Energy and Buildings 58 (2013) 227–236

Contents lists available at SciVerse ScienceDirect

Energy and Buildings

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

A novel ant colony optimization-based maximum power point tracking for

photovoltaic systems under partially shaded conditions
Lian Lian Jiang a,∗ , Douglas L. Maskell a , Jagdish C. Patra b
a
School of Computer Engineering, Nanyang Technological University, Singapore
b
Faculty of Engineering & Industrial Sciences, Swinburne University of Technology, Australia

a r t i c l e i n f o a b s t r a c t

Article history: In order to achieve maximum efficiency a photovoltaic (PV) arrays should operate at their maximum
Received 18 September 2012 power point (MPP). Therefore, an MPP tracking (MPPT) scheme is implemented between the PV system
Received in revised form and the load to obtain maximum power. When the irradiance distribution on the PV arrays is uniform,
26 November 2012
many traditional MPPT techniques can track the MPP effectively. However, when the PV arrays are par-
Accepted 2 December 2012
tially shaded, multiple MPPs show up, which usually results in the failure of finding the global MPP. Some
researchers have reported this problem and tried to solve it, but most of the MPP control schemes are
Keywords:
relatively complicated or fail to guarantee the MPP under all shading circumstances. In order to over-
Maximum power point tracking
Photovoltaic systems
come this difficulty, this paper presents a novel ant colony optimization (ACO)-based MPPT scheme for
Ant colony optimization PV systems. A new control scheme is also introduced based on the proposed MPPT method. This heuristic
Partially shaded conditions algorithm based technique not only ensures the ability to find the global MPP, but also gives a simpler
control scheme and lower system cost. The feasibility of this proposed method is verified with the irra-
diance of various shading patterns by simulation. In addition, the performance comparison with other
traditional MPPT techniques, such as: constant voltage tracking (CVT), perturb and observe (P&O), parti-
cle swarm optimization (PSO), is also presented. The results show that the proposed algorithm can track
the global MPP effectively, and is robust to various shading patterns.
© 2012 Elsevier B.V. All rights reserved.

1. Introduction arrays mainly depends on the characteristic of the load. In order to

maximize the power from the PV system, an MPP tracker is usually
The use of renewable energy has experienced consistent growth inserted between the PV arrays and the load to make sure that the
due to the limited supply of fossil fuel based energy and the con- system operating point is adjusted to be located at the MPP.
stantly growing environmental concerns associated with burning Over the years, many researchers have studied MPPT algo-
fossil fuels. Solar energy systems as a renewable energy source is of rithms. Buciarelli et al. [1] proposed the perturb and observe (P&O)
particular interest due to their lower maintenance, the abundance (or hill climbing) based MPPT algorithm which is widely used in
of the energy source, almost zero post-production pollution and commercial products. Hussein et al. [2] proposed the incremen-
the advancements in semiconductor and power electronic devices. tal conductance (IncCond) method, which is more efficient under
Among the different ways of converting sunlight into electricity, rapidly changing conditions as it uses the fact that the derivative
the PV systems directly convert solar radiation into electricity by of the power with respect to the voltage (dP/dV) at the MPP is zero.
the photovoltaic effect. Assemblies of solar cells make solar mod- However, when the PV systems are operated under partially shaded
ules and the large-scale PV systems usually consist of many solar conditions, which are usually caused by the passing clouds, nearby
modules connected in series or parallel. Due to the nonlinearity trees or buildings, long-lasting dust, etc., the characteristic of P–V
between the PV output voltage and current, there is a unique maxi- curve shows multiple peaks. This results in these two traditional
mum power point (MPP) in the power–voltage (P–V) characteristics MPPT algorithms becoming trapped at a local maximum, causing a
under uniform weather conditions. When the PV arrays are directly significant energy loss of up to 70% [3].
connected with the load, potential power that could be extracted To tackle this problem, Kobayashi et al. [4] proposed a two-stage
from the PV arrays is wasted because the power output of the PV MPPT controller to track the global maximum by moving the oper-
ating point to the vicinity of the real MPP at the first stage and,
thereafter, shifting this point to the real MPP using the traditional
∗ Corresponding author. IncCond algorithm. The advantage of this method is that the algo-
E-mail addresses: [email protected] (L.L. Jiang), [email protected] rithm is capable of tracking the global MPP in most cases. However,
(D.L. Maskell), [email protected] (J.C. Patra). the additional circuits for measuring the short circuit current and

0378-7788/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.enbuild.2012.12.001
228 L.L. Jiang et al. / Energy and Buildings 58 (2013) 227–236

