Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
KT
_ _
1
_ _
I
S2
exp
qV IR
S
AKT
_ _
1
_ _
V IR
S
R
P
. 1
A solution of the above equation requires a knowledge
of the values for the six parameters, namely I
ph
, I
S1
, I
S2
,
A, R
S
and R
P
, which are variables of the cell types,
irradiation and temperature. A commonly used ap-
proach to identify these parameters involves treating the
ARTICLE IN PRESS
Nomenclature
A ideality factor ( 02)
I
or
solar cell reverse saturation current
( 19.9693 10
7
A)
I
ph
solar cell photon current (A)
I
S1
solar cell saturation current (A)
I
S2
solar cell recombination current (A)
K Boltzman constant ( 1.380658 10
23
)
N
P
number of solar cells connected in parallel
N
S
number of solar cells connected in series
q electron charge 1:062 10
19
R
S
solar cell series resistance (O)
R
P
solar cell shunt resistance (O)
S solar radiation (mw/cm
2
)
T cell temperature (1K)
I
Ph
I
S1
I
S2
I
+
V
Rs
R
P
D
1
D
2
--
Fig. 1. The equivalent circuit of a solar cell.
L. Zhang, Y. Fei Bai / Engineering Applications of Articial Intelligence 18 (2005) 833844 834
cell/panel as a black box. By testing it under different
environmental conditions and taking inputoutput
samples values for the six parameters may be estimated
through applying a curve-tting mechanism. Such a
procedure leads to a set of values suitable only to one
weather condition and cell type, hence is not useful for
general application. In order to develop a model for
general use, it is necessary to dene a set of expressions
with which variations of the parameters with environ-
ment conditions can be described. Gow and Manning
(1999) developed a set of formulae
I
ph
K
0
S1 K
1
T; I
S1
K
2
T
3
exp
K
3
T
_ _
,
I
S2
K
4
T
3=2
exp
K
5
T
_ _
; A K
6
K
7
T,
R
S
K
8
K
9
S
K
10
T; R
P
K
11
expK
12
T,
where constants K
0
K
12
, given in Appendix 1, were
derived using a curve-tting method. They are indepen-
dent of the type of PV cells. Details of their derivation
are discussed by Gow and Manning (1999).
Applying these expressions to represent the corre-
sponding parameters in Eq. (1), we obtain a simulation
model of a real PV cell.
For a PV array with N
S
cells in series and N
P
cells in
parallel, the output current can be expressed as
I
T
N
P
I
ph
I
S1
exp
qV=N
S
I=N
P
R
S
KT
_ _
1
_ _ _
I
S2
exp
qV=N
S
I=N
P
R
S
AKT
_ _
1
_ _
V=N
S
I=N
P
R
S
R
P
_
. 2
The output power can be calculated by multiplying I
T
and V. Eq. (2) can be applied to simulate the I V and
P V characteristics of a PV array. In this study, a small
PV panel is simulated with its details given in Section
3.3. To evaluate the voltage at the maximum output
power point corresponding to a particular weather
condition, it is necessary to nd the voltage value
yielding dP/dV
max
0. This involves using an iterative
algorithm to solve a differential equation which is
implicit.
3. Modelling of PV arrays using RBFNs trained by a GA
3.1. Structure of RBFNs for PV arrays
The RBFNs proposed to represent a PV array are
shown in Fig. 2. This array comprises three layers: the
input, hidden and output layers. The input layer consists
of a three-dimensional vector, X, whose elements are
radiation, ambient temperature and load voltage. The
output layer, Y, has only one element, the load current,
though in general it can be a vector of any dimension.
