A novel multi-model neuro-fuzzy-based MPPT for three-phase

grid-connected photovoltaic system
Aymen Chaouachi

, Rashad M. Kamel, Ken Nagasaka
Department of Electronic and Information Engineering, Tokyo University of Agriculture and Technology, 2-24-16, Nakamachi,
Koganei-shi, Tokyo 184-8588, Japan
Received 8 March 2010; received in revised form 9 July 2010; accepted 13 August 2010
Available online 29 September 2010
Communicated by: Associate Editor Elias Stefanakos
Abstract
This paper presents a novel methodology for Maximum Power Point Tracking (MPPT) of a grid-connected 20 kW photovoltaic (PV)
system using neuro-fuzzy network. The proposed method predicts the reference PV voltage guarantying optimal power transfer between
the PV generator and the main utility grid. The neuro-fuzzy network is composed of a fuzzy rule-based classifier and three multi-layered
feed forwarded Artificial Neural Networks (ANN). Inputs of the network (irradiance and temperature) are classified before they are fed
into the appropriated ANN for either training or estimation process while the output is the reference voltage. The main advantage of the
proposed methodology, comparing to a conventional single neural network-based approach, is the distinct generalization ability regard-
ing to the nonlinear and dynamic behavior of a PV generator. In fact, the neuro-fuzzy network is a neural network based multi-model
machine learning that defines a set of local models emulating the complex and nonlinear behavior of a PV generator under a wide range
of operating conditions. Simulation results under several rapid irradiance variations proved that the proposed MPPT method fulfilled the
highest efficiency comparing to a conventional single neural network and the Perturb and Observe (P&O) algorithm dispositive.
Ó 2010 Elsevier Ltd. All rights reserved.
Keywords: Neuro-fuzzy; MPPT; Multi-model; Grid-connected; Photovoltaic
1. Introduction
The Kyoto agreement on global reduction of greenhouse
gas emission has prompted interest on renewable energy
systems worldwide. Nowadays photovoltaic energy is one
of the most popular renewable sources since it is clean,
inexhaustible and requires little maintenance. However, it
still presents a vast area of competition comparing to con-
ventional energy resources due to its high cost and low effi-
ciency during energy conversion. Regarding to this, it is
necessary to optimize the performance of PV systems
through the operation of conversion systems to increase
the output efficiency of the overall system (Enrique et al.,
2007; Chua and Chen, 2009). This approach is commonly
named as Maximum Power Point Tracking (MPPT), sev-
eral methods are referred in the bibliography: the P&O
method is the most commonly used in practice due to its
simplicity and ease of implementation (Enrique et al.,
2010), however, this algorithm can fail or oscillate around
the Maximum Power Point (MPP) under sudden sunlight
changes (Hohm and Ropp, 2003; Salas et al., 2006). Incre-
mental conductance is also commonly used as it can over-
come some aspects of the P&O algorithm instabilities.
Nevertheless this method involves current and voltage dif-
ferentiation which requires a relatively complex decision
making process and therefore needs more complex calcula-
tion capacity and memory (Esram and Chapman, 2007).
More recently, Artificial Neural Network (ANN) tech-
niques are being employed for photovoltaic applications,
mainly because of their symbolic reasoning, flexibility
0038-092X/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.solener.2010.08.004

Corresponding author. Tel./fax: +81 42 388 7481.
E-mail addresses: A.chaouachi@gmail.com(A. Chaouachi), r_m_kamel
@yahoo.com (R.M. Kamel), bahman@cc.tuat.ac.jp (K. Nagasaka).
www.elsevier.com/locate/solener
Available online at www.sciencedirect.com
Solar Energy 84 (2010) 2219–2229
and explanation capabilities that are useful to deal with
strong nonlinearities and complex systems (Mellit and Kal-
ogirou, 2008). Mummadi et al. proposed an ANN-based
identification of the optimal operating point for a DC
motor supplied by PV generator (Veerachary and Yadaiah,
2000). Theodore et al. proposed a MPPT for solar vehicle
based on ANN reference voltage estimation (OCRAN
et al., 2005). Bahgat et al. proposed a neural network based
MPPT for a PV module supplying a DC motor that drives
an air fan (Bahgat et al., 2005). Preliminary results show
high MPPT efficiency for such methods (Hohm and Ropp,
2003). On the other hand, neural networks are still consid-
ered as unstable learning model regarding to the presence
of noisy sets, large training set, underfitting and overfitting
problems that causes a lack of generalization and trapping
on local minima solutions (Gao et al., 2005; Angueraet al.,
2007). Moreover, due to the high nonlinearity and close
dependence on weather conditions of PV generators, a neu-
ral network based model requires the use of a rich training
set covering a wide range of climatic conditions which leads
to a delicate generalization ability.
This paper presents a new MPPT methodology based on
a more robust estimator consisting of multi-model machine
learning. The multi-model approach defines a set of three
models by means of fuzzy classification, each model con-
sists of a single ANN that emulates the behavior of the
PV generator for a specific climatic conditions, so-called
“class”. The proposed classifier consists of transparent
rules-based fuzzy classifier, such rules, unlike opaque or
uninterruptable classification is straight forwarded and
simple to understand. The multi-model approach aims to
decrease the process complexity of the system under wide
operating conditions frame, therefore the proposed estima-
tor offers a distinct improvement regarding to the general-
ization capability comparing to the single ANN estimator.
Moreover, the proposed neuro-fuzzy predictor can be suc-
cessfully implemented for a real time MPPT that can han-
dle rapid weather conditions variations while a P&O
algorithm approach and a historical meteorological data
based prediction, yet easier to implement but usually fail
in case of sudden weather conditions variation.
2. Overview of the grid-connected PV system
Fig. 1 shows the overall of the studied system, composed
of a 20 kW photovoltaic generator connected to the main
electric grid via two power electronics stages. The first stage
consists of a DC–DC converter that insures impedance
adaptation between the generator and the load by tracking
the reference voltage estimated by the neuro-fuzzy net-
work. While the second stage is composed of a DC bus
and a PQ inverter that injects the power generated by the
PV generator into the main grid, where the inverter is act-
ing as power controller ensuring a high power factor. In the
following subsections, all the components of the grid-con-
nected PV system are described in details.
2.1. Photovoltaic system
The studied grid-connected PV system comprises a
20 kW photovoltaic generator installed at the Engineering
Campus of the Tokyo University of Agriculture and Tech-
nology (Fig. 2). A dynamic simulation model of the photo-
voltaic system is designed under Matlab Simulink
Ò
environment to analyze the behavior and the performances
of the proposed MPPT algorithm.
2.1.1. Simulation model of a PV generator
A simple equivalent circuit for a PV cell (Fig. 3a) con-
sists of a real diode associated to a parallel ideal current
source delivering a proportional current to the solar irradi-
ance, from where the following relations can be expressed
(Chaouachi et al., 2009):
I ¼ I
SC
ÀI
d
ð1Þ
where I
d
¼ I
0
ðe
eV
d
=gKT
c
À1Þ ð2Þ
Therefore I ¼ I
SC
ÀI
0
ðe
eV
d
=gKTc
À1Þ ð3Þ
Nomenclature
Photovoltaic parameters superscripted with “M” refers to a
PV module while those superscripted with “C”
are referring to a solar cell and each parameter un-
der scripted with “o” refers to standard conditions
e electron charge, e = 1.602 Â 10
19
(C)
I
d
diode current (A)
I
0
reverse saturation current (A)
I
sc
short-circuit current (A)
I
M
sc
o
short-circuit current under standard conditions
(A)
V
M
OC
o
open voltage under standard conditions (V)
V
T
thermal voltage in the semiconductor (V)
P
M
max
o
power under standard conditions (W)
N
S
cells in series in a module
N
P
parallel branches in a module
N
M
modules in series in an array
N
B
parallel branches in an array
T
a
ambient temperature (°C)
T
c
cell temperature (°C)
G
a
irradiance (W/m
2
)
R
s
series resistance (Q)
g diode ideality factor
K Boltzmann constant, if = 1.381 Â 10
À23
(J/K)
FF Fill Factor
2220 A. Chaouachi et al. / Solar Energy 84 (2010) 2219–2229
To deal with the dynamic behavior of a PV generator,
some parallel leakage resistance R
p
and a series resistance
R
s
(associated to the resistance of the semiconductor)
should be included (Fig. 3a). However, for the proposed
simulation, taking into account both of series and parallel
resistances results in computational limitations subjected to
the implicit nonlinear nature of the simulation model.
Therefore, such approach is usually avoided in the photo-
voltaic systems bibliography. For the proposed simulation
model, the parallel resistance is neglected while the series
resistance R
s
is considered which represents the internal
resistance and the connections between cells. The mathe-
matical relations established afterwards describing the PV
generator model is expressed as following:
I ¼ I
SC
ÀI
0
exp
eðV þI Á R
s
Þ
gKT
c
_ _
À1
_ _
ð4Þ
Considering the case of a photovoltaic generator module
(panel), composed of N
P
parallel branches, each including
N
S
solar cells associated in series, the current delivered by a
photovoltaic module under constant weather conditions
can be expressed as:
I
M
¼ I
M
sc
1 Àexp
V
M
ÀV
M
oc
þR
M
S
Á I
M
N
S
Á V
C
T
_ _ _ _
ð5Þ
The module current has an implicit expression depend-
ing on the following variables expressed in function of a
single cell parameters:
Short circuit current I
M
SC
¼ N
P
Á I
C
SC
Open circuit voltage V
M
oc
¼ N
S
Á V
C
oc
Thermal voltage in the
semiconductor
V
C
T
¼ mKT
c
=e
Equivalent serial resistance
of the module
R
M
s
¼ R
C
s
Á N
P
=N
S
As the first step, under the assumption that all the PV cells
are identical, the fundamental simulation parameters for a
PV PQ
Inverter
Neuro-
Fuzzy
network
PI
Irradiance
Temperature
PV
generator
DC-DC
converter
DC
Bus
Main grid
I
PV
V
PV
Inverter
*
PV
V
PV
L
PV
R
PV
C
DC
C
DC
V
PV
V
i
L
I
PV
PI
*
L
i
i
L
/
*
m
V
DC
V
1-m*
m*
D*
m
V
Fig. 1. Overview of the neuro-fuzzy MPPT grid-connected photovoltaic system.
Fig. 2. The studied photovoltaic field installed at Tokyo University of
Agriculture and Technology.
+
-
I
I
sc
V
I
d
R
p
-
+
-
Load
PV
Load
R
s
+
-
I
I
sc
V
I
d
Load
(a) (b) (c)
Fig. 3. (a) PV generator model, (b) Equivalent circuit with added R
p
parallel and R
s
series resistance (c) photovoltaic generator equivalent simple circuit.
A. Chaouachi et al. / Solar Energy 84 (2010) 2219–2229 2221
solar cell (in standard conditions) are expressed using the
module characteristics provided by the manufacturer’s cat-
alogue (Appendix A):
P
C
maxo
¼
P
M
maxo
N
S
N
P
ð6Þ
V
C
oco
¼
V
M
oc
o
N
S
ð7Þ
I
C
sco
¼
I
M
sco
N
P
ð8Þ
Now, the instantaneous series resistance of a solar cell can
be expressed as:
R
C
S
¼ 1 À
FF
P
C
maxo
V
C
oco
I
C
sco
_
_
_
_
Á
V
C
oco
I
C
sco
ð9Þ
where FF is the Fill Factor and is given by:
FF ¼
V
C
oco
V
C
To
ÀLn
V
C
oco
V
C
To
þ0:72
_ _ _ _
V
C
oco
V
C
To
þ1
ð10Þ
Operating condition parameters can be calculated for a
fixed voltage V
M
, temperature T
a
and irradiance G
a
using
the cell parameters in standard conditions. The short
circuit current is proportional to the irradiance and can
be expressed as:
I
C
SC
¼
I
C
SC
O
G
ao
Á G
a
ð11Þ
The open circuit voltage of the cell is depending on the
nominal open circuit voltage and the actual weather condi-
tions which can be expressed using the following relation:
V
C
oc
¼ V
C
oco
þ0:03 Á ðT
a
þ0:03 Á G
a
ÀT
C
o
Þ ð12Þ
Now, the instantaneous current debited by a photovol-
taic module can be finally expressed for the fixed parame-
ters (V
M
, T
a
, G
a
):
I
M
¼ N
P
Á I
C
SC
1 Àexp
V
M
ÀN
S
V
C
oc
þI
M
R
C
S
N
S
N
P
N
S
V
C
T
_ _ _ _
ð13Þ
Considering the case of a photovoltaic array that consists
in N
B
parallel branches, each containing N
M
modules asso-
ciated in series. For a V
M
applied voltage, the array current
is equal to I ¼