open circuit voltage online add complexity to the system, and it

requires a large amount of storage. Patel and Agarwal [5] proposed Id1 Id2 Ip I +
Rs
an observation method to track the global MPP under partially Ipv
shaded conditions by searching two sides of the local MPP with
disturbance step Vlarge , which is less than the minimum possible Rp V
displacement between the two successive peaks. This method can
find the global MPP, but the MPP tracking speed is limited because
all the local MPPs are required to be checked in order to get the
-
global MPP. In addition, when the irradiance is non-uniform, with
a complex distribution, the multiple local maximum points (which
Fig. 1. Equivalent circuit for two-diode model.
tend to be close to each other) make the selection of the step size
for voltage perturbation quite difficult. Another method associated
with the MPPT under partially shaded conditions was proposed in cost. It also guarantees to find the global MPP under various par-
[6]. The basic idea of this algorithm is to move the operation point tially shaded conditions. In addition, it features fast convergence
decided by a linear function (Vpv = (Vo /Io ) × Ipv , where Vpv is the ref- speed, convergence independent of the initial conditions, and no
erence voltage at the next perturbation step, Vo is the open circuit requirement of knowledge about the characteristic of PV array.
voltage and Io is the short circuit current), and then, use the tradi- The reminder of the paper is organized as follows. Section 2
tional MPPT algorithms, such as P&O or IncCond, to track the global introduces the PV system and our proposed configuration structure
MPP. The advantage of this method is that it works very well under that we investigated. Section 3 briefly presents the main principle
both uniform irradiance and non-uniform irradiance with slow of the ACO algorithm and how it is applied to MPPT for PV sys-
irradiance changes. However, under non-uniform irradiance with tems. Section 4 provides experimental results and a discussion of
rapidly changing conditions, the traditional MPPT algorithm works the proposed approach. Comparisons with other methods are also
better than the proposed algorithm. Miyatake et al. [7] propose presented in this section. Finally, Section 5 presents the conclusions
another MPPT method based on the Fibonacci search algorithm. The and directions for future work.
main principle is to iteratively restrict and shift the search range to
contain the optimal point in the range by using the line search algo-
2. Photovoltaic system
rithm with an improved Fibonacci sequence. This algorithm can be
used under partially shaded or rapidly changing irradiance con-
2.1. PV array models
ditions, but does not guarantee to find the global MPP under all
conditions. A dividing rectangle (DIRECT) search method in con-
In order to test our proposed MPPT method, a two-diode model
junction with the P&O method has also been proposed in [8]. The
[20] is used to model the characteristics of the PV arrays. It reduces
algorithm searches the Lipschitz function, which describes the PV
the computational time and improves the accuracy of the perfor-
power and voltage relationship in an interval. However, the selec-
mance especially under low irradiance. The equivalent circuit of
tion of the initial operating point heavily influences performance in
the two-diode model for a solar cell is shown in Fig. 1.
finding the global MPP. In addition, since the P&O algorithm is used
Suppose that the PV array contains Nss PV modules in series and
under uniform conditions, the oscillation around the MPP reduces
Npp PV strings in parallel, the output current of the PV array is given
the efficiency of the system.
by the equation [20]
Some other methods are also proposed [9–11]. For instance,
the fractional-order incremental conductance method shortens the   V + IR  
s
tracking time [9]. However, it is only suitable for small-scale PV I = Ipv Npp − I01 Npp exp −1
a1 Vt Nss
systems. When applying this method to large-scale PV systems   V + IR   V + IRs
under partially shaded conditions, it fails to find the global MPP. The s
− I02 Npp exp −1 − (1)
power or voltage compensation method improves the efficiency of a2 Vt Nss Rp
the system [10,11], but requires additional devices which increase
where I is the photovoltaic output current, V is the photovoltaic
the complexity and cost of the system.
output voltage,  = Nss /Npp , Rs and Rp are the series and parallel
Compared to these direct search methods, computational intel-
resistances, respectively, Vt is the thermal voltage of the two diodes
ligence based methods, such as fuzzy logic (FL) [12], artificial
(Vt = Ns k T/q), Ns is the number of solar cells connected in series in
neural network (ANN) [13–15], particle swarm optimization (PSO)
each PV module, k is the Boltzman constant, q is the electron charge,
[16] and so forth, offer significant benefits. These can include:
and a1 , a2 are the diode ideal constants. The light generated current
no requirement for knowledge of internal system parameters,
is given by
reduced computational effort and a compact solution for multi-
variable problems. However, for fuzzy logic based methods, the G
Ipv = (Ipv STC + Ki (T − TSTC )) (2)
fuzzy rule base, which is dependent on the experience of algorithm GSTC
developers, significantly influences the performance of MPPT. For
ANN based methods, it is only suitable for the system that can where Ipv STC represents the light generated current under standard
get sufficient training data. The PSO based method is efficient for test conditions (STC) with temperature TSTC = 25 ◦ C, and irradiance
non-uniform weather conditions. However, its convergence signif- GSTC = 1000 W/m2 , and the constant Ki is the short circuit current
icantly depends on the initial place of the agents [16]. coefficient. The reverse saturation current of the diode is given by
At present, ant colony optimization (ACO) has been widely used Isc STC + Ki T
in scheduling [17], image processing [18], power electronic circuit I01 = I02 = (3)
exp((Voc STC + Kv T )/Vt ) − 1
design [19], and many other fields. In this paper, we propose a novel
ACO-based MPPT to track the MPP for a large-scale PV system under where the constant Kv is the open circuit voltage coefficient, Ipv STC
partially shaded conditions. We also design a PV array configura- is the short circuit current under STC, and Voc STC is the open circuit
tion structure along with this new MPPT algorithm. With this MPPT voltage under the STC. In fact, Eq. (1) can be represented as I = f
control scheme, only one pair of current and voltage sensors are (I, V). Therefore, for a given current value, this nonlinear equation
required which simplifies the PV system and reduces the system can be solved using the standard Newton–Raphson method. The
 L.L. Jiang et al. / Energy and Buildings 58 (2013) 227–236 229

Table 1
The specifications of the PV module.
No bypass diode
6 With bypass diode
Maximum power (Pmax ) 60 W Short circuit current (Isc ) 3.8 A Uniform
Optimum voltage (Vmp ) 17.1 V Number of cells (Ns in series) 36 M1
4