The hidden layer is composed of L radial basis functions
(RBFs), f
j
j 1; . . . ; L; that are connected directly to
all the elements in the input layer. For a data set
consisting of N
T
input vectors together with the
corresponding output currents there are N
T
such hidden
units, each corresponding to one data point. Thus,
we have
X
n
R
n
ad
T
n
a
V
n
_
_
_
_; Y
n
I
n
n 1; 2; . . . ; N
T
,
the hidden unit can be expressed as a matrix:
U
f
1
1
f
1
2
. . . f
1
L
f
2
1
f
2
2
. . . f
2
L
.
.
.
f
N
T
1
f
N
T
2
. . . f
N
T
L
_
_
_
_
and the weight vector
W
W
1;k
W
2;k
.
.
.
W
L;k
_
_
_
_
k 1.
The function f
j
X
n
takes the form of a non-linear
distribution. The one commonly used is the Gaussian
function of the form
U
j
X
n
exp
kX
n
l
j
k
2
s
2
j
_ _
, (3)
where l
j
j 1; 2; . . . ; L is a vector having the same
dimension as X and representing the centre of the RBF,
U
j
and s is a scalar dening the width of an RBF,
ARTICLE IN PRESS
dd
1
b
1
dd
a
1
dd
b
1
2
a
2
2
b
1
1
a
1
1
b
2
s
b
2
2
b
2
s
w
sdd
w
1.1
a
a
a
y
P
y
2
y
1
X
N
X
2
X
1
Input
layer
layer layer
RBF
hidden
Linear
output
Fig. 2. RBFN network structure.
L. Zhang, Y. Fei Bai / Engineering Applications of Articial Intelligence 18 (2005) 833844 835
sometimes called the spread width. Each element of the
output vector (in the PV array application it is a scalar),
Y
n
, is taken to be a linearly weighted sum of L basis
functions given by
y
n
k
X
n
L
j1
w
jk
U
j
X
n
b
j
, (4)
where y
n
k
X
n
is the kth element of Y
n
(X
n
), w
jk
are the
weighting factors b
j
is the biases of the linear layer,
which compensates for the difference between the
average value over the data set of the RBFs activation
and the corresponding average value of the target
outputs. With the above-described structure the trans-
formation from the input space to the hidden layer is
non-linear due to the use of Gaussian functions for
RBFs. The relationship of the hidden layer to the output
layer, however, is linear.
3.2. GA-based RBFN training scheme
Training an RBF network involves selecting the
appropriate number of basis functions and their
parameters and estimating the weights and biases. The
criterion is to minimise the mean square errors (MSE)
dened as
MSE
1
N
T
N
T
n1
y
n
k
y
n
k
X
n
2
, (5)
where y
k
n
(in this application k 1) are the target values
of the network output when the network is presented
with input vector X
n
.
Since the RBFs have been motivated from the
perspective of function approximation, regularisation,
noisy interpolation, etc. (Abdulhadi et al., 2004; Haykin,
1999) a two-stage procedure has been commonly used
for training RBFNs, which can be substantially faster
than the ones for training multi-layer perceptron
(BPNN). In the rst stage, the number of RBFs and
parameters governing them are determined using
unsupervised methods employing only the input data
ignoring the output data. The second stage involves
estimating the weights from the hidden layer to output
layer which requires the solution of a linear problem.
The proposed new training scheme adopts the two-stage
training procedure. However, in the rst stage, a GA is
employed to search for a single parameter crucial in
determining the structure of the RBFNs, subsequently a
self-growing algorithm is used to progressively increase
the number of RBFs and adjust their positions. The
advantage of this approach lies in its ability to select
RBFNs which is optimal in minimising MSE of Eq. (5)
without requiring the user to pre-select the RBFs
number, which may be inappropriate. Details of the
method are described below.