N
B
i¼1
I
M
i
. If we assume that all the panels are
identical and under the same temperature and irradiance,
then the total current generated by the array can be ex-
pressed as following:
I ¼ N
B
Á I
M
ð14Þ
2.1.2. Influence of temperature and irradiance on PV
operating
Fig. 4 shows the behavior of a photovoltaic panel
simulation in accordance to temperature and irradiance
0 5 10 15 20 25 30 35
0
1
2
3
4
5
Voltage [V]
C
u
r
r
e
n
t

[
A
]
(a) Effect of temperature under
constant irradiance of 1000 W/m2
0 5 10 15 20 25 30 35
0
20
40
60
80
100
120
(c) Effect of temperature under
constant irradiance of 1000 W/m2
Voltage [V]
P
o
w
e
r

[
W
]
0 5 10 15 20 25 30 35
0
1
2
3
4
5
(b) Effect of irradiance under
constant temperature of 25 C
Voltage [V]
C
u
r
r
e
n
t

[
A
]
45 C
35 C
25 C
45 C
35 C
25 C
0 5 10 15 20 25 30 35
0
20
40
60
80
100
120
(d) Effect of irradiance under
constant temperature of 25 C
1000 W/m2
700 W/m2
500 W/m2
700 W/m2
1000 W/m2
500 W/m2
Fig. 4. Temperature (a) and irradiance (b) effects on the IV curve of a PV generator temperature (c) and irradiance (d) effects on the power curve of a PV
generator.
2222 A. Chaouachi et al. / Solar Energy 84 (2010) 2219–2229
variations under respectively constant irradiance and tem-
perature. In fact, a PV generator connected to a load can
operate in a large margin of current and voltage depending
on weather conditions (Masters, 2004). Fig. 4 shows that
the open circuit voltage is increasing following a logarith-
mic relationship with the irradiance and decreasing slightly
as the cell temperature increases. On the other hand, the
short circuit current is linearly depending on the ambient
irradiance in direct proportion, while the open circuit volt-
age decrease slightly as the cell temperature increases.
Therefore, the maximum power that could be generated
by a PV system is slightly depending on the temperature
and irradiance variations: the maximum power increases
as the irradiance increases and vice versa, on the other
hand a PV generator performs better for low temperature
than raised one.
2.2. DC–DC boost converter
The operating point of a PV generator connected to a
load is the intersection of the load curve and the PV cur-
rent–voltage characteristics. In case of resistive load, the
load characteristic is a straight line with slop eðI ¼
1
R
V Þ,
while the power delivered to the load depends on the value
of the resistance only. It should be pointed out that if the
load R is equal to a certain optimal load R
opt
, the PV gen-
erator delivers the maximum power (MP). However, if the
load R is noticeably larger or smaller than R
opt
the PV gen-
erator operates respectively as constant voltage source or
constant current source.
To overcome this undesired effects on the PV power out-
put, an electrical tracking have to be achieved through a
power conditioning converter (DC–DC converter) inserted
between the load and the source to insure an impedance
adaptation by matching the load impedance with the vary-
ing PV source. In other words, the DC–DC boost converter
is used to maximize the energy transfer from the photovol-
taic generator to an electrical system by adjusting the PV
generator output voltage to a reference value (V
ref
) at
which the PV generator supplies the maximum power.
The proposed DC–DC converter includes two power
accumulation elements and thus two controllable variables
which are the V
pv
voltage and the inductor current (i
L
), a
mathematical model describing the DC–DC converter
can be expressed in the following relation (Kamel et al.,
2009):
V
m
i
dc
_ _
¼ m
V
DC
i
L
_ _
ð15Þ
where m = 1 À D and D is the duty cycle of the converter
switch expressed as:
D ¼
t
on
T
¼ t
on
f
s
ð16Þ
where f
s
is the switching frequency of the converter switch,
t
on
the time which the switch is turned on during a complete
period T.
di
L
t
¼
1
L
pv
ðV
m
ÀV
pv
Þ À
R
pv
L
pv
ð17Þ
dV
pv
dt
¼
1
pv
ði
L
Ài
pv
Þ ð18Þ
The voltage control loop with the PV current compensa-
tion gives the reference current i
Ã
L
that can be calculated
using Eq. (17):
i
Ã
L
¼ PIðV
Ã
pv
ÀV
pv
Þ þi
pv
ð19Þ
where V
Ã
pv
is the reference voltage calculated by the neuro-
fuzzy logic controller (Fig. 1), PI is the parameters of the
proportional-integral controller
The optimal switching voltage is expressed using (18)
and (19):
V
Ã
m
¼ PIðV
Ã
pv
ÀV
pv
Þ þV
pv
þ
R
pv
L
pv
i
L
ð20Þ
The controller parameters are chosen to maintain con-
stant PV voltage and to minimize the current ripple. The
DC–DC converter command is obtained by the inversion
of Eq. (15), expressed as following:
m
Ã
¼
V
Ã
m
V
DC
ð21Þ
With optimal duty cycle D
Ã
¼ 1 Àm
Ã
.
A DC link insures an energy balance between the power
generated by the PV generator and the power injected into
the network by charging or discharging the capacitor, that
oscillate between two levels depending on the actual cli-
matic conditions and the power injected.
2.3. PQ inverter
The inverter plays a vital role in grid-connected systems,
by interfacing the PV generator with the main utility power
system, the configuration of a basic inverter connected to a
PV generator is shown in Fig. 5.
The basic and minimum requirement of voltage source
inverter is to control the flow of active and reactive power
between the PV source and the main utility, the mathemat-
ical relations for P&Q magnitudes can be expressed as fol-
lowing in Eqs. (22) and (23) Lasseter et al., 2000:
P ¼
VE
wL
sinðd
V
Àd
E
Þ ð22Þ
Q ¼
V
2
wL
À
V
wL
cosðd
V
Àd
E
Þ ð23Þ
As the main objective of the current work is to analyze
the MPPT proposed methodology during steady-state
operation, it would be of no advantage to consider the
details of inverter switching. Moreover, if such details are
considered, then extra efforts should be done to compen-
sate the undesired effects of the generated harmonics that
leads to more heavy computational simulations. Thus,
the proposed inverter is represented by its control function
while the fast transients related with the commutations of
A. Chaouachi et al. / Solar Energy 84 (2010) 2219–2229 2223
the switches are not considered, therefore the power
injected at the AC side does not include harmonics, losses
or delays. The inverter voltage is represented using the fol-
lowing three controlled sinusoidal voltage sources defined
as:
V
a
¼
ffiffiffi
2
p
V Á sinðwt þd
V
Þ ð24Þ
V
b
¼
ffiffiffi
2
p
V Á sin wt þd
V
À
2p
3
_ _
ð25Þ
V
c
¼
ffiffiffi
2
p
V Á sin wt þd
V
þ
2p
3
_ _
ð26Þ
where the control variables are V and d
v
The basic structure of PQ inverter controller is shown in
Fig. 6, two PI controllers would suffice to control the flow
of active and reactive power by generating the proper val-
ues of voltage (V) and angle (d
v
) based on the instanta-
neous value of voltage and current. The reference power
(P
ref
) in Fig. 6 represents the amount of active power pro-
duced by the photovoltaic generator interfaced to the main
utility through the inverter, while Q
ref
represents the
amount of reactive power desired to be injected into or
absorbed from the main utility. In the present case, the
inverter is operating at unity power factor (Q
ref
= 0) there-