I(A)
Optimum current (Imp ) 3.5 A Temp. coefficient of Isc A (◦ C) 0.003 M2
Open circuit voltage (Voc ) 21.1 V Temp. coefficient of Voc V (◦ C) −0.08
2

relationship between the current and voltage of the bypass diode

is described as 0
0 10 20 30 40 50
 Ib
 V(V)
Vb = −a3 Vt ln 1 + (4)
Iob Fig. 3. The I–V curves of PV modules (M1 and M2 ) and the total PV string, with
where Ib , Iob , Vb , and a3 is the current passing through the bypass and without bypass diodes, for a small PV system with two PV modules con-
nected in series. For the partial shading condition, the irradiance on each PV module
diode, the saturation current, the voltage across the bypass diode, 2 2
is: GM1 = 300 W/m and GM2 = 1000 W/m . For the uniform irradiance condition:
and the ideal constant of the bypass diode, respectively. The PV GM1 = GM2 = 1000 W/m .
2

modules used in this work are BP MSX 60. The specification of the
PV module is listed in Table 1.
PV module receives a different irradiance value (GM1 = 300 W/m2 ,
2.2. The partially shaded problem definition GM2 = 1000 W/m2 ). Fig. 4 gives the P–V curves, and shows that
there is only one MPP in the P–V curve for the module string without
As indicated earlier, PV cells are extremely sensitive to partial bypass diodes. Thus without bypass diodes, the number of MPPs is
shading. If a cell or a small portion of a module in a series string the same as when the irradiance on these two modules is uniform
is shaded, then instead of contributing to the power output, the (at 300 W/m2 ). However, when the bypass diodes are added, each
shaded cell(s) will absorb the power from the other cells in the module contributes to the total power, and as such two MPPs are
string. This absorbed power is converted into heat, contributing to observed in the P–V curve of the module string. Here, the number of
hot spots that can damage the cell(s). Most commercial modules use MPPs is related to the uniformity of the irradiance on each PV mod-
bypass diodes across a series of cells to overcome this effect. Thus, ule and the number of the modules (bypass diodes). For example,
the shaded cells are bypassed and as a result, only the power from for a PV system with Nss modules (as in Fig. 2) connected in series
the shaded series (of cells) is lost. A typical PV module consists of with each module receiving a different irradiance, the number of
a series connection of cells with a bypass diode across each 18–20 MPPs in the P–V curve will equal Nss . Some researchers [21,22] have
cells. This group of cells is called a series with a bypass diode. A suggested using a bypass diode for each solar cell to reduce the
typical module will have 2 or 3 series. power loss caused by partial shading within a PV module, but this
When only a single cell (or a small number of cells) in a series is would add to the cost and complexity of the module. These differ-
partially shaded the series behaves the same as if the entire series ent bypass diode configurations and how they could be optimized
is partially shaded. To simplify the analysis, and to avoid confusion to maximize power is not within the scope of this paper.
in our partial shading analysis, we have considered a module to
have just a single series with bypass diode. Then a more realistic 2.3. MPPT under partially shaded conditions
commercial module could be considered as a series connection of
two or more of these simple modules, as shown in Fig. 2. Using this As indicated above, bypass diodes are added to commercial PV
simple module model means that there is no difference in power modules to protect the PV panel from hotspots due to the mod-
output, irrespective of whether a single cell or the complete module ule acting as a load and consuming power. In order to protect the
is shaded, so long as the irradiance is uniform. PV panel, the bypass diodes (D1 and D2 ) are added to the PV sys-
In order to illustrate the characteristics of the PV modules under tem by connecting them parallel with the PV modules, as shown in
half-shading and full shading conditions, the I–V and P–V curves Fig. 2. In addition, the blocking diode (Db ) is usually connected in
of the PV module with and without bypass, for the small PV sys- the PV string in series to prevent the reverse current from the load
tem shown in Fig. 2, are given in Figs. 3 and 4 respectively. Fig. 3 or unbalanced current flow from other paralleled PV strings.
shows the I–V curves for each PV module and for the combined Since the current through these two modules with different irra-
PV string without bypass diodes and with bypass diodes. Each diance is equal due to their series connection for the PV string with
bypass diodes, the I–V curve of this assembly (in green dotted line)
can be obtained by adding the voltages across them while keeping
the current value constant as shown in Fig. 3. From the I–V curve of

150 No bypass diode

With bypass diode G3
Uniform P1
100 M1
P(W)

M2

50 G1 G2
P2

0
0 10 20 30 40 50
V(V)

Fig. 4. The P–V curves of the modules with the same series assembly and the same
Fig. 2. A small PV system with two PV modules, module M1 is shaded and M2 is irradiance as in Fig. 3. P1 and P2 represent the PV string power for the uniform and
receiving normal irradiance (1000 W/m2 ). non-uniform irradiance conditions, respectively.
230 L.L. Jiang et al. / Energy and Buildings 58 (2013) 227–236

these two modules, it is evident that two knee-points are formed, the shortest path through the exchange of such information. Ini-
generating two local maxima (G1 and G2 ) in the resultant P–V curve tially, the ACO algorithm is used to tackle combinational problems.
as shown in Fig. 4. When these two modules are exposed to uni- Nowadays, more and more researchers have extended this tech-
form irradiance, there is only one MPP (G3 ) in the P–V curve. In nology into the continuous domain [25–27], among which Socha
order to extract the maximum power from the PV array, the sys- et al. [25] show that the ACOR gives better performance than other
tem should be operated at the global MPP (G3 or G1 , depending on related ACO algorithms. For continuous problems, the solution con-
the distribution of the irradiance). However, when the irradiance struction is a little different from the combinational problems. An
changes from uniform to non-uniform distribution with the power archive is defined in the process of solution generation as shown in
curve changing from P1 to P2 , the operating point of the PV system Fig. 6. Suppose that there is an N-dimensional problem in which
with a traditional MPPT algorithm, such as P&O and IncCond, will N parameters need to be optimized. The procedure for solution
move to the local maxima (G2 ). Thus, the potential power (PG1 –PG2 ) construction can be described as follows [25].
from the PV arrays is wasted. Therefore, in order to maximize the Initially, K random solutions are generated and stored in the
power output from PV systems, an efficient MPP tracking method solution archive (see Fig. 6) with size (K ≥ N). Then, these K solutions
is needed. are ranked according to the evaluation value (from best to worst, in
the minimization problem, f(s1 ) ≤ f(s2 ) ≤ · · · ≤ f(sl ) ≤ · · · ≤ f(sK )). For
each dimension, the probability density function is a Gaussian func-
2.4. Proposed PV array configuration structure
tion which consists of multiple Gaussian sub-functions. The new
solutions are generated by sampling this Gaussian kernel for each
PV modules can be connected in series, parallel or a hybrid,
dimension, which is given by
among which the hybrid connections are popular. Generally, the
 