3.2.1. Selecting the cluster distance factor by a GA
The structure of RBFNs can be obtained by estimat-
ing the optimal cluster distance factor, ; which is
dened as the maximum allowed distance between an
input sample to a prototype RBF centre. Using the
estimated the input data space can be classied into L
units each of which denes an RBF prototype. The
approach to optimal selection involves searching
through a class of RBFNs trained by using different
values, and selecting the one which gives the best
value in terms of specied performance criteria. A GA is
used to perform this selection as it is capable of
searching different areas of the parameter space and
directing the search to regions where the probability of
nding the optimal solution is high. In addition, the
searching mechanism involves manipulating a popula-
tion of binary strings which represent the unknown
parameters and the computational procedure is simple
and hence efcient. The procedure of GA-based
selection is outlined below, while details are given in
Bai (2003).
The GA works with a set of strings, called population.
Each string is a binary coding of a set of parameters
which may be the solution to the problem. The search
for starts from creating an initial population having
randomly generated values. They are then encoded
into binary strings with a bit-length between 8 and 32,
called chromosomes. The tness of each chromosome is
evaluated by applying a non-linear tness function
dened by
f Y
n
1
MSEY
n
1
1
1=N
T
N
T
n1
y
n
k
y
n
k
X
n
2
_ _
1
6
The optimal solution should give the minimum MSE
between the sampled data output y
n
k
and the predicted
output by the RBFNs,
y
n
k
X
n
; namely, the maximum
f(Y
n
). Thus, those with higher tness values will be
chosen to be the parents of the next generation
whereas those with lower tness will be rejected. To
create the new offspring the selected parent
strings undergo a reproduction process which involves
the application of genetic operators such as cross-
over and mutation. Details of these are discussed in
Bai (2003). The new generation thus generated has
the same population size as its predecessor and is
considered tter. Their tness values can be evaluated
through applying the tness function of Eq. (6), and
the above-described selection process is then repeated.
Continuing such a procedure the newer and tter
chromosomes evolve until a predened stopping
condition, either in terms of reaching the maximum
tness value or the maximum number of iterations, is
satised.
ARTICLE IN PRESS
L. Zhang, Y. Fei Bai / Engineering Applications of Articial Intelligence 18 (2005) 833844 836
3.2.2. The self-growing RBFN training algorithm
The appropriate value obtained can be used as the
highest possible resolution to partition the input data
space progressively into L different feature spaces
centred at l
i
i 1; 2; . . . ; L: In detail, learning begins
by determining the rst RBF centre, l
1
: This is
accomplished by setting l
1
equal to an input vector,
x
1
, which is chosen randomly from the input training
data set and evaluating the Euclidean norm between l
1
and the next input data x
2
, namely kl
1
x
2
k
2
: The
result is compared with : If it is smaller, l
1
is established
as the centre of an RBF prototype and its elements are
updated using a formula expressed as
l
1i
new l
1i
old akx
2i
l
1i
oldk
i 1; 2; . . . ; N, 7
where m
1i
and x
2i
represents the ith components of
vectors l
1
and x
2
; respectively, and 0oao1 is the
updating ratio, which is usually chosen as 0.5. On the
other hand, if the Euclidean norm is greater than a new
prototype, l
2
; locating at x
2
is created. Applying this
procedure to process the next sample x
3
; the Euclidean
norm values for l
1
and x
3
; plus l
2
and x
3
are calculated,
respectively. Their results are compared and the pair
giving the smaller value is selected. Comparing this
selected Euclidean norm value with ; if it is larger, a
new prototype, l
3
x
3
; is generated otherwise the
components of l
1
=l
2
are rescaled to the value calculated
using Eq. (7). Carrying on in this manner to process the
remaining input samples, the number of clusters grows
and their centres self-adjust continuously until all the
samples are processed. Thus the number and locations
of RBFs in the hidden layer are determined.
3.2.3. The supervised-learning scheme for the weights
Following determination of the RBFs in the hidden
layer, a supervised learning scheme is employed to train
the weights between the hidden and the output layers.