fore no reactive power is exchanged and the total MP
extracted from the PV generator is injected to the grid.
3. Neuro-fuzzy maximum power point estimation
The neuro-fuzzy network consists of two stages; the first
one is a fuzzy-rule-based classifier while the second one is
composed of three multi-layered feed forwarded ANNs
(Fig. 7). The three ANNs have similar architecture, com-
posed of three layers: input, hidden and output layers.
3.1. Fuzzy rules-based classification
Classification task consists of assigning a class C
j
from a
predefined class set C = {C
1
, C
2
, . . . , C
M
} to an object
belonging to a certain feature space x e S
N
, so the problem
comes to find a mapping defined as (Kuncheva, 2000):
D : S
N
!C
A fuzzy rule-based classification system is composed of
fuzzy reasoning method (FRM) and a Knowledge Base
(KB) consisting in rule base (RB) and data base that
Fig. 5. Basic inverter connected to a PV generator.
Fig. 6. Basic structure of the inverter PQ control scheme.
ANN
1
ANN
2
ANN
3
Temp
Temp
Temp
Irrd
Irrd
Irrd
Reference
Voltage
Fuzzy Logic
Classifier
Irrd.
Temp.
Class 1
Class 2
Class 3
Fig. 7. Architecture of the Neuro–fuzzy network.
2224 A. Chaouachi et al. / Solar Energy 84 (2010) 2219–2229
describes the semantic of the fuzzy subsets associated to
linguistic labels on the if-then part of the rules. Basing on
the KB, the FRM determines a label class for admissible
patterns as shown in Fig. 8 (Cordon and Jesus, 1999).
Let us consider a rule base R = {R
1
, R
2
, . . . , R
L
} and r
j
rules in R, the RBs are generated based on fuzzy rules with
a class in the consequent according to the following
structure:
R
k
if a
1
is L
k
1
and . . . a
N
then O is C
j
ð27Þ
where a
1
, . . . , ,a
N
is the outstanding selected features for
the classification task, L
k
1
; . . . ; L
k
N
is the linguistic labels to
discretize the continuous domain of the variables, O is
the class C
j
to which the pattern belongs.
3.1.1. Fuzzy reasoning method
A fuzzy reasoning method is an inference procedure that
derives conclusions from a set of if-then rules and a pattern,
such methodology aims to improve the generalization
capability of a classification system (Kuncheva, 2000).
While each rule r
j
in R generates a class C
j
related to a pat-
tern P
t
¼ ðp
t
1
; . . . p
t
N
Þ, the FRM considers the rule with the
highest combination between the matching degree of the
pattern with the if-then part and the certainty degree for
the classes. In fact, for each class C
j
, the association degree
of the pattern with the class O
j
can be expressed as
following:
O
j
¼ max
k2L
R
k
ðP
t
Þ ð28Þ
where R
k
(P
t
) is the strength of activation (matching rule)
associated to the kth rule, obtained by applying a t-norm
to the degree of satisfaction of the inputs patterns to the
clause (a
1
is L
k
1
):
R
k
ðP
t
Þ ¼ Tðl
k
L
1
ðp
t
1
Þ; . . . ; l
k
L
N
ðp
t
N
ÞÞ ð29Þ
Finally the classification for the pattern E
t
is the class C
h
such expressed in:
O
h
¼ max
j¼1;...;M
O
j
ð30Þ
This approach can be graphically represented by Fig. 9.
3.1.2. Fuzzy controller design
The fuzzy rule-based classifier was build and designed
under the Matlab Fuzzy Logic Toolbox and lately exported
into Simulink environment as a part of the neuro-fuzzy net-
work and the whole simulation system.
3.1.2.1. Membership functions. A membership function is a
curve that defines how each point in the input or output
space is mapped to a membership value between 0 and 1.
Fig. 10a and b shows the temperature and irradiance mem-
bership functions in trapezoidal shape. The temperature
membership function is sorted into three categories labeled
as: hot, warm and cold, while the irradiance membership
function is sorted into three categories labeled as sunny,
normal and cloudy.
The output of the fuzzy system is the class associated to
the inputs pattern. Fig. 10c shows the output membership
function in triangular shape, sorted into three categories:
class1, class2 and class3.
3.1.2.2. Fuzzy rules. The influence of climatic conditions on
the model or behavior of a PV generator can be interpreted
through linguistic conditional statements that describe the
correlation between the climatic conditions and the PV
operating behavior. Indeed, if-then statement based trans-
parent rules, similar to those used in fuzzy control (Lee,
1990) are established based on human expertise on the
influence of weather conditions on PV systems operating
characteristics that are summarized in Table 1.
Fig. 11 shows the clustering output of the fuzzy classifier
as a function of output power, the training data set was
split into three classes each one is used for the training pro-
cess of its appropriate ANN. Unlike conventional crisp
classification, the proposed fuzzy classifier presents several
patterns with similar power level that are classified into dis-
similar classes as they belongs to different weather condi-
tions (temperature, irradiance).
3.2. Multi-layred perceptron neural network
Multi-layered perceptrons neural network (MLPNN)
has been applied successfully to solve some difficult and
diverse problems based on a preliminary supervised train-
ing with error back propagation algorithm using an error
Class Pattern
Reasoning
Method
Data
Base
Rule
Base
Knowledge Base
Fig. 8. Fuzzy rule-based classification system.
R
1
R
2
.
.
.
.
.
R
k
.
.
.
R
L
Class
Pattern
Fig. 9. Fuzzy reasoning method.
A. Chaouachi et al. / Solar Energy 84 (2010) 2219–2229 2225
correction learning rule. Basically, error back learning con-
sists in two pass through the different layers of the network,
a forward pass and backward pass. In the forward pass, an
activity pattern (input vector) is applied to the sensory
nodes of the network, its effect propagates through the net-
work layer by layer to produce an output as actual
response. During the backward pass synaptic weights are
adjusted in accordance to an error correction-rule. The
error signal (subtracted from a desired value) is then prop-
agated backward through the network against the direction
of the synaptic connections (Haykin, 1999; Hornik et al.,
1989).
To evaluate the performance of the proposed neuro-
fuzzy network comparing to a conventional neural network
based MPPT method, a single ANN estimator was devel-
oped besides the three ANN’s (ANN1, ANN2 and
ANN3) constituting the neuro-fuzzy learning machine.
The networks inputs are temperature and irradiance while
the output is the optimal reference voltage.
The training data base was recorded during the year
2008 for the training and testing process the original set
was split into three subsets: training set (70%), validation
set (15%) used during the training process for early stop-
ping to avoid overfitting problems and testing set (15%)
to evaluate the generalization performance of the devel-
oped ANNs. To improve the training convergence perfor-
mances, mean 0 and standard deviation 1 based across
channel normalization (Chaouachi et al., 2010) was used
for the input training set rescaling, basing on the following
relation:
S
i
¼
x
i
Àmean
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