PV modules are not connected totally in parallel or series because 
K

K
1 (x − il )
2

it may not meet the load current and voltage requirements. In addi- G i (x) = ωl gli (x) = ωl √ exp − (5)
i
2li
2
tion, it reduces the system stability. For instance, when one of the l 2
l=1 l=1
PV modules connected in series is shut down, all the PV modules
where Gi (x) is the Gaussian kernel for the ith dimension of the
will be influenced by this single troubled module. In this work, in
solution, gli (x) is the lth sub-Gaussian function for the ith dimen-
order to illustrate the system performance clearly, we consider a PV
system with two PV stings and three PV modules in each PV string. sion of the solution, il and li are the ith dimensional mean value
Usually, there are two modes of MPPT control: the centralized and and the standard deviation for the lth solution, respectively. With
distributed control methods. Centralized control means that the K solutions in the archive, three parameters of the Gaussian ker-
entire PV array is controlled by a single MPPT controller. Each PV nels (mean, i ; standard deviations,  i ; weight, ωl ), are calculated
string has an equal value for the control variables (voltage or cur- based on following equations.
rent). Distributed control requires that each PV panel is equipped Mean:
with an MPPT which are controlled individually. The advantage of i = {i1 , . . . , il , . . . iK } = {s1i , . . . , sli , . . . sKi } (6)
the former mode is that it has a simple structure. However, it is
easy to lose energy and cannot maximize the energy harvest. For where sli is the ith dimensional value for the lth solution.
example, when each of the PV strings is assigned the same control Standard deviation:
value, it cannot guarantee that each PV string is working at the best

K
|sji − sli |
operating point. The latter method can maximize the energy out- li =  (7)
put, but has a complex structure and a higher system cost, as each K −1
j=1
PV module requires its own MPPT controller.
In this work, we use centralized control by applying computa- where the standard deviation for the ith dimension of the lth solu-
tional intelligence optimization algorithms, but the control value tion is calculated by multiplying the average distance from the
of each PV string converter is different. The difference to the tradi- chosen solution sl to other solutions in the archive with the parame-
tional centralized control is that, in our proposed PV array structure ter, , which represents the speed of convergence (higher the value
as shown in Fig. 5 each PV string is installed with a DC–DC converter of , the lower the convergence speed) [25].
instead of each PV module. Moreover, only one pair of current and Weight:
voltage sensors are required, which reduces the system cost. Either 2
the current or the voltage of these PV strings can be regarded as the 1 (l − 1)
ωl = √ exp − , (ωK ≤ · · · ≤ ωl ≤ · · · ≤ ω2 ≤ ω1 )
control variable to maximize the power output. Therefore, this con- QK 2 2Q 2 K 2
figuration structure not only improves the reliability of the system, (8)
but also increases the efficiency of the energy output. where ωl is the weight of solution sl , l is the rank of solution sl , Q is
the algorithm parameter which represents the importance of the
3. Ant colony optimization applied to MPPT in a PV system best ranked solutions. When Q is small, the possibility of choosing
the best ranking solution is larger. When it is large, the chance of
3.1. Ant colony optimization choosing all the solutions is equal.
The process of sampling a Gaussian kernel with multiple
The ant colony optimization (ACO) is a probabilistic algorithm, Gaussian sub-functions is divided into two steps: first, choose
which is used to find the global optimal solution for a nonlinear the Gaussian function; second, sample the chosen Gaussian
problem. ACO mimics the foraging behavior of the ants to achieve sub-function randomly according to the parameterized normal dis-
optimization of the path in a graph. The collective behaviors of tribution. The probability of choosing the lth Gaussian function can
a large number of ants form a positive feedback phenomenon be calculated by
[23,24]: ants initially search the path randomly and lay down ωl
pl = (9)
pheromone for other ants to follow. The more ants that travel a r=K
ω
r=1 r
path, the higher density of pheromone on the path, and as a result,
the greater likelihood that subsequent ants will choose the path. The sampling process is repeated for each dimension (i = 1, . . .,
Finally, most of the ants follow the trail until ant individuals find N) of each new solution until M (also called the number of ants)
 L.L. Jiang et al. / Energy and Buildings 58 (2013) 227–236 231

Fig. 5. Proposed ACO based MPPT control structure.

new solutions are created. After this, these M new solutions are where st is the current vector at the tth step, INt is the current control
added to the original solutions. All the M + K solutions are ranked value for Nth PV string at the tth step. The objective function of this
again and only the K best solutions are kept into the archive. Thus, optimization problem f (st ) is the summation of the power output
the whole procedure is repeated until the maximum iteration is from each PV sub-string after applying current variable to each PV
reached or the termination conditions are satisfied. With this solu- string. It can be described by
tion construction scheme, all the ants will be effectively guided to Npp
the optimized point in the search process. 
t
f (s ) = (Ijt × Vjt ) (11)
The ACO combines a positive feedback mechanism, distributed
j=1
computing, and a greedy search algorithm. It has a strong ability
to search the optimal solution. The positive feedback mechanism where Ijt and Vjt are the current control variable value and the volt-
ensures that the ant colony algorithm is capable of early detection age value for jth PV string at the tth step respectively. Assuming
of the optimal solution. Distributed computing helps the ACO to that there are M new solutions generated in each iteration and the
avoid premature convergence. The greedy search helps to quickly
find an acceptable solution, thereby enhancing the efficiency of the
system. s1 s11 s12 ... s1i ... s1n ω1 f(s1)