The aim is for the RBFN to give the desired responses
using only the output of the training data when
presented with the output of the RBFs generated by
the available prototypes. A straightforward procedure
for accomplishing this is to use the pseudo-inverse
method. In this case, the formal solution of the weights
in matrix form is given by
W
U
T
U
1
U
T
Y
. (8)
In the proposed scheme, an iterative procedure is
applied to update the weights. Thus, for the kth
iteration, the jth element, w
j
(k), of the weight vector
W(k) can be updated on the basis of the (k1)th value,
w
j
(k1), and is given as
w
j
k w
j
k 1
ekU
j
k
L
i1
U
2
j
k
j 1; 2; . . . ; L. (9)
This algorithm can be proved to converge to a
unique solution, W
: It is computationally efcient
and can give good learning and generalisation of
performance.
The schematics of the above training scheme are
illustrated by the owchart shown in Figs. 3a and b.
3.3. An RBFN network for a PV panel
To derive an RBFN for a PV array, the equations
given in Section 2 are implemented to represent a
small real PV panel. The panel is composed of two
cell branches in series and each branch has 20 cells in
parallel. Under the standard weather condition
(100 mW/cm
2
and 25 1C), the current and voltage values
at the peak power point are 15.791 A and 0.393 V,
respectively. So each cell is capable of generating the
maximum power of 6.206 W. The open circuit voltage
of the cell is 0.775 V, and short circuit current is
30.496 A. A total of 60 different weather conditions
are generated randomly with radiation varying in
the range 10100 mW/cm
2
and temperature 040 1C.
For each weather condition, 50 IV samples
are calculated; thus, 3000 data are collected for train-
ing and testing the RBFN. The model trained is
tested by applying eight different weather conditions
data measured in Riyadh Solar Village, Saudi Arabia,
in 1998 as listed in Table 1. In applying the GA-RBFN
training scheme the search range of the cluster distance
factor, ; is set between 5 and 50, and the GA para-
meters are established as given in Appendix 2. The
optimal found by the GA is 30.664063, giving 29 RBF
units. To establish the qualitative judgement of the
method the GA-trained RBFN is compared with a
different RBFN trained by the conventional k-means
algorithm. Fig. 4 shows the IV and PV curves
constructed by both the GA-RBFN and the conven-
tional RBFN for two sets of radiation and temperature
values. It can be observed that both the IV and PV
curves produced by the GA-RBFN t well with the
curves of the real panel. Thus, the GA trained RBFN
model is signicantly better than the RBFN using k-
means algorithm.
It is worthy of note that training an RBFN model
requires a large amount of data measured either on-line
or off-line. The results shown in Fig. 4 were obtained by
off-line implementation of the GA-RBFN method. For
real-time operation, the training scheme may run
periodically when sufcient samples are measured from
the PV panel. Investigation in this aspect has been
carried out and for discussions one can refer to Bai
(2003). Using the trained RBFNs the MPPs for each of
the eight weather are obtained, which are listed in
Table 1. As is shown clearly, the GA-trained RBFN
gives, on average, more accurate predictions of MPPs
than its counter model.
ARTICLE IN PRESS
L. Zhang, Y. Fei Bai / Engineering Applications of Articial Intelligence 18 (2005) 833844 837
4. MPP modelling and prediction using RBFNs trained
by a GA
4.1. MPP modelling
The RBFN can also be applied to directly construct
an MPP model for a PV panel. This is accomplished
by having a two-dimensional vector consisting of
the radiation and temperature as the input layer.
The corresponding output vector is comprised of
samples of current (I
max
) and voltage (V
max
) at the
MPPs. The hidden layer again consists of a group
of RBF units as shown in Fig. 2. The network can be
trained using the above described GA-self-growing
ARTICLE IN PRESS
END
Producing new by GA evolution
(reproduction, crossove, mutation)
N
?
Stop evolution
Y
GA fitness evaluation by
f(y)=
1 1
=
MSE(y)+1 1
N
r
N
r
A-1
y
n
(k) - y
n
(k)
+1
i
2
(k)
i-1
( j = 1,2
...