N
i¼1
ðxi ÀmeanÞ
2
NÀ1
_ ð31Þ
where
mean ¼

N
iÀ1
x
i
N
x
i
the raw input variable X in the ith training case, S
i
the
standardized value corresponding to x
i
, N the number of
training case
The target data set was linearly normalized in order to
force the network values to be within the range of output
activation functions using upper (Y
max
) and lower bounds
(Y
min
) for the values:
Z
i
¼
Y
i
ÀðY
max
ÀY
min
Þ=2
ðY
max
ÀY
min
Þ=2
ð32Þ
where Y
i
the raw target variable Y for the ith training case.
Z
i
the standardized value corresponding to Y
i
.
The backpropagation momentum algorithm was used
for the training process of the developed networks, the
framework of the ANN’s was set out basing on trial and
error approach (Table 2). In fact, the networks were
trained up to the stopping validation point while their
parameters were continuously changed until no significant
error performance decrease is observed. It should be
noticed that a single hidden layer was sufficient for the pro-
posed task, as adding a second hidden layer did not
improve the training performances.
Class 1 Class 2 Class 3
1.0
0 0.2 0.4 0.6 0.8 1
Temperature °C
Warm Cold Hot
1.0
10 20 30 40 0
Cloudy Normal
0.1 0.2 0.4 0.6 0.8 1 0
Sunny
1.0
Irradiance kW/m
2
(a) (b) (c)
Fig. 10. Inputs and output membership functions of the fuzzy classifier.
Table 1
Fuzzy rules assignment.
Temperature Irradiance
Cloudy Normal Sunny
Cold Class2 Class3 Class3
Warm Class1 Class2 Class3
Hot Class1 Class1 Class2
0 2 4 6 8 1 12 14 16 18 20
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Fuzzy classification
Power generated [kWh]
C
l
a
s
s

o
u
t
p
u
t
Class 3
Class 2
Class 1
Fig. 11. Clustering of the input patterns (training data).
Table 2
Neural networks training parameters.
ANN1 ANN2 ANN3 Single ANN
Hidden neurons 11 14 11 16
Learning rate 0.1 0.1 0.1 0.2
Momentum 0.3 0.3 0.3 0.35
2226 A. Chaouachi et al. / Solar Energy 84 (2010) 2219–2229
4. Simulation results
Several performance criteria are reported in the ANN
literature as: the training time, the modeling time and the
prediction error. In the present study, as the training pro-
cess is in offline mode, the first two criteria are not consid-
ered to be relevant. Thereby, the estimation performances
of the neuro-fuzzy network and the single ANN will be
evaluated only in term of estimation error defined as the
difference between the experimental and the estimated val-
ues based on statistical approach. In fact the mean absolute
error (MAE), was applied as statistical error test criteria.
On the other hand, the performance of the proposed esti-
mator was compared with the experimental algorithm
(P&O) and the conventional ANN estimator regarding to
the maximum power extracted during a whole day.
Fig. 12a and b respectively shows the reference voltage
estimation by the neuro-fuzzy network and a single ANN
comparing to the experimental data for an unsteady
weather conditions during a cloudy day. We can see that
the proposed machine learning achieved the most accurate
estimation. Moreover, for the total testing process mean
absolute error for the neuro-fuzzy estimator is 0.139 while
the single ANN estimator’s error is equal to 2.496. In fact,
the complex nature of the PV generator and its nonlinear
behavior versus the atmospheric changes involves the use
of a large training data base to cover most of eventual
weather conditions that leads to classification degradation
and loss in the generalization performance caused by over-
fitting problem. Nevertheless, the proposed neuro-fuzzy
06 08 10 12 14 16 18
260
270
280
290
300
310
Time [hour]
V
o
l
t
a
g
e

[
V
]
Voltage reference
Voltage estimation
06 08 10 12 14 16 18
260
270
280
290
300
310
Time [hour]
V
o
l
t
a
g
e

[
V
]
Voltage estimation
Votage reference
06 08 10 12 14 16 18
0
4
8
12
16
Time [hour]
P
o
w
e
r

[
k
W
]
P&O
Neuro−fuzzy MPPT
Single ANN MPPT
(a)
(b)
(c)
Fig. 12. One day operation performances: (a) Neuro-fuzzy reference
voltage estimation. (b) Single ANN reference voltage estimation. (c)
Power generation comparison using neuro-fuzzy, single ANN, and P&O
algorithm.
0 0.05 0.1
0
20
40
60
Time [s]
C
u
r
r
e
n
t

[
A
]
PV generated current
0 0.05 0.1
0
100
200
300
400
Tracking of the reference voltage
Time [s]
V
o
l
t
a
g
e

[
V
]
0 0.05 0.1
0
5
10
15
Power generated and injected into the grid
Time [s]
P
o
w
e
r

[
k
W
]
0 0.05 0.1
0
100
200
300
400
RMS Voltage
Time [s]
V
o
l
t
a
g
e

[
V
]
PV power
Injected power
PV voltage
Ref voltage
(a)
(c)
(b)
(d)
Fig. 13. System operation during a weather condition step: (a) Tracking of the reference voltage. (b) Power generated and injected into the grid. (c) PV
generated current. (d) RMS voltage.
A. Chaouachi et al. / Solar Energy 84 (2010) 2219–2229 2227
machine learning can overcome these limitations through a
better definition of the model complexity based on the
fuzzy classification.
Fig. 12c presents the simulation output of the PV system
(extracted power) during 1 day using the neuro-fuzzy net-
work and the single ANN comparing to the measured
experimental data (P&O). It can be seen that the neuro-
fuzzy MPPT methodology accomplished better perfor-
mances then the single ANN or the P&O’s algorithm that
can fail to track the MPP or oscillates around it under rap-
idly changing climatic conditions. In fact, the proposed
neuro-fuzzy-based method achieved the highest power effi-
ciency with 6.85% of extra generated power comparing to
the single ANN and 2.73% comparing to the experimental
dispositive using the P&O algorithm.
In order to validate the stability performances of the
proposed MPPT based neuro-fuzzy estimator, we pro-
ceeded to observe the tracking of the MP during the grid
connection operation. Fig. 13a shows the tracking of the
reference voltage estimated by the neuro-fuzzy machine
learning after a weather condition step change (irradiance
650 ?500 kW/m
2
; temperature 26.5 ?24.5 °C) basing on
experimental measures. In fact, the neuro-fuzzy network
supplies a new value of reference voltage, tracked through
the DC–DC stage so that the overall system can keep
operating under optimal conditions, that is to say that
the power injected into the main utility is equal to the
MP Fig. 13b throughout the steady-state next to the
climatic disturbance. Fig. 13c shows the variation of the
current generated by the PV array after the climatic per-
turbation basing on the new set point of the system opti-
mal operation meanwhile the RMS voltage keep stable
(Fig. 13d).
Fig. 14 shows the current injected into the main grid and
the grid side voltage before and after the weather condi-
tions step change. As it can be seen, the voltage and current
are in phase which means that the MP extracted from the
PV array can pass into the main grid as the whole system
operates at unity power factor (Q
ref
= 0) with no reactive
power exchange. The peak occurring approximately at
0.016 s is due to the dynamic state during the initialization
of the system (DC–DC converter, DC bus, inverter. . .).
After reaching the steady-state (0.02 s) such peak will not
occur even during sudden weather condition changes like
the one occurring at 0.07 s.
5. Conclusions
In this paper a new MPPT methodology was applied to
a grid-connected photovoltaic system based on a pro-
posed neuro-fuzzy estimator. The whole system was sim-
ulated under Matlab Simulink
Ò
environment. In this
study, the developed neuro-fuzzy network consists of
two stages; the first one is a fuzzy rule-based classifier,
the second one is composed of three multi-layered feed
forwarded ANNs trained offline using experimental data
from a real PV system installed at the engineering campus
of Tokyo University of Agriculture and Technology. The
proposed neuro-fuzzy estimator showed the ability to
faithfully emulate the dynamic and nonlinear behavior
of a photovoltaic generator under a large wide of climatic
conditions. In fact the multi-model aspect of the proposed
machine learning confer it a distinct generalization ability
comparing to a conventional single ANN-based MPP pre-
dictor. Maximum power operation was achieved by track-
ing the reference voltage estimated by the neuro-fuzzy
network through a DC–DC converter. The whole grid-
connected system performance was tested during a cloudy
day with several rapid irradiance variations. The simula-
tion results showed that the proposed system perfor-
mances was not degraded, as the MPPT dispositive was
able to track the reference voltage insuring an optimal
operating condition under any rapid changing meteoric
conditions.
Appendix A. PV module specification
Manufacturer SHARP Maximum power 120 W
Type NE-LO1A V
OC
o
30.90 V
Maximum
voltage
600 V I
OC
o
4.17 A
Weight 12.50 Kg V
pm
26.70 V
0 0.02 0.04 0.06 0.08 0.1
−200
0
200
400
600
Time [ s ]
V
o
l
t
a
g
e

[

V

]
Grid voltage
0 0.02 0.04 0.06 0.08 0.1
−40
−20
0
20
40
60
Injected current
Time [ s ]
C
u
r
r
e
n
t