3.2. Application of ACO to the MPPT problem s2 s2 1 s2 2 ... s2 i ... s2 n

ω2 f(s2)
. . . . . . . .
The ACO-based optimization is now modified to solve the MPPT
problem in PV systems. The system structure that we considered . . . . . . . .
is shown in Fig. 5. The control variable is the current of each PV . . . . . . . .
string. The 3-D plot for the relationship between power and the sl sl1 sl2 ... sli ... sln
PV array current (I1 , I2 ) is shown in Fig. 7. From the figure, we
ωl f(sl)
can see that there are multiple local peaks when the PV arrays are . . . . . . . .
illuminated under non-uniform irradiance. Actually, the MPPT for
. . . . . . . .
PV systems under partially shaded conditions with the proposed
PV array structure can be simplified as a global optimization in a . . . . . . . .
multidimensional space. sK sK1 sK2 ... sKi ... sKn ωK f(sK)
In order to illustrate the control process, we define the control
vector as G1 G2 Gi Gn
st = [I1t , I2t , . . . , INt ] (10) Fig. 6. The archive of solution generation process in ACOR .
232 L.L. Jiang et al. / Energy and Buildings 58 (2013) 227–236

Initialize the solutions in the

180 A archive; Set parameters
135
(MaxIter, K, M, Q, EP);
P(W)

90
45
B Sense V(k), I(k);
0 Compute power P(k) for the ant;
3
2 3
2
I2(A) 1 1
0 0 I1(A)
Flag_ant=Flag_ant+1;Store
Fig. 7. 3-D image of the power versus the current in each PV substring under the C power for each ant; Generate a
shading pattern of [M11 , M12 , M13 ; M21 , M22 , M23 ] = [1000, 800, 400; 1000, 400, new solution; Eqn. (5)-(9)
200] W/m2 .

maximum number of iterations is T, the values of current for each No Reinitialize

Flag_ant≥ M?
sub-string (st ) change in the following series order I solutions in the
D archive;
(s11 → s12 → · · · → s1M ) → · · · → (st1 → st2 → · · · → stM ) Yes
Rank M+K solutions and keep
(st+1
1
→ st+1
2
··· → st+1
M
) → ··· → (sT1 → sT2 → ··· → sTM ) (12) best K solutions in the archive;
E
Flag_Iter=Flag_Iter+1;
The power of each current vector, f(st ), is calculated and evalu- Flag_ant=1;
ated at each stage simultaneously. During the search process, the
solution archive is updated iteratively with the newly generated
solutions. All the processes are repeated until the global MPP is
No Flag_Iter≥ MaxIter?
obtained. Thus, for a particular shading pattern, the ACO-based
or power remains
searching method can find the global MPP. unchanged?
F
Usually, the real working environment of the PV system is
always changing due to the varying weather or the load conditions,
and as a result, the global MPP is always changing. This requires Yes
that the MPPT algorithm has the ability to search for global MPP
for the new weather condition. In order to realize this purpose, Keep operating point at the MPP;
the search process is restarted within a period, and the reinitial- G
ization is executed depending on the detected weather conditions.
Here, we use the following strategy to detect the change of shading
patterns, which is given by No Yes
Sudden change in irradiance
I(si+1 ) − I(si ) or timer interrupt? Eqn. (13)
> I (13) H
I(si )
When (13) is satisfied or the period set by the timer is reached,
Fig. 8. Flow chart of ACO-based MPPT algorithm for PV system.
the search process will be executed again. Thus, it ensures that the
MPPT can always find the global MPP under various working envi-
ronmental conditions. The flow chart of the proposed ACO-based be reinitialized (blocks I and A) so that it can find the global MPP
MPPT algorithm for PV systems is shown in Fig. 8. It can be described for the new environment conditions.
as follows: In the proposed MPPT algorithm, the parameters required to be
First, initialize the solutions in the archive range and set the algo- decided by the user include the number of ants (M), the size of
rithm parameters (block A), which include the maximum number of the solution archive (K), convergence speed constant (), and the
iterations (MaxIter), the size of the archive (K), the number of ants locality of the search process (Q). The number of ants influences the
(M), the balance coefficient (Q), and the convergence speed con- convergence speed and the accuracy of the optimization process.
stant (EP). Then, sense the voltage and current from the entire PV More ants mean it is easier to find the global MPP under differ-
array and calculate the power (block B). After that, mark the corre- ent irradiance conditions, while it requires more time to transfer
sponding ant by using a sign variable (Flag ant) and generate a new all the ants into the MPP. Fewer ants gives a better convergence
solution based on the sampling process by using Eqs. (5)–(9) (block speed, however, they can easily become trapped into one of the
C). By repeating process blocks, B and C, M new solutions are cre- local power peaks. The size K of the solution archive should not be
ated. Thereafter, these M new solutions and the original solutions smaller than the dimension of the problem [25]. The relationship
in the archive are put together and ranked. As a result, keep the K between the size of the archive and the percentage of the extracted
best solutions into the solution archive. The number of the iteration power for all the cases investigated in the experiments is shown in
(Flag Iter) is also marked in each step (block E). The processing of Fig. 9. Therefore, the tradeoff between fast convergence speed and
blocks (B, C, D, E, F) is repeated until the maximum iteration number tracking accuracy should be made when choosing the number of
(MaxIter) is reached or the power value remains unchanged (with ants. Since our objective for the MPPT algorithm is to get fast con-
very small change) within a specified number of successive iter- vergence with a minimum number of ants, we choose an archive
ations. With the above procedures, the global MPP can be found. size of 7. The convergence speed constant () acts like the rate of
Finally, keep this working point (block G) until a timer interrupt pheromone evaporation while solving combination problems. The
occurs or a sudden irradiance change is sensed by (13) (block H). higher the value of , the lower the convergence speed because, in
Whenever these two occurrences happen, the whole algorithm will this case, the search is less biased toward the space that has been
 L.L. Jiang et al. / Energy and Buildings 58 (2013) 227–236 233