, L)
Constructing RBFN structure
by Rbf-growing alorithm
Initilising randomly
Determining the first RBF center
1
by
randomly selecting one date sample x
1
(n = 1, p = 1)
For the next n-th sample x
n
, finding the
closest RBF center
i
by
min( x
n
1
2
, ..., x
n
P
2
) || ||
|| ||
|| ||
|| ||
x
n
i
> E
?
Y
Y
N
N
P = P+1,
P
= x
n
P
(new) =
P
(old) + x
n
P
(old)
Last sample data ?
END
(a) (b)
L
Fig. 3. Flowcharts of GA-RBFN training algorithm. (a) RBF self-growing; (b) combining GA for RBFN training.
Table 1
MPP prediction results by GA-RBFN and RBFN
Time (h) Radiation (mW/cm
2
) Temperature (1C) P
max
(W) Prediction results by GA-RBFN Prediction results by RBFN
P
max
Error P
max
Error
89 55.267 29.170 5.142 4.838 5.91% 4.843 5.81%
910 76.213 31.394 6.405 6.368 0.58% 6.350 0.85%
1011 90.950 32.930 7.257 7.288 0.44% 7.372 1.60%
1112 96.900 34.967 7.926 7.910 0.21% 8.066 1.76%
1213 95.703 35.884 8.101 8.088 0.16% 8.250 1.85%
1314 86.797 36.524 7.849 7.922 0.94% 8.034 2.36%
1415 70.299 37.176 7.065 6.946 1.68% 6.975 1.28%
516 49.329 37.002 5.518 4.974 9.86% 4.989 9.60%
L. Zhang, Y. Fei Bai / Engineering Applications of Articial Intelligence 18 (2005) 833844 838
algorithm when provided with measured input and
output data.
4.2. An MPP-RBFN for a PV panel in Riyadh, Saudi
Arabia
An RBFN was trained to predict the MPPs for a grid-
connected PV panel in Riyadh, Saudi Arabia. The
system consists of two branches each rated at 3 kW. A
commercial MPP tracker is employed to locate the
MPPs in response to changes of environmental condi-
tions. The method used is the P&O scheme. In this
present work, the set of data collected in 1998 from only
one branch of this PV system was used to train the
RBFN. Samples of radiation, temperature, output
current and voltage are taken once per ve minutes
from sunrise at 7:40 to sunset at 17:00 for a day. Their
average values per hour are used for constructing the
RBFN for MPP modelling. In this way a set of 510 data
extracted from the total was used for training and the
other 110 data for model testing. Figs. 5a and b show
clouds of training and testing data with the range of
isolation level varying from 10 to 100 mW/cm
2
. At the
radiation levels below 10 mW/cm
2
, the power output
from the PV panel became insignicantly low, and hence
have not been used for training. Subsequently, the
RBFN model derived would not be suitable for weather
conditions with very low solar insolation due to the
highly non-linear feature of the PV cells.
To apply the GA-based training method the GA
parameters are set in Appendix 2. The evolution is
carried out in 10 generations and the optimal value for
is 13.5430. The number of RBF units subsequently
derived is 26.
The RBFN thus trained using the GA-based method
was then applied to predict the PV panel output power
giving the input data from the testing data set. The
predicted power values are plotted in the graph shown in
Fig. 6b. To compare the performance of this RBFN
a different RBFN, trained by using the conventional
k-means method with the same set of inputoutput
training data, is constructed. The number of RBF
units is set as 118 after a number of test runs. The
result derived from the latter, shown in Fig. 6a, is
contrasted with that of the GA-trained network. It can
be seen clearly that the GA-trained RBFN ts
signicantly better the measured data from the real PV
plant than the RBFN derived using the k-means training
scheme. The degree of accuracy may be measured by
using the percentage mean relative errors (PMRE)
expressed as
PMRE
1
N
T
N
T
k1
jekj
yk
100%
_ _
,
where N
T
is the number of data samples, y(k) is the kth
measured output, and jekj represents the absolute
value of the difference between the kth sampled and
predicted power values. Thus the PMRE for GA-trained
RBFN is 3.6995%, whereas that for the k-means scheme
is 5.4204%, which is higher. The results are also shown
in Fig. 6a.