[

A

]
Ia
Ib
Ic
Va
Vb
Vc
Fig. 14. Grid side voltage and injected current during a weather condition step.
2228 A. Chaouachi et al. / Solar Energy 84 (2010) 2219–2229
References
Anguera, X., Shinozaki, T., Wooters, C., Hernando, J., 2007. Model
complexity selection and cross-validation EM training for Robust
speaker diarization. In: Proc of ICASSP.
Bahgat, A.B.G., Helwa, N.H., Ahmad, G.E., El Shenawy, E.T., 2005.
Maximum power point tracking controller for pv systems using neural
networks. Renewable Energy 30, 1257–1268.
Chaouachi, A., Kamel, R.M., Nagasaka, K. 2009. Modeling and simulation
of a photovoltaic field based on matlab simulin. In: Conference on Energy
System, Economy and Environment, Tokyo, pp. 70–74.
Chaouachi, A., Kamel, R.M., Nagasaka, K., 2010. Neural network
ensemble-based solar power generation short-term forecasting. Journal
of Advanced Computational Intelligence and Intelligent Informatics
14 (1), 69–75.
Chua, C.C., Chen, C.L., 2009. Robust maximum power point tracking
method for photovoltaic cells: a sliding mode control approach. Solar
Energy 83, 1370–1378.
Cordon, O., Jesus, F.H.M.J., 1999. A proposal on reasoning methods in
fuzzy rule-based classification systems. International Journal of
Approximate Reasoning 20, 21–45.
Enrique, J.M., Duran, E., de-Cardona, M.S., Andujar, J.M., 2007.
Theoretical assessment of the maximum power point tracking
efficiency of photovoltaic facilities with different converter topologies.
Solar Energy 81, 31–38.
Enrique, J.M., Andujar, J.M., Bohorquez, M.A., 2010. A reliable, fast and
low cost maximum power point tracker for photovoltaic applications.
Solar Energy 84, 79–89.
Esram, T., Chapman, P.L., 2007. Comparison of photovoltaic array
maximum power point tracking techniques. IEEE Transactions on
Energy Conversion 22 (2), 439–449.
Gao, Z., Ming, F., Hongling, Z., 2005. Bagging neural networks for
predicting water consumption. Journal of Communication and Com-
puter 2 (3), 19–24 (Serial No. 4).
Haykin, S., 1999. Neural Networks, A Comprehensive Foundation,
Second ed. Springer.
Hohm, D.P., Ropp, M.E., 2003. Comparative study of maximum power
point tracking algorithms. Progress in Photovoltaics: Research and
Applications, 47–62.
Hornik, K.M., Stinchcombe, M., White, H., 1989. Multilayer feedforward
networks are universal approximators. Neural Networks 2 (2), 359–
366.
Kamel, R.M., Chaouachi, A., Nagasaka, K., 2009. Design and imple-
mentation of various inverter controls to interface distributed gener-
ators (DGs) in micro grids. In: Conference on Energy System,
Economy and Environment, Tokyo, pp. 60-64.
Kuncheva, L.I., 2000. Fuzzy Classifier Design. Physica-Verlag,
Heidelberg.
Lasseter, R. Tomsovic, K., Piagi, P., 2000. Scenarios for Distributed
Technology Applications with Steady State and Dynamic Models of
Loads and Micro-Sources. CERTS Report.
Lee, C.C., 1990. Fuzzy logic in control systems. IEEE Transactions on
Systems, Man & Cybernetics SMC 20 (2), 404–435.
Masters, G.M., 2004. Renewable and Efficient Electric Power Systems.
Wiley Interscience.
Mellit, A., Kalogirou, S.A., 2008. Artificial intelligence techniques for
photovoltaic applications: a review. Progress in Energy and Combus-
tion Science 34, 574–632.
OCRAN, T.A., Junyi, C., Binggang, C., Xinghua, S., 2005. Artificial
neural network maximum power point tracker for solar electric
vehicle. Tsinghua Science and Technology 10 (2), 204–208 (ISSN 1007-
0214 12/23).
Salas, V., Olias, E., Barrado, A., Lazaro, A., 2006. Review of the
maximum power point tracking algorithms for stand-alone pho-
tovoltaic systems. Solar Energy Materials & Solar Cells 90, 1555–
1578.
Veerachary, M., Yadaiah, N., 2000. ANN based peak power tracking for
PV supplied DC motors. Solar Energy 69, 343–350.
A. Chaouachi et al. / Solar Energy 84 (2010) 2219–2229 2229

all the components of the grid-connected PV system are described in details.1. therefore the proposed estimator offers a distinct improvement regarding to the generalization capability comparing to the single ANN estimator. 2005)..2220 A. 3a) consists of a real diode associated to a parallel ideal current source delivering a proportional current to the solar irradiance. Simulation model of a PV generator A simple equivalent circuit for a PV cell (Fig. Theodore et al. Mummadi et al. 2005. 2. proposed a neural network based MPPT for a PV module supplying a DC motor that drives an air fan (Bahgat et al. a neural network based model requires the use of a rich training set covering a wide range of climatic conditions which leads to a delicate generalization ability. 2. so-called “class”. due to the high nonlinearity and close dependence on weather conditions of PV generators.. 2003). 2008). 2009): I ¼ I SC À I d where I d ¼ I 0 ðe eV d =gKT c ð1Þ À 1Þ eV d =gKT c ð2Þ À 1Þ ð3Þ Therefore I ¼ I SC À I 0 ðe . 2005). 2000).. if = 1. This paper presents a new MPPT methodology based on a more robust estimator consisting of multi-model machine learning. The first stage consists of a DC–DC converter that insures impedance adaptation between the generator and the load by tracking the reference voltage estimated by the neuro-fuzzy network. 2.1.. underfitting and overfitting problems that causes a lack of generalization and trapping on local minima solutions (Gao et al. 2007). large training set. such rules. neural networks are still considered as unstable learning model regarding to the presence of noisy sets. A dynamic simulation model of the photovoltaic system is designed under Matlab SimulinkÒ environment to analyze the behavior and the performances of the proposed MPPT algorithm. The proposed classifier consists of transparent rules-based fuzzy classifier. Moreover. The multi-model approach aims to decrease the process complexity of the system under wide operating conditions frame. On the other hand. The multi-model approach defines a set of three models by means of fuzzy classification. yet easier to implement but usually fail in case of sudden weather conditions variation. Bahgat et al. Preliminary results show high MPPT efficiency for such methods (Hohm and Ropp. each model consists of a single ANN that emulates the behavior of the PV generator for a specific climatic conditions. Chaouachi et al. Photovoltaic system The studied grid-connected PV system comprises a 20 kW photovoltaic generator installed at the Engineering Campus of the Tokyo University of Agriculture and Technology (Fig. unlike opaque or uninterruptable classification is straight forwarded and simple to understand.. where the inverter is acting as power controller ensuring a high power factor. composed of a 20 kW photovoltaic generator connected to the main electric grid via two power electronics stages. / Solar Energy 84 (2010) 2219–2229 Nomenclature Photovoltaic parameters superscripted with “M” refers to a PV module while those superscripted with “C” are referring to a solar cell and each parameter under scripted with “o” refers to standard conditions e electron charge. In the following subsections.602 Â 1019 (C) Id diode current (A) I0 reverse saturation current (A) Isc short-circuit current (A) I Mo short-circuit current under standard conditions sc (A) VM o open voltage under standard conditions (V) OC VT thermal voltage in the semiconductor (V) PM o max NS NP NM NB Ta Tc Ga Rs g K FF power under standard conditions (W) cells in series in a module parallel branches in a module modules in series in an array parallel branches in an array ambient temperature (°C) cell temperature (°C) irradiance (W/m2) series resistance (Q) diode ideality factor Boltzmann constant.1. Anguera et al. proposed a MPPT for solar vehicle based on ANN reference voltage estimation (OCRAN et al. Overview of the grid-connected PV system Fig. Moreover. e = 1. the proposed neuro-fuzzy predictor can be successfully implemented for a real time MPPT that can handle rapid weather conditions variations while a P&O algorithm approach and a historical meteorological data based prediction. 1 shows the overall of the studied system. from where the following relations can be expressed (Chaouachi et al.381 Â 10À23 (J/K) Fill Factor and explanation capabilities that are useful to deal with strong nonlinearities and complex systems (Mellit and Kalogirou. 2). proposed an ANN-based identification of the optimal operating point for a DC motor supplied by PV generator (Veerachary and Yadaiah. While the second stage is composed of a DC bus and a PQ inverter that injects the power generated by the PV generator into the main grid.

the current delivered by a photovoltaic module under constant weather conditions can be expressed as:   M  V À V M þ RM Á I M oc S I M ¼ I M 1 À exp ð5Þ sc NS Á V C T The module current has an implicit expression depending on the following variables expressed in function of a single cell parameters: Short circuit current Open circuit voltage Thermal voltage in the semiconductor Equivalent serial resistance of the module IM ¼ NP Á IC SC SC V M ¼ NS Á V C oc oc V C ¼ mKT c =e T RM ¼ RC Á N P =N S s s Fig. composed of NP parallel branches. 3a). (a) PV generator model. / Solar Energy 84 (2010) 2219–2229 2221 PV generator DC-DC converter DC Bus Inverter Irradiance PV Temperature IPV iL C PV LPV RPV Main grid VDC C DC Vm VPV IPV NeuroFuzzy network D* PQ Inverter iL PI VPV PI 1-m* m* / * Vm V * PV i * L VDC Fig. The matheI Id Isc + As the first step. (b) Equivalent circuit with added Rp parallel and Rs series resistance (c) photovoltaic generator equivalent simple circuit. 2.A. Chaouachi et al. the fundamental simulation parameters for a V Rs + Isc Id I + V Rp Load PV - (a) Load (b) - (c) Load - Fig. The studied photovoltaic field installed at Tokyo University of Agriculture and Technology. taking into account both of series and parallel resistances results in computational limitations subjected to the implicit nonlinear nature of the simulation model. To deal with the dynamic behavior of a PV generator. the parallel resistance is neglected while the series resistance Rs is considered which represents the internal resistance and the connections between cells. Overview of the neuro-fuzzy MPPT grid-connected photovoltaic system. For the proposed simulation model. However. some parallel leakage resistance Rp and a series resistance Rs (associated to the resistance of the semiconductor) should be included (Fig. Therefore. such approach is usually avoided in the photovoltaic systems bibliography. for the proposed simulation. 3. 1. each including NS solar cells associated in series. under the assumption that all the PV cells are identical. . matical relations established afterwards describing the PV generator model is expressed as following:     eðV þ I Á Rs Þ À1 ð4Þ I ¼ I SC À I 0 exp gKT c Considering the case of a photovoltaic generator module (panel).