Table 2
Percentage ofthe extracted power

1 The shading patterns for the experiment conditions.

0.8 Pattern no. Shading pattern [M11 , M12 , M13 ; M21 , M22 , M23 ] (W/m2 )

SP1 [1000, 1000, 1000; 1000, 1000, 1000]

0.6
SP2 [1000, 800, 400; 1000, 400, 200]
SP3 [500, 200, 400; 800, 400, 200]
0.4
SP4 [600, 1000, 1000; 800, 400, 200]

0.2
Table 3
0 Algorithm parameters used in the experiment.
1 2 3 4 5 6 7 8 9 10
Size ofthe solution archive Parameter Symbol Value

Solution archive size K 7

Fig. 9. Extracted power versus the archive size.
No. of ants in an iteration M 4
Dimension of the problem N 2
Convergence speed constant  0.82
searched [25]. The locality of the search process (Q) determines Locality of the search process Q 0.45
the balance between the diversification and intensification. Diver-
sification in the search avoids being stuck in the local peaks, and
the intensification ensures a fast search process. When Q is small, Table 4
The performance of the proposed ACO-based MPPT under various shading patterns.
the best-ranked solutions in the archive are more likely to be cho-
sen, while when it is large, the probability of choosing solutions Pattern no. ACO-based MPPT
is uniform [25]. In order to illustrate this phenomenon, we tested I1 (A) I2 (A) Pideal (W) P (W)
the relationship between the number of iterations for ACO-based
SP1 3.488 3.488 359.10 359.09
MPPT algorithm to converge and the value of Q the box-plot of the
SP2 2.843 3.459 150.10 150.08
number of iterations to converge for different Q values is illustrated SP3 1.376 1.381 95.20 95.17
in Fig. 10. It shows that the larger value of Q requires more iteration SP4 3.497 1.382 164.47 162.62
to converge. Therefore, when choosing the values of Q, the tradeoff
between these two contradictory factors should also be made.
By applying the ACO-based MPPT algorithm for these four cases,
the current in each PV string (I1 and I2 ), the ideal power (Pideal ) and
4. Experimental results the average power measured by running the proposed method (P)
200 times are shown in Table 4. It shows that the proposed ACO-
In this section, we verify the feasibility of the proposed ACO- based MPPT can track the global MPP so that the power output is
based MPPT using MATLAB simulation. Comparisons with other almost the same as the ideal power under all four different shading
existing MPPT algorithms, under dynamic irradiance conditions, cases. Note that this power output is under the assumption that
are also conducted. there is no power leakage in the PV system due to the conversion
efficiency of the DC–DC converter, power losses from the wire con-
4.1. Experiment 1 nection or other factors. When implementing this MPPT in a real
system, the power harvest will be less than this ideal case. The
This experiment demonstrates the ability of the proposed ACO- box-plot distribution of the power extracted from the PV arrays for
based MPPT to track the global MPP under steady and transient these four shading patterns is given in Fig. 11. From the figure, we
shading patterns. The structure of the PV system for this test is can see that after 200 runs for each shading pattern, the distribution
shown in Fig. 5. Since it is very difficult to test all the non-uniform of the final power extracted from the PV system using our proposed
irradiance conditions, we select some circumstances to illustrate MPP tracking method is located around the corresponding global
the tracking ability of the proposed MPPT method. The four shading MPP. It also indicates that the efficiency of the proposed method
patterns (SPs) we tested are listed in Table 2. For SP1, the irradiance is not heavily dependent on the initial conditions of the searching
on all the PV panels is uniform, as a result, only one peak exists in process.
the P–I characteristic curve of the PV array. For the other three SPs, The ability to find the global MPP for the new weather con-
there are multiple peaks. The parameters of the proposed MPPT ditions is very important especially for the places with frequent
method in the experiments are set as shown in Table 3. weather changes, such as in the tropics [28]. In order to illustrate
the tracking ability of the proposed ACO-based MPPT algorithm
under transient irradiance conditions, we considered three cases –
30
Number of iterations to converge

Case 1: SP changes from SP1 to SP2; Case 2: SP changes from SP1 to

25
400
20
Extracted power (W)

15 300

10
200
5
100
0 0.2 0.4 0.6 0.9
Q
0
A B C D
Fig. 10. Box-plot of the relationship between the number of iterations to converge Shading patterns
and the value of Q for the irradiance Case B in 200 runs; the edges of the box are the
25th and 75th percentiles of the iterations needed to converge, and the line inside Fig. 11. Power extracted by the ACO-based MPPT under four different shading
the box indicates the average number of iterations to converge. patterns.
234 L.L. Jiang et al. / Energy and Buildings 58 (2013) 227–236

Shading pattern SP1 Shading pattern SP2

400

Power (W)
200

0
1

0.5
D1

0
1
D2

0.5

0
4
I1(A)

0
4
I2(A)

0
0 1 2 3 4 5 6 7 8
Time (s)

Fig. 12. Transient response of the proposed MPPT method for Case 1: SP change from SP1 to SP2, K = 7, M = 4,  = 0.82, Q = 0.45.

Shading pattern SP1 Shading pattern SP3

400
Power (W)

200

0
1
D1

0.5

0
1
D2

0.5
0
4
I1(A)

0
4
I2(A)

2
0
0 1 2 3 4 5 6 7 8
Time (s)

Fig. 13. Transient response of the proposed MPPT method for Case 2: SP change from SP1 to SP3, K = 7, M = 4,  = 0.82, Q = 0.45.

Fig. 14. Transient response of the proposed MPPT method for Case 3: SP change from SP1 to SP4, K = 7, M = 4,  = 0.82, Q = 0.45.
 L.L. Jiang et al. / Energy and Buildings 58 (2013) 227–236 235

Table 5
The power obtained by different MPPT algorithms under different shading patterns.

