tolerable.
However, it is important to take note in the sense of min-
imization of emissions, the PSO method seems supreme to
all other methods including the HNN method due to its
capabilities in handling nonlinearities.
The HNN, HNN-PSO, and the PSO methods make about
300, 9490, and 85 function evaluations to reach the optimal
solution. Thus, the PSO method is the most efcient in
program execution as compared to the HNN and the HNN-
PSO methods. The reason for this can be attributed to the
difference in terms of algorithmic complexity where the
HNN-PSO method has the highest degree of complexity
followed by the HNN and PSO methods. Due to the network
energy component in the HNN segment in methods (HNN
and HNN-PSO), the stability and the convergence of the
computations are assured in both these methods. As for the
PSOmethod, the oscillations dampen out with the number of
iterations due to the balancing of the social and the personal
interaction mechanisms for each particle. Thus, the stability
of the computations and the convergence of the solutions are
also assured in the PSO method (see Figs. 12 and 13).
A global solution with a better computational time may
be obtained by implementing the HNN, PSO, and the
HNN-PSO methods on a high-performance computer. It
was also inferred that the HNN, PSO, and the HNN-PSO
methods were easily implemented using the language of
C??. However, The HNN-PSO method is slightly more
complicated to implement compared with the PSO and the
HNN approach since it has high algorithmic complexity.
All three methods perform very well in terms of feasibility.
This can be concluded since all three methods do not break
any of the given constraints in this problem. In terms of
computational efciency, it can be concluded that the
PSO method outperforms all three methods. The PSO
method is easily executable and taking very low compu-
tational time.
The HNN method takes about 122 ms, HNN-PSO
method takes 127 ms, and the PSO method takes 67 ms of
computational time to reach the optimal solution. The
computational time is closely related to the algorithmic
complexity. Hence, the variation of this factor in the
methods proposed in this work heavily inuences the
computational time.
Articial neural networks have been used extensively in
solving issues in power systems operations and as control
strategies [2224]. Articial neural networks [25, 26] and
HNN [27, 28] have been specically used for solving the
economic dispatch problem. DG with multiple renewable
energy sources is also a power system related problem and
thus has many similarities with the economic dispatch
problem in terms of its multivariate nature and degree on
nonlinearities. Thus, the similarity between the DG prob-
lem and the economic dispatch problem may be the reason
why the HNN algorithm performs very well in this work.
6 Conclusions and recommendation
The overall optimization of the objective function was
carried very well by the HNN method, followed by the
HNN-PSO, and the PSO methods. The HNN-PSO method
also performs fairly well and hence ranks second to the
HNN method. The PSO method seems to singularly min-
imize the emissions (PE) very well although not as well in
the optimization of the other objectives. In this work, a new
local optimum was reached for the objective functions
using the HNN, PSO, and the HNN-PSO method. The
HNN method can be considered to be the most robust
followed by the HNN-PSO and the PSO methods. The
HNN method can be applicable for a wide range of non-
linear-class optimization problems provided that the
degrees of nonlinearities are not too high. In the event such
a scenario comes, then the PSO or its hybrids would do
well in the optimization. The HNN method is the most
efcient optimizer in this problem as compared to the PSO
and the HNN-PSO approaches although in compromises in
terms of computational time with the PSO method. In
terms of constraint satisfaction, all the methods perform
almost on an equal level. The PSO method is more easily
implemented compared with the HNN and the HNN-PSO
methods in terms of ease of implementation. To improve
the constraint satisfaction while optimizing the objective
functions further approaching the global optimal, other
optimization techniques such as genetic programming and
simulator annealing should be implemented and tested in
their pure form or as hybrids.
Acknowledgments This work was supported by STIRF grant
(STIRF CODE NO: 90/10.11) of University Technology Petronas
(UTP), Malaysia. The authors sincerely thank the anonymous referees
for their valuable and constructive comments and suggestions for the
improvement of the overall quality on this paper.