. the instantaneous series resistance of a solar cell can be expressed as: 0 1 C FF A V oco RC ¼ @ 1 À P C Á C ð9Þ S maxo I sco C C V oco I sco Now. Chaouachi et al. temperature Ta and irradiance Ga using the cell parameters in standard conditions. Ta. each containing NM modules associated in series. 4 shows the behavior of a photovoltaic panel simulation in accordance to temperature and irradiance (a) Effect of temperature under constant irradiance of 1000 W/m2 5 4 5 4 (b) Effect of irradiance under constant temperature of 25 C Current [A] Current [A] 3 2 1 0 0 5 10 15 20 45 C 35 C 25 C 3 2 1 1000 W/m2 700 W/m2 500 W/m2 25 30 35 0 0 5 10 15 20 25 30 35 Voltage [V] Voltage [V] (c) Effect of temperature under constant irradiance of 1000 W/m2 120 100 120 100 80 45 C 35 C 25 C (d) Effect of irradiance under constant temperature of 25 C Power [W] 80 60 40 20 0 0 5 10 15 20 60 40 20 30 35 0 0 5 10 15 20 700 W/m2 1000 W/m2 500 W/m2 25 25 30 35 Voltage [V] Fig. 4. Temperature (a) and irradiance (b) effects on the IV curve of a PV generator temperature (c) and irradiance (d) effects on the power curve of a PV generator. The short circuit current is proportional to the irradiance and can be expressed as: 2. Influence of temperature and irradiance on PV operating Fig. If we assume that all the panels are i identical and under the same temperature and irradiance. Ga): " !# NS V M À N S V C þ I M RC N P oc S M C I ¼ N P Á I SC 1 À exp ð13Þ NSV C T Considering the case of a photovoltaic array that consists in NB parallel branches. then the total current generated by the array can be expressed as following: I ¼ NB Á IM ð14Þ where FF is the Fill Factor and is given by:    VCo VC oc À Ln V oco þ 0:72 C VC To To FF ¼ C V oco þ1 VC To ð10Þ Operating condition parameters can be calculated for a fixed voltage VM. the array current For NB is equal to I ¼ i¼1 I M .1. / Solar Energy 84 (2010) 2219–2229 solar cell (in standard conditions) are expressed using the module characteristics provided by the manufacturer’s catalogue (Appendix A): PC o max PM ¼ maxo N SN P V Mo oc NS ð6Þ ð7Þ ð8Þ IC ¼ SC IC O SC Á Ga Gao ð11Þ The open circuit voltage of the cell is depending on the nominal open circuit voltage and the actual weather conditions which can be expressed using the following relation: V C ¼ V C o þ 0:03 Á ðT a þ 0:03 Á Ga À T Co Þ oc oc ð12Þ VCo ¼ oc IC o ¼ sc I Mo sc NP Now.2.P a VM applied voltage.2222 A. the instantaneous current debited by a photovoltaic module can be finally expressed for the fixed parameters (VM.

the maximum power that could be generated by a PV system is slightly depending on the temperature and irradiance variations: the maximum power increases as the irradiance increases and vice versa. On the other hand. PI is the parameters of the proportional-integral controller The optimal switching voltage is expressed using (18) and (19): V à ¼ PIðV à À V pv Þ þ V pv þ m pv Rpv iL Lpv ð20Þ The controller parameters are chosen to maintain constant PV voltage and to minimize the current ripple. The DC–DC converter command is obtained by the inversion of Eq. on the other hand a PV generator performs better for low temperature than raised one. 1). It should be pointed out that if the load R is equal to a certain optimal load Ropt. However. The proposed DC–DC converter includes two power accumulation elements and thus two controllable variables which are the Vpv voltage and the inductor current (iL). the 1 load characteristic is a straight line with slop eðI ¼ R V Þ. 2004). Thus. the proposed inverter is represented by its control function while the fast transients related with the commutations of . an electrical tracking have to be achieved through a power conditioning converter (DC–DC converter) inserted between the load and the source to insure an impedance adaptation by matching the load impedance with the varying PV source. In other words. (17): ià ¼ PIðV à À V pv Þ þ ipv L pv ð19Þ where V à is the reference voltage calculated by the neuropv fuzzy logic controller (Fig. Fig. it would be of no advantage to consider the details of inverter switching. 2009):     Vm V DC ¼m ð15Þ idc iL where m = 1 À D and D is the duty cycle of the converter switch expressed as: D¼ ton ¼ ton fs T ð16Þ diL 1 Rpv ¼ ðV m À V pv Þ À Lpv t Lpv dV pv 1 ¼ ðiL À ipv Þ pv dt ð17Þ ð18Þ The voltage control loop with the PV current compensation gives the reference current ià that can be calculated L using Eq. (15). Therefore. 5. the mathematical relations for P&Q magnitudes can be expressed as following in Eqs. As the main objective of the current work is to analyze the MPPT proposed methodology during steady-state operation.3. 2. 4 shows that the open circuit voltage is increasing following a logarithmic relationship with the irradiance and decreasing slightly as the cell temperature increases. 2.A. To overcome this undesired effects on the PV power output.2. that oscillate between two levels depending on the actual climatic conditions and the power injected.. A DC link insures an energy balance between the power generated by the PV generator and the power injected into the network by charging or discharging the capacitor. In fact. In case of resistive load. ton the time which the switch is turned on during a complete period T. the short circuit current is linearly depending on the ambient irradiance in direct proportion. (22) and (23) Lasseter et al. DC–DC boost converter The operating point of a PV generator connected to a load is the intersection of the load curve and the PV current–voltage characteristics. 2000: P¼ Q¼ VE sinðdV À dE Þ wL V2 V cosðdV À dE Þ À wL wL ð22Þ ð23Þ where fs is the switching frequency of the converter switch. / Solar Energy 84 (2010) 2219–2229 2223 variations under respectively constant irradiance and temperature. by interfacing the PV generator with the main utility power system. while the open circuit voltage decrease slightly as the cell temperature increases. the configuration of a basic inverter connected to a PV generator is shown in Fig. if the load R is noticeably larger or smaller than Ropt the PV generator operates respectively as constant voltage source or constant current source.. expressed as following: mà ¼ Và m V DC ð21Þ With optimal duty cycle Dà ¼ 1 À mà . PQ inverter The inverter plays a vital role in grid-connected systems. Chaouachi et al. Moreover. the DC–DC boost converter is used to maximize the energy transfer from the photovoltaic generator to an electrical system by adjusting the PV generator output voltage to a reference value (Vref) at which the PV generator supplies the maximum power. a mathematical model describing the DC–DC converter can be expressed in the following relation (Kamel et al. a PV generator connected to a load can operate in a large margin of current and voltage depending on weather conditions (Masters. the PV generator delivers the maximum power (MP). while the power delivered to the load depends on the value of the resistance only. then extra efforts should be done to compensate the undesired effects of the generated harmonics that leads to more heavy computational simulations. The basic and minimum requirement of voltage source inverter is to control the flow of active and reactive power between the PV source and the main utility. if such details are considered.