ARTICLE IN PRESS
Fig. 4. IV and PV characteristics reconstruction by (a) RBFN (b) GA-RBFN (solid linereal data, dashed linepredicted data).
L. Zhang, Y. Fei Bai / Engineering Applications of Articial Intelligence 18 (2005) 833844 839
4.3. An MPP-RBFN for a PV array in Southampton,
UK
The above-described training scheme was also applied
to construct an MPP-RBFN for one of the PV arrays
installed in Southampton University (the Southampton
Test and Reference Facility), UK. This facility was
constructed with the support from the UK Department
of Trade and Industrys New and Renewable Energy
programme. There are three solar arrays, and they are
connected to the grid through commercial power
conditioners which are tted with maximum power
ARTICLE IN PRESS
Fig. 5. Cloud of sampled data from Riyadh Solar Village for GA-RBFN (a) training data and (b) testing data.
Fig. 6. MPP prediction results using Riyadh data by (a) RBFN and (b) GA-RBFN (solid linereal data, dashed linepredicted data).
L. Zhang, Y. Fei Bai / Engineering Applications of Articial Intelligence 18 (2005) 833844 840
point tracking facilities. For this present work, data
from Array2 are employed.
The set of data used for training was collected
during the period from 1 to 10 April 2000. They
are samples of radiation, temperature and Array2s
current and voltage. The data were measured at one
per 2 s and their average over a 5-min interval was
logged. Therefore, the amount of data within 1 day
(24 h) is 288 samples, and for 5 days there are 1440
samples in total. Half of these data (day 1 to day 5)
is used to train the model, the remaining half (day 6
to day 10) for testing. With GA-based training
scheme, the GA parameters chosen are as given in
Appendix 2.
ARTICLE IN PRESS
Fig. 7. MPP prediction results using STaR data (on April 7th) (solid linereal data, dashed linepredicted data). (a) Prediction by RBFN-based
model and (b) prediction by GA-RBFN-based model.
L. Zhang, Y. Fei Bai / Engineering Applications of Articial Intelligence 18 (2005) 833844 841
The value derived after 10 evolution cycles is 7.2578
and the corresponding number of RBFs is 18. The
trained RBFN is then applied to predict MPPs for
Array2, Fig. 7 shows the comparison between the
measured and RBFN predicted voltage, current and
power over a 1-day period. Excellent matches have been
obtained for all three variables. In contrast, the result
from another RBFN trained by using the k-means
method shown in Fig. 7a is not as accurate as the GA-
RBFN (shown in Fig. 7b). The PMRE for both models
are evaluated and results are listed in Table 2. Clearly,
the PMREs for the one trained by the GA-RBFN are
consistently and signicantly lower than those given by
its counter part.
To compare further the performance of achieving
MPPT by using RBFN model, Figs. 8a and b show the
currentradiation characteristic curves of the STaR data
using conventional P&O method for MPPT and that
predicted by the GA-RBFN, respectively. As can be
seen the former depicts a persistently oscillatory
relationship between the current and isolation while
the later demonstrates a steady linear feature. The
signicant difference between the two implies that the
P&O method cannot always track the PV maximum
output power points for different radiation levels. In
contrast, the RBFN method would predict the MPP
with good precision and speed for a practical PV
power generator irrespective of the weather condition
variations.
