ARTICLE IN PRESS

Energy 32 (2007) 1513–1523

www.elsevier.com/locate/energy

Prediction of hourly and daily diffuse fraction using neural network,

as compared to linear regression models
Hamdy K. Elminir, Yosry A. Azzam, Farag I. Younes
Department of Solar and Space Research, National Research Institute of Astronomy and Geophysics,
El-Marsad Street, P.O. Box 11421 Helwan, Cairo, Egypt
Received 15 July 2006

Abstract

For most of the locations all over Egypt the records of diffuse radiation in whatever scale are non-existent. In case that it exists, the
quality of these records is not as good as it should be for most purposes and so an estimate of its values is desirable. To achieve such a
task, an artificial neural network (ANN) model has been proposed to predict diffuse fraction (KD) in hourly and daily scale.
A comparison between the performances of the ANN model with that of two linear regression models has been reported. An attempt was
also done to describe the ANN outputs in terms of first order polynomials relating KD with clearness index (KT) and sunshine fraction
(S/S0). If care is taken in considering the corresponding regional climatic differences, these correlations can be generalized and
transferred to other sites. The results hint that the ANN model is more suitable to predict diffuse fraction in hourly and daily scales than
the regression models in the plain areas of Egypt.
r 2006 Elsevier Ltd. All rights reserved.

Keywords: Diffuse fraction; Linear regression model; Neural network; Back-propagation algorithm

1. Introduction and objectives
Solar energy is being seriously considered for satisfying
part of the energy demand in Egypt [1,2]. In this respect,
the importance of solar radiation data for the design and
efficient operation of solar energy systems has been
recognized [3–8]. The increasing demand for information
on the availability of solar radiation has highlighted the
inadequacy of the current observational networks. This
inadequacy occurs in three basic ways: limited spatial
coverage; limited length of record; and predominance of
global radiation data, while few data are available for the
diffuse component. Given global solar radiation measure-
ments, the diffuse component, essential for most energy
applications, can be obtained through various theoretical
or empirical correlations. These correlations are usually
expressed in terms of linear and/or polynomial fittings
relating the diffuse fraction with, (i) the clearness index
[9–13] and/or (ii) combining both clearness index and
sunshine fraction (S/S0) [14–16]. One of the limitations
imposed by these correlations is that they show a latitude
dependency [17,18]. As a result, Hamdy [19] checked the
adequacy of eight widely used linear regression models for
the prediction of diffuse fraction in different scales. An
integral view of this study shows that
1. For the observed range of errors, it was determined that
Gopinathan and Soler’s model [15] produces the lower
error, while Collares-Pereira and