Case # Power obtained in Watts

Ideal power CVT P&O PSO Proposed ACO

Case 1: SP1 to SP2 150.10 110.30 103.40 150.07 150.10

Case 2: SP1 to SP3 95.20 69.72 66.10 95.16 95.20
Case 3: SP1 to SP4 164.47 144.21 156.60 164.47 164.43

SP3, and Case 3: SP changes from SP1 to SP4. The sampling period of of PV array. From Table 5, we can see that PSO and ACO are both
the MPPT algorithm is set to 0.01 s. The power transient character- good choices for MPPT under partial shading conditions with rapid
istics, the corresponding duty cycle, and the current variable of each irradiance change. Both of PSO and ACO have four parameters to
PV string for Case 1. Cases 2 and 3 are shown in Figs. 12–14, respec- be determined. As the choice of the algorithm parameters influ-
tively. From these three performance figures for different transition ences the optimization performance of the applied method, there
cases, we can see that when the shading pattern changes from a is a slight difference in the power obtained by these two meth-
uniform condition to a partially shaded condition at 4 s (middle of ods as shown in Table 5. However, compared to the PSO-based
x-axis), the proposed MPPT algorithm can find the global MPP for MPPT method, our method has the advantage of faster conver-
the new shading pattern. For example, when the case changes are gence speed. It is reported that the required convergence iterations
SP1 to SP2, SP1 to SP3 and SP1 to SP4, power changes from 359 W of the PSO-based MPPT algorithm is less than 20 [16], while our
to 150 W, 359 W to 95 W and 359 W to 162 W are respectively method only requires less than 15 iterations to reach convergence.
obtained. By simulating all the cases, we found that the average In addition, ACO has not previously been applied to the problem of
iterations required to converge was less than 15. Since the samp- MPPT for PV systems, the proposed new method provides another
ling period of the MPPT algorithm influences the tracking speed, it optional method to maximize the power output from PV systems.
should be determined appropriately, according to the place where
PV systems are located. When setting the sampling period of the 5. Conclusions and future work
MPPT algorithm as 0.01 s, we measure that the average execution
time of the ACO-based MPPT algorithm is 7.3 ms by running the We proposed a novel ACO-based MPPT for PV systems under
algorithm 30 times, and the average runtime for the proposed MPPT partially shaded conditions. The configuration structure of the PV
algorithm to find the global MPP is less than 0.4 s. This means that array is also presented in order to work with the proposed method.
it can find the global MPP in a very short response time. This capa- The feasibility of the proposed MPPT for PV arrays under steady and
bility of fast response is very important for the systems located transient irradiance conditions are verified by simulation with var-
where there are rapidly changing irradiance conditions, such as in ious shading patterns. The results show that, compared to some
the tropics [28]. traditional MPPT methods, such as P&O and CVT, the proposed
ACO-based MPPT algorithm provides better performance to find the
4.2. Experiment 2 global MPP under various shading conditions. As a result, it ensures
to abstract more power from the PV arrays. In addition, it shows
This experiment compares the proposed ACO-based MPPT algo- some advantages over another recent popularly used method –
rithm with other existing MPPT techniques. As the efficiency of PSO-based MPPT. For instance, it requires less number of iterations
the MPPT algorithms significantly depends on the transient perfor- to converge and it is convergence independent of the initial con-
mance, we consider the cases when shading pattern changes from ditions. Since only a pair of current and voltage sensors are used,
a uniform distribution to other non-uniform pattern. Here, the pro- the proposed PV array structure also reduces the system cost. In
posed ACO-based MPPT algorithm is compared with three existing our future work, we plan to extend this work from a simulation
methods: constant voltage tracking (CVT) [29], P&O [1] and PSO study of the proposed ACO-based MPPT algorithm, to a prototype
[16]. Another popular MPPT technique (IncCond) [2] is not given implementation of a physical system.
here because it has the same disadvantage as the P&O algorithm, in
its lack of ability to track the global MPP. When testing the perfor- References
mance of the P&O and CVT based MPPT, each PV string is installed
with a MPPT. When testing PSO and ACO based MPPT, PV strings [1] L.L. Buciarelli, B.L. Grossman, E.F. Lyon, N.E. Rasmussen, The energy balance
associated with the use of a MPPT in a 100 kW peak power system, in: IEEE
share the same MPPT and only one pair of current and voltage Photovoltaic Specialists Conference, 1980, pp. 523–527.
sensors are required, which largely reduces the system costs. [2] K.H. Hussein, I. Muta, T. Hoshino, M. Osakada, Maximum photovoltaic power
Table 5 shows the power harvest for the four MPPT methods tracking: an algorithm for rapidly changing atmospheric conditions, in: IEE
Proceedings of Generation, Transmission and Distribution, vol. 142, 1995, pp.
when the shading pattern changes from case A to cases B, C, and D, 1350–2360.
respectively. Since the traditional P&O method lacks the ability to [3] T. Esram, P.L. Chapman, Comparison of photovoltaic array maximum power
track the global MPP, it easily becomes stuck at the local optimum point tracking techniques, IEEE Transactions on Energy Conversion 22 (2)
(2007) 439–449.
when the irradiance pattern changes. CVT assumes that the MPP [4] K. Kobayashi, I. Takano, Y. Sawada, A study of a two stage maximum power
is always located at the place of the 70–90% of the open voltage point tracking control of a photovoltaic system under partially shaded inso-
(Voc ). However, the ratio of the open circuit voltage and optimum lation conditions, Solar Energy Materials and Solar Cells 90 (18–19) (2006)
2975–2988.
voltage at the MPP under different irradiance conditions is not con-
[5] H. Patel, V. Agarwal, Maximum power point tracking scheme for PV systems
stant. A large amount of potential power is lost when applying this operating under partially shaded conditions, IEEE Transactions on Industrial
technology to PV systems located at places with large irradiance Electronics 55 (4) (2008) 1689–1698.
[6] Y.H. Ji, D.Y. Jung, J.G. Kim, J.H. Kim, T.W. Lee, C.Y. Won, A real maximum power
and temperature changes. Therefore, the power produced by these
point tracking method for mismatching compensation in PV array under par-
two methods is less than the power obtained by other two heuris- tially shaded conditions, IEEE Transactions on Power Electronics 26 (4) (2010)
tic search techniques. Compared to traditional MPPT methods, the 1001–1009.
heuristic algorithms such as PSO and ACO have many advantages. [7] M. Miyatake, T. Inada, I. Hiratsuka, H. Zhao, H. Otsuka, M. Nakano, Control char-
acteristics of a Fibonacci-search-based maximum power point tracker when a
For instance, they can track the global MPP with well-tuned param- photovoltaic array is partially shaded, in: The 4th International Power Elec-
eters and no requirement of knowledge about the characteristic tronics and Motion Control Conference, 2004, pp. 816–821.
236 L.L. Jiang et al. / Energy and Buildings 58 (2013) 227–236