The three ANNs have similar architecture. 3. the first one is a fuzzy-rule-based classifier while the second one is composed of three multi-layered feed forwarded ANNs (Fig. therefore the power injected at the AC side does not include harmonics. 3. In the present case. CM} to an object belonging to a certain feature space x e SN. two PI controllers would suffice to control the flow of active and reactive power by generating the proper values of voltage (V) and angle (dv) based on the instantaneous value of voltage and current.2224 A. . . 5. .1. Fig. Chaouachi et al. The reference power (Pref) in Fig. composed of three layers: input. / Solar Energy 84 (2010) 2219–2229 Fig. the inverter is operating at unity power factor (Qref = 0) therefore no reactive power is exchanged and the total MP extracted from the PV generator is injected to the grid. Basic structure of the inverter PQ control scheme. 7. while Qref represents the amount of reactive power desired to be injected into or absorbed from the main utility. 7). 6 represents the amount of active power produced by the photovoltaic generator interfaced to the main utility through the inverter. . Architecture of the Neuro–fuzzy network. Neuro-fuzzy maximum power point estimation The neuro-fuzzy network consists of two stages. Fuzzy rules-based classification Classification task consists of assigning a class Cj from a predefined class set C = {C1. the switches are not considered. C2. hidden and output layers. Basic inverter connected to a PV generator. The inverter voltage is represented using the following three controlled sinusoidal voltage sources defined as: pffiffiffi V a ¼ 2V Á sinðwt þ dV Þ ð24Þ   pffiffiffi 2p ð25Þ V b ¼ 2V Á sin wt þ dV À 3   pffiffiffi 2p V c ¼ 2V Á sin wt þ dV þ ð26Þ 3 where the control variables are V and dv The basic structure of PQ inverter controller is shown in Fig. 2000): D : SN ! C A fuzzy rule-based classification system is composed of fuzzy reasoning method (FRM) and a Knowledge Base (KB) consisting in rule base (RB) and data base that Irrd Class 1 Temp ANN1 Irrd. . Fuzzy Logic Classifier Class 2 Irrd Temp ANN2 Reference Voltage Irrd Class 3 Temp ANN3 Fig. so the problem comes to find a mapping defined as (Kuncheva. losses or delays. Temp. 6. 6.

warm and cold. . such methodology aims to improve the generalization capability of a classification system (Kuncheva. The influence of climatic conditions on the model or behavior of a PV generator can be interpreted through linguistic conditional statements that describe the correlation between the climatic conditions and the PV operating behavior. .M Oj ð30Þ This approach can be graphically represented by Fig. .A. Let us consider a rule base R = {R1.2. The output of the fuzzy system is the class associated to the inputs pattern. 3. ptN Þ. . Fuzzy rules. Lk is the linguistic labels to 1 N discretize the continuous domain of the variables. . .2. . . . the FRM determines a label class for admissible patterns as shown in Fig. 8. while the irradiance membership function is sorted into three categories labeled as sunny. / Solar Energy 84 (2010) 2219–2229 2225 describes the semantic of the fuzzy subsets associated to linguistic labels on the if-then part of the rules. Fuzzy reasoning method. Lk . 10a and b shows the temperature and irradiance membership functions in trapezoidal shape. The temperature membership function is sorted into three categories labeled as: hot. .1. 3. . Fuzzy reasoning method A fuzzy reasoning method is an inference procedure that derives conclusions from a set of if-then rules and a pattern. if-then statement based transparent rules. Chaouachi et al. . irradiance). . RL Pattern Class where a1. Indeed. In fact. 10c shows the output membership function in triangular shape. aN then O is C j 1 ð27Þ R1 R2 . 9. . . lk N ðptN ÞÞ L L ð29Þ Finally the classification for the pattern Et is the class Ch such expressed in: Oh ¼ maxj¼1. . . . 1990) are established based on human expertise on the influence of weather conditions on PV systems operating characteristics that are summarized in Table 1. . Multi-layered perceptrons neural network (MLPNN) has been applied successfully to solve some difficult and diverse problems based on a preliminary supervised training with error back propagation algorithm using an error . 2000). Fuzzy rule-based classification system.1. .2. . . Fuzzy controller design The fuzzy rule-based classifier was build and designed under the Matlab Fuzzy Logic Toolbox and lately exported Knowledge Base Rule Base Data Base Pattern Reasoning Method Class Fig.1. the association degree of the pattern with the class Oj can be expressed as following: Oj ¼ maxk2L Rk ðP t Þ k t Fig.1.1.aN is the outstanding selected features for the classification task. Fig. . the training data set was split into three classes each one is used for the training process of its appropriate ANN. While each rule rj in R generates a class Cj related to a pattern P t ¼ ðpt1 . 11 shows the clustering output of the fuzzy classifier as a function of output power.1.. Unlike conventional crisp classification. normal and cloudy.. . 3. the proposed fuzzy classifier presents several patterns with similar power level that are classified into dissimilar classes as they belongs to different weather conditions (temperature.. . similar to those used in fuzzy control (Lee. Rk . O is the class Cj to which the pattern belongs. RL} and rj rules in R. into Simulink environment as a part of the neuro-fuzzy network and the whole simulation system. Fig. for each class Cj. Multi-layred perceptron neural network ð28Þ where R (P ) is the strength of activation (matching rule) associated to the kth rule. Basing on the KB. 8 (Cordon and Jesus.2. . class2 and class3. Membership functions. the RBs are generated based on fuzzy rules with a class in the consequent according to the following structure: Rk if a1 is Lk and . R2.. 9. obtained by applying a t-norm to the degree of satisfaction of the inputs patterns to the clause (a1 is Lk ): 1 Rk ðP t Þ ¼ T ðlk 1 ðpt1 Þ. 3. sorted into three categories: class1. the FRM considers the rule with the highest combination between the matching degree of the pattern with the if-then part and the certainty degree for the classes. . A membership function is a curve that defines how each point in the input or output space is mapped to a membership value between 0 and 1. . 1999). 3.2. Fig.

mean 0 and standard deviation 1 based across channel normalization (Chaouachi et al. Temperature Irradiance Cloudy Cold Warm Hot Class2 Class1 Class1 Normal Class3 Class2 Class1 Sunny Class3 Class3 Class2 Fuzzy classification 0. To evaluate the performance of the proposed neurofuzzy network comparing to a conventional neural network based MPPT method.3 ANN3 11 0.5 0. Basically.4 0. error back learning consists in two pass through the different layers of the network. N the number of training case The target data set was linearly normalized in order to force the network values to be within the range of output activation functions using upper (Ymax) and lower bounds (Ymin) for the values: Zi ¼ Y i À ðY max À Y min Þ=2 ðY max À Y min Þ=2 ð32Þ correction learning rule. 1989). The networks inputs are temperature and irradiance while the output is the optimal reference voltage. The backpropagation momentum algorithm was used for the training process of the developed networks. a single ANN estimator was developed besides the three ANN’s (ANN1. Zi the standardized value corresponding to Yi.2 0. its effect propagates through the network layer by layer to produce an output as actual response. validation where Yi the raw target variable Y for the ith training case. an activity pattern (input vector) is applied to the sensory nodes of the network.6 0. / Solar Energy 84 (2010) 2219–2229 1. In the forward pass. Inputs and output membership functions of the fuzzy classifier.1 0.3 0. Table 1 Fuzzy rules assignment.6 0.4 0.7 Class 1 set (15%) used during the training process for early stopping to avoid overfitting problems and testing set (15%) to evaluate the generalization performance of the developed ANNs.4 0. 10.9 0. xi the raw input variable X in the ith training case. Table 2 Neural networks training parameters.2226 A. Clustering of the input patterns (training data). the networks were trained up to the stopping validation point while their parameters were continuously changed until no significant error performance decrease is observed. ANN1 Hidden neurons Learning rate Momentum 11 0. basing on the following relation: xi À mean S i ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð31Þ PN 2 i¼1 ðxi ÀmeanÞ N À1 where Class 2 Class output 0. ANN2 and ANN3) constituting the neuro-fuzzy learning machine.6 0. as adding a second hidden layer did not improve the training performances. 11.1 0 PN mean ¼ iÀ1 xi N Class 3 2 4 6 8 1 12 14 16 18 20 Power generated [kWh] Fig.2 0.1 0.8 2 (c) 1 0 0. To improve the training convergence performances. Hornik et al.8 1 Irradiance kW/m Fig.0 Cloudy Normal Sunny 1. 2010) was used for the input training set rescaling. Si the standardized value corresponding to xi.0 Class 1 Class 2 Class 3 (a) 0 10 20 30 40 Temperature °C (b) 0 0. During the backward pass synaptic weights are adjusted in accordance to an error correction-rule.2 0. 1999.8 0. the framework of the ANN’s was set out basing on trial and error approach (Table 2). The training data base was recorded during the year 2008 for the training and testing process the original set was split into three subsets: training set (70%). a forward pass and backward pass..1 0. It should be noticed that a single hidden layer was sufficient for the proposed task.2 0. In fact. Chaouachi et al.3 Single ANN 16 0..35 .1 0. The error signal (subtracted from a desired value) is then propagated backward through the network against the direction of the synaptic connections (Haykin.3 ANN2 14 0.0 Cold Warm Hot 1.