5. Conclusion
This paper has presented an RBFN-based approach
to model the IV characteristics and maximum power
variations of PV panels. The main advantage of this
model is that it does not require a priori knowledge of
the PV panels, and the models derived are sufciently
accurate for real-time MPPT applications. The method
was veried using data collected from two real PV
panels. The trained MPP model for the panel in Riyadh
Solar Village gives a PMRE as low as 3.6995% and the
other in Southampton Test and Reference Facility gives
ARTICLE IN PRESS
Table 2
Solar array MPP prediction results by using STaR sample data
Date P
max
error by RBFN P
max
error by GA-RBFN
V
max
9.54% 8.96%
I
max
7.70% 5.71%
P
max
5.37% 4.18%
Fig. 8. I
max
radiation relation. (a) STaR data (on April 7th), (b) data predicted by GA-RBFN.
L. Zhang, Y. Fei Bai / Engineering Applications of Articial Intelligence 18 (2005) 833844 842
ARTICLE IN PRESS
a PMRE of 4.18%. Although in this work the training
of RBFNs were conducted ofine, for applying to a real
PV power generation system, it is feasible to employ the
scheme online to derive the corresponding MPP-RBFN.
This model can then be used to predict the MPPs in real-
time for the changing weather conditions.
Acknowledgement
The authors would like to thank Dr. Tom Markvart
at Southampton University for providing data for one
of the PV arrays in the Southampton Test and Reference
Facility, UK.
Appendix 1. Values for constants K
0
to K
12
K
0
5:729 10
7
; K
1
0:1098,
K
2
44:5355; K
3
1:264 10
4
,
K
4
11:8003; K
5
7:3174 10
3
,
K
6
2; K
7
0; K
8
1:07,
K
9
1:6126 10
3
,
K
10
3:27 10
3
; K
11
2:303 10
6
,
K
12
2:812 10
2
.
Appendix 2. GA parameters for RBFN training
Probability of
mutation
Probability of
crossover
Chromosome
length
Population size Maximum
generation
number
For simulated PV panel 0.01 0.8 8-bit 30 10
For Riyadh PV panel 0.1 0.6 8-bit 20 10
For Southampton PV
panel
0.01 0.8 8-bit 20 10
References
Abdin, E.S., Osheiba, A.M., Khater, M.M., 1999. Modeling and
optimal controllers design for a stand-alone photovoltaic-diesel
generating unit. IEEE Transactions on Energy Conversion 14 (3),
560565.
Abdulhadi, M., Al-Ibrahim, A.M., Virk, G.S., 2004. Neuro-fuzzy
based solar cell models. IEEE Transactions on Energy Conversion
19 (3), 619624.
Al-Amoudi, A., Zhang, L., 1998. Optimal control of a grid-connected
PV system for maximum power point tracking and utility power
factor. IEE PEVD98, 2123 September 1998, pp. 325330.
Yunfei Bai, 2003. Internet-based modelling and control for grid-
connected photovoltaic power generation systems using articial
neural networks. Ph.D. Thesis, University of Leeds, UK, June
2003.
Bodur, M., Ermis, M., 1994. Maximum power point tracking for low
power photovoltaic solar panels. Seventh Mediterranean Electri-
calTechnical Conference Proceedings, vol. 2, pp. 758761.
Brambilla, A., Gambarara, M., Garatti, A., Ronchi, F., 1999. New
approach to photovoltaic arrays maximum power point tracking.
Power Electronics Specialists Conference, 1999. PESC99. Thirtieth
Annual IEEE, vol. 2, pp. 632637.
In-Su Cha, Jang-Gyun Choi, Gyun-Jong Yu, Myung-woong Jung,
Hyung-Lae Baek, Dong-Hyee Kim, 1997. MPPT for temperature
compensation of photovoltaic system with neural networks.
Photovoltaic Specialists Conference, 1997, Conference Record of
the 26th IEEE, pp. 13211324.
El-Shibini, M.A., Rakha, H.H., 1989. Maximum power point tracking
technique. Proceedings of Integrating Research, Industry and
Education in Energy and Communication Engineering, MELE-
CON 89, Mediterranean, pp. 2124.