Neural Comput & Applic
1 3
Author's personal copy
References
1. EI-Sayes MA, Osman MG, Kaddah SS (1993) Assessment of the
economic penetration levels of photovoltaic panels, wind turbine
generators and storage batteries. Electric Power Syst Res 27:
233246
2. Kandil MS, Farghal SA, EL-Alfy AE (1991) Optimum operation
of an autonomous energy system suitable for new communities in
developing countries. Electric Power Syst Res 21:137146
3. Kellogg W, Nehrir MH, Venkataramanan G, Gerez V (1996)
Optimal unit sizing for a hybrid wind/photovoltaic generating
system. Electric Power Syst Res 39:3538
4. Kellogg W, Nehrir MH, Venkataramanan G, Gerez V (1998)
Generation unit sizing and cost analysis for stand-alone wind,
photovoltaic, and hybrid wind/PV systems. IEEE Trans Energy
Convers 13(1):7075
5. Lund H (2005) Large-scale integration of wind power into dif-
ferent energy systems. Energy 30:24022412
6. Notton G, Cristofari C, Poggi P, Muselli M (2001) Wind hybrid
electrical supply system: behaviour simulation and sizing opti-
mization. Wind Energy 4:4359
7. Wang L, Singh C (2007) PSO-based multi-criteria optimum
design of a grid-connected hybrid power system with multiple
renewable sources of energy. In: Proceedings of the IEEE Swarm
intelligence symposium, pp. 250157
8. Chedid R, Akiki H, Rahman S (1998) A decision support tech-
nique for the design of hybrid solar-wind power systems. IEEE
Trans Energy Convers 13(1):7683
9. Chedid R, Rahman S (1997) Unit sizing and control of hybrid
wind-solar power systems. IEEE Trans Energy Convers 12(1):
7985
10. Hopeld JJ (1982) Neural networks and physical systems with
emergent collective computational abilities. Proc Natl Acad Sci
USA 29:25542558
11. Hopeld JJ (1984) Neurons with graded response have collective
computational properties like those of two-state neurons. Proc
Natl Acad Sci USA 81:30883092
12. Lee KY, Sode-Yome A, Park JH (1998) Adaptive hopeld neural
networks for economic load dispatch. IEEE Trans Power Syst
USA 13(2):519526
13. Tank DW, Hopeld JJ (1986) Simple Neural optimization
network: an A\D converter, signal decision circuit and a linear
programming circuit. IEEE transactions on circuits and systems,
CAS-33, pp 533541
14. Kennedy J, Eberhart R (1995) Particle swarm optimization. In:
IEEE proceedings of the international conference on neural net-
works, Perth, Australia, pp 19421948
15. Phuangpornpitak N, Prommee W, Tia S, Phuangpornpitak W
(2010) A study of particle swarm technique for renewable energy
power systems. PEA-AIT international conference on energy and
sustainable development: issues and strategies, Thailand, pp 17
16. Shi Y, Eberhart R (1998) A modied particle swarm optimizer.
In: Proceedings of the IEEE international conference on evolu-
tionary computation, pp 6973
17. Frans van den Bergh (2001) An analysis of particle swarm
optimizers. PhD thesis, University of Pretoria
18. Zitzler E, Laumanns M, Bleuler S (2002) A tutorial on evolu-
tionary multiobjective optimization
19. Barton JP, Ineld DG (2004) Energy storage and its use with
intermittent renewable energy. IEEE Trans Energy Convers 19
(2):441448
20. Talaq JH, El-Hawary F, El-Hawary ME (1994) A summary of
environmental/economic dispatch Algorithms. IEEE Trans Power
Syst 9:15081516
21. Bayon L, Grau JM, Ruiz MM, Suarez PM (2006) Optimization of
SO2 and NOx emissions in thermal plants. J Math Chem 40(1):
2941
22. Moreno J, Esquivel A (2000) Neural network based approach for
the computation of harmonic power in real time microprocessor-
based vector for an induction motor drive. In: IEEE transactions
on industry application, pp 277282
23. Vazquez JR, Salmeron PR (2000) Three phase active power lter
control using neural networks. In: 10th Mediterranean electro-
technical conference, vol 3, pp 924927
24. Bahbah AG, Girgis AA (2004) New method for generators
angles and angular velocities prediction for transient stability
assessment of multi machine power systems using recurrent
neural network. IEEE Trans Power Syst 19:10151022
25. Singh G, Srivastava SC, Kalra PK, Vinod Kumar DM (1995) Fast
approach to articial neural network training and its application
to economic load dispatch. Electric Mach Power Syst 23:1324
26. Kumar J, Sheble GB (1995) Clamped state solution of articial
neural network for real-time economic dispatch. IEEE Trans
Power Syst 10(2):925931
27. King TD, El-Hawary ME, El-Hawary F (1995) Optimal envi-
ronmental dispatching of electric power systems via an improved
hopeld neural network model. IEEE Trans Power Syst 10(3):
15591565
28. Abe S, Kawakami J, Hirasawa K (1992) Solving inequality
constrained combinatorial optimization problems by the hopeld
neural networks. Neural Netw 5:663670
Neural Comput & Applic
1 3
Author's personal copy

You might also like