Thereby. Moreover. In fact the mean absolute error (MAE). / Solar Energy 84 (2010) 2219–2229 2227 (a) Voltage [V] 310 300 290 280 270 260 06 08 10 12 14 16 18 Voltage reference Voltage estimation 4. One day operation performances: (a) Neuro-fuzzy reference voltage estimation. 12.139 while the single ANN estimator’s error is equal to 2. (c) Power generation comparison using neuro-fuzzy. the modeling time and the prediction error. (d) RMS voltage. 13.05 0. was applied as statistical error test criteria. Chaouachi et al. and P&O algorithm. 12a and b respectively shows the reference voltage estimation by the neuro-fuzzy network and a single ANN comparing to the experimental data for an unsteady weather conditions during a cloudy day. the estimation performances of the neuro-fuzzy network and the single ANN will be evaluated only in term of estimation error defined as the difference between the experimental and the estimated values based on statistical approach. the first two criteria are not considered to be relevant.05 0. single ANN. Tracking of the reference voltage Power generated and injected into the grid (a) Voltage [V] 400 300 200 100 0 PV voltage Ref voltage 0 0.1 Time [s] PV generated current Time [s] RMS Voltage (c) Current [A] 60 (d) Voltage [V] 0 0. the complex nature of the PV generator and its nonlinear behavior versus the atmospheric changes involves the use of a large training data base to cover most of eventual weather conditions that leads to classification degradation and loss in the generalization performance caused by overfitting problem. In the present study.05 0. the performance of the proposed estimator was compared with the experimental algorithm (P&O) and the conventional ANN estimator regarding to the maximum power extracted during a whole day. (c) PV generated current. (b) Power generated and injected into the grid.05 0. Nevertheless. . System operation during a weather condition step: (a) Tracking of the reference voltage.A. Fig. for the total testing process mean absolute error for the neuro-fuzzy estimator is 0. We can see that the proposed machine learning achieved the most accurate estimation.1 400 300 200 100 40 20 0 0 0 0. On the other hand.1 Time [s] Time [s] Fig.1 (b) Power [kW] 15 10 5 PV power Injected power 0 0 0. as the training process is in offline mode. Simulation results Several performance criteria are reported in the ANN literature as: the training time.496. (b) Single ANN reference voltage estimation. the proposed neuro-fuzzy Time [hour] (b) Voltage [V] 310 300 290 280 270 260 06 08 10 12 14 16 18 Voltage estimation Votage reference Time [hour] (c) Power [kW] 16 12 8 4 0 06 08 10 12 14 16 18 P&O Neuro−fuzzy MPPT Single ANN MPPT Time [hour] Fig. In fact.

50 Kg Maximum power V OC o I OCo Vpm 120 W 30.90 V 4. Chaouachi et al. the second one is composed of three multi-layered feed forwarded ANNs trained offline using experimental data from a real PV system installed at the engineering campus of Tokyo University of Agriculture and Technology. the proposed neuro-fuzzy-based method achieved the highest power efficiency with 6. . After reaching the steady-state (0.016 s is due to the dynamic state during the initialization of the system (DC–DC converter. Fig. Fig.85% of extra generated power comparing to the single ANN and 2. .1 Time [ s ] Time [ s ] Fig. the voltage and current are in phase which means that the MP extracted from the PV array can pass into the main grid as the whole system operates at unity power factor (Qref = 0) with no reactive power exchange. 13d). In order to validate the stability performances of the proposed MPPT based neuro-fuzzy estimator. 13a shows the tracking of the reference voltage estimated by the neuro-fuzzy machine learning after a weather condition step change (irradiance 650 ? 500 kW/m2. as the MPPT dispositive was able to track the reference voltage insuring an optimal operating condition under any rapid changing meteoric conditions. Fig. temperature 26. The whole system was simulated under Matlab SimulinkÒ environment.04 0.02 0. 12c presents the simulation output of the PV system (extracted power) during 1 day using the neuro-fuzzy network and the single ANN comparing to the measured experimental data (P&O). PV module specification Manufacturer Type Maximum voltage Weight SHARP NE-LO1A 600 V 12. the neuro-fuzzy network supplies a new value of reference voltage. The peak occurring approximately at 0. 14. Maximum power operation was achieved by tracking the reference voltage estimated by the neuro-fuzzy network through a DC–DC converter.2228 A. The simulation results showed that the proposed system performances was not degraded. tracked through the DC–DC stage so that the overall system can keep operating under optimal conditions. 14 shows the current injected into the main grid and the grid side voltage before and after the weather conditions step change.06 0. / Solar Energy 84 (2010) 2219–2229 Grid voltage 600 Va Vb Vc Injected current 60 40 Ia Ib Ic Voltage [ V ] Current [ A ] 400 200 0 −200 0 0. In this study.73% comparing to the experimental dispositive using the P&O algorithm.5 °C) basing on experimental measures. Grid side voltage and injected current during a weather condition step. In fact.02 s) such peak will not occur even during sudden weather condition changes like the one occurring at 0. The proposed neuro-fuzzy estimator showed the ability to faithfully emulate the dynamic and nonlinear behavior of a photovoltaic generator under a large wide of climatic conditions. The whole gridconnected system performance was tested during a cloudy day with several rapid irradiance variations. machine learning can overcome these limitations through a better definition of the model complexity based on the fuzzy classification. 13b throughout the steady-state next to the climatic disturbance.).04 0.17 A 26.08 0.08 20 0 −20 0. As it can be seen.07 s. Appendix A.70 V . It can be seen that the neurofuzzy MPPT methodology accomplished better performances then the single ANN or the P&O’s algorithm that can fail to track the MPP or oscillates around it under rapidly changing climatic conditions. the developed neuro-fuzzy network consists of two stages. we proceeded to observe the tracking of the MP during the grid connection operation. inverter. Fig.5 ? 24.06 0. 13c shows the variation of the current generated by the PV array after the climatic perturbation basing on the new set point of the system optimal operation meanwhile the RMS voltage keep stable (Fig.02 0. DC bus. the first one is a fuzzy rule-based classifier. In fact the multi-model aspect of the proposed machine learning confer it a distinct generalization ability comparing to a conventional single ANN-based MPP predictor.1 −40 0 0. In fact. 5. that is to say that the power injected into the main utility is equal to the MP Fig. Conclusions In this paper a new MPPT methodology was applied to a grid-connected photovoltaic system based on a proposed neuro-fuzzy estimator.

. 47–62. / Solar Energy 84 (2010) 2219–2229 2229 References Anguera. P. C. Bagging neural networks for predicting water consumption.A. S. R. Solar Energy 83. Ahmad. Nagasaka.. Duran. Neural network ensemble-based solar power generation short-term forecasting. 204–208 (ISSN 10070214 12/23). 1989. 1990.. E. 1370–1378. Man & Cybernetics SMC 20 (2). IEEE Transactions on Energy Conversion 22 (2). 1999. A. Masters.A. E.M.M. 2000. Hohm. ANN based peak power tracking for PV supplied DC motors.. Economy and Environment. Kuncheva... A. 2005. Veerachary. 2009. A. Chaouachi. Hongling..M. White... J. Second ed.C.. Junyi.. Lasseter. Binggang. Chaouachi. X. O.L. A. T. Andujar.. Nagasaka. K.. . C. pp.E. Artificial neural network maximum power point tracker for solar electric vehicle... Wooters. K. Andujar. A. Kamel. 2004.. 2008. 439–449. D.M.H. J. Ming. Progress in Energy and Combustion Science 34. N. P. El Shenawy.. In: Conference on Energy System. Review of the maximum power point tracking algorithms for stand-alone photovoltaic systems. 31–38. 574–632. Kamel. 2005.. Helwa. A. 2003. 2000... Bahgat. Olias.M. F.. Fuzzy Classifier Design.B. 343–350. Enrique. Comparison of photovoltaic array maximum power point tracking techniques..G. J.. Bohorquez. Mellit. Solar Energy Materials & Solar Cells 90. Jesus. Stinchcombe. 2010. Xinghua. Modeling and simulation of a photovoltaic field based on matlab simulin.. 359– 366. 2009. R. 1257–1268. Solar Energy 84. Nagasaka.L.. Tokyo... 4). 2005. Maximum power point tracking controller for pv systems using neural networks. 21–45.E. 2009.. G... Yadaiah. de-Cardona. Lee.. Theoretical assessment of the maximum power point tracking efficiency of photovoltaic facilities with different converter topologies.M. 69–75. Model complexity selection and cross-validation EM training for Robust speaker diarization.. S. M. 1999. Journal of Communication and Computer 2 (3). Kalogirou. A Comprehensive Foundation.. V. 60-64. Lazaro. C.S. M. Solar Energy 81. Springer. 19–24 (Serial No. Haykin. Gao. K. Design and implementation of various inverter controls to interface distributed generators (DGs) in micro grids. Chaouachi. Neural Networks 2 (2). Scenarios for Distributed Technology Applications with Steady State and Dynamic Models of Loads and Micro-Sources.. Heidelberg.M. Z. J. K. In: Proc of ICASSP. 1555– 1578. C. Tokyo. Renewable Energy 30. Enrique. T. S.P. M. R. Z. IEEE Transactions on Systems. Kamel. 2007. Shinozaki. Piagi. Fuzzy logic in control systems.. H. Physica-Verlag.. 2007. pp. F... Neural Networks. CERTS Report. 2000. R. G. M. Robust maximum power point tracking method for photovoltaic cells: a sliding mode control approach.M. Tsinghua Science and Technology 10 (2). K. Solar Energy 69.M. fast and low cost maximum power point tracker for photovoltaic applications. Wiley Interscience. M... OCRAN. Chaouachi et al... 2007. Chapman. T. International Journal of Approximate Reasoning 20. C. Artificial intelligence techniques for photovoltaic applications: a review. L. Economy and Environment.I. E. 2006. Journal of Advanced Computational Intelligence and Intelligent Informatics 14 (1). Barrado.T.C.J. Chua. Esram. In: Conference on Energy System.. A. N. J... 404–435.A.. Progress in Photovoltaics: Research and Applications. Multilayer feedforward networks are universal approximators.. C. Comparative study of maximum power point tracking algorithms.. Hernando... Cordon. Chen. 2010. Tomsovic.. Ropp.A. Salas..H. Hornik. A proposal on reasoning methods in fuzzy rule-based classification systems. 79–89. Renewable and Efficient Electric Power Systems..M. 70–74.. A reliable..

Sign up to vote on this title
UsefulNot useful