Giraud, F., Salameh, Z.M., 1999. Analysis of the effects of a passing
cloud on a grid-interactive photovoltaic system with battery
storage using neural networks. IEEE Transactions on Energy
Conversion 14 (4), 15721577.
Gow, J.A., Manning, C.D., 1999. Development of a photovoltaic
array model for use in power electronics simulation study.
IEEE Proceedings on Electrical Power Application 146 (2),
193200.
Gow, J.A., Manning, CD., 2000. Controller arrangement for boost
converter systems sourced from solar photovoltaic arrays or other
maximum power sources. Electric Power Applications, IEE
Proceedings 147 (1), 1520.
Hamdy, M., 1994. A new model for the IV output characteristics of
photovoltaic modules. Journal of Power Source 50, 1120.
Hassoun, M.H., 1995. Fundamentals of Articial Neural Networks.
The MIT Press, Cambridge, MA, London, England.
Haykin, S., 1999. Neural Networksa Comprehensive Foundation,
second edition. Prentice Hall, New Jersey, US.
Takashi Hiyama, Ken Kitabayashi, 1997. Neural network based
estimation of maximum power generation from PV module using
environmental information. IEEE Transactions on Energy Con-
version 12 (3), 241247.
Takashi Hiyama, Shinichi Kouzuma, Tomofumi Imakubo, 1995.
Identication of optimal operating point of PV modules using
neural network for real time maximum power tracking control.
IEEE Transactions on Energy Conversion 10 (2), 360366.
Chihchiang Hua, Chihming Shen, 1998. Comparative study of peak
power tracking techniques for solar storage system. Applied Power
L. Zhang, Y. Fei Bai / Engineering Applications of Articial Intelligence 18 (2005) 833844 843
Electronics Conference and Exposition, APEC 98, 13th Annual
Conference Proceedings, vol. 2, pp. 679685.
Chihchiang Hua, Jongrong Lin, Chihming Shen, 1998. Implementation
of a DSP-controlled photovoltaic system with peak power tracking.
Industrial IEEE Transactions on Electronics 45 (1), 99107.
Hussein, K.H., Muta, I., Hoshino, T., Osakada, M., 1995. Maximum
photovoltaic power tracking: an algorithm for rapidly changing
atmospheric condition. IEE ProceedingsGeneral Transmission
Distribution 142 (1), 5964.
Matsui, M., Kitano, T., De-hong Xu, Zhong-qing Yang, 1999. A new
maximum photovoltaic power tracking control scheme based on
power equilibrium at DC link. Industry Applications Conference,
1999. Thirty-Fourth IAS Annual Meeting. Conference Record of
the 1999 IEEE, vol. 2, pp. 804809.
Merten, J., Asensi, J.M., et al., 1998. Improved equivalent circuit
and analytical model for amorphous silicon solar cells and
modules. IEEE Transactions on Electron Devices 45 (2),
423429.
Merwe, L., Merwe, G., 1998. Maximum power point tracking
implementation strategies. Proceedings. ISIE 98. IEEE Interna-
tional Symposium, vol. 1, pp. 214217.
Takashima, T., Tanaka, T., Amano, M., Ando, Y., 2000. Maximum
output control of photovoltaic (PV) array. Energy Conversion
Engineering Conference and Exhibit, 2000 (IECEC) 35th Inter-
society, vol. 1, pp. 380383.
Vanhtesevanos, G., Kalaitzakis, K., 1987. A hybrid photovoltaic
simulator for utility interactive studies. IEEE Transactions on
Energy Conversion 2, 352357.
Wong, T.K.P., Chan, P.C.H., 1995. An equivalent circuit approach to
solar cell modeling. IEEE Tencon 95 on Microelectronics and
VLSI, November 1995, pp. 1217.
ARTICLE IN PRESS
L. Zhang, Y. Fei Bai / Engineering Applications of Articial Intelligence 18 (2005) 833844 844