[8] T.L. Nguyen, K.S. Low, A global maximum power point tracking scheme employ- [18] D. Picard, M. Cord, A. Revel, Image retrieval over networks: active learn-
ing DIRECT search algorithm for photovoltaic systems, IEEE Transactions on ing using ant algorithm, IEEE Transactions on Multimedia 10 (7) (2008)
Industrial Electronics 57 (10) (2010) 3456–3467. 1356–1365.
[9] C.H. Lin, C.H. Huang, Y.C. Du, J.L. Chen, Maximum photovoltaic power tracking [19] J. Zhang, H. Chung, W.L. Lo, T. Huang, Extended ant colony optimization algo-
for the PV array using the fractional-order incremental conductance method, rithm for power electronic circuit design, IEEE Transaction on Power Electronics
Applied Energy 88 (12) (2011) 4840–4847. 24 (1) (2009) 147–162.
[10] T. Mishima, T. Ohnishi, Power compensation system for partially shaded PV [20] K. Ishaque, Z. Salam, Syafaruddin, A comprehensive MATLAB Simulink PV sys-
array using electric double layer capacitors, in: IEEE Annual Conference on tem simulator with partial shading capability based on two-diode model, Solar
Industrial Electronics Society, 2002, pp. 3262–3267. Energy 85 (9) (2011) 2217–2227.
[11] E. Karatepe, T. Hiyama, M. Boztepe, M. Colak, Voltage based power com- [21] Woyte, J. Nijs, R. Belmans, Partial shadowing of photovoltaic arrays with differ-
pensation system for photovoltaic generation system under partially shaded ent system configurations: literature review and field test results, Solar Energy
insolation conditions, Energy Conversion and Management 49 (8) (2008) 74 (3) (2003) 217–233.
2307–2316. [22] E. Diaz-Dorado, A. Suarez-Garcia, C. Carrillo, J. Cidras, Influence of the shad-
[12] S. Subiyanto, A. Mohamed, M.A. Hannan, Intelligent maximum power point ows in photovoltaic systems with different configurations of bypass diodes, in:
tracking for PV system using Hopfield neural network optimized fuzzy logic International Symposium on Power Electronics Electrical Drives Automation
controller, Energy and Buildings 51 (2012) 29–38. and Motion, Pisa, Italy, 2010, pp. 134–139.
[13] A.B.G. Bahgat, N.H. Helwa, G.E. Ahmad, E.T. Shenawy, Maximum power point [23] M. Dorigo, Optimization, learning and natural algorithms, Ph.D. Thesis, Politec-
tracking controller for PV systems using neural networks, Renewable Energy nico di Milano, Italy, 1992.
30 (8) (2005) 1257–1268. [24] M. Dorigo, T. Stützle, Ant Colony Optimization, MIT Press, Cambridge, MA, 2004.
[14] L.L. Jiang, D.L. Maskell, J.C. Patra, Chebyshev functional link neural network- [25] K. Socha, M. Dorigo, Ant colony optimization for continuous domains, European
based modeling and experimental verification for photovoltaic arrays, in: Journal of Operational Research 185 (3) (2008) 1155–1173.
IEEE International Joint Conference on Neural Network, Australia, 2012, [26] T. Liao, M.A.M.d. Oca, D. Aydin, T. Stutzle, M. Dorigo, An incremental ant colony
pp. 1–8. algorithm with local search for continuous optimization, in: Proceedings of
[15] L.L. Jiang, D.L. Maskell, J.C. Patra, A FLANN - based controller for maximum the 13th annual conference on Genetic and evolutionary computation, ACM,
power point tracking in PV systems under rapidly changing conditions, in: IEEE Dublin, Ireland, 2011, pp. 125–132.
International Conference on Acoustics, Speech, and Signal Processing (ICASSP), [27] L. Chen, J. Shen, H. Chen, An improved ant colony algorithm in continuous opti-
Kyoto, Japan, 2012, pp. 2141–2144. mization, Journal of Systems Science and Systems Engineering 12 (2) (2003)
[16] M. Miyatake, M. Veerachary, F. Toriumi, N. Fujii, H. Ko, Maximum power point 224–235.
tracking of multiple photovoltaic arrays: a PSO approach, IEEE Transactions on [28] R. Jayaraman, D.L. Maskell, Temporal and spatial variations of the solar radiation
Aerospace and Electronic Systems 47 (1) (2011) 367–380. observed in Singapore, Energy Procedia 25 (2012) 108–117.
[17] D. Martens, M. De Backer, R. Haesen, J. Vanthienen, M. Snoeck, B. Baesens, [29] J.J. Schoeman, J.D.V. Wyk, A simplified maximal power controller for terres-
Classification with ant colony optimization, IEEE Transactions on Evolutionary trial photovoltaic panel arrays, in: IEEE Annual Power Electronics Specialists
Computation 11 (5) (2007) 651–665. Conference, Cambridge, MA, 1982, pp. 361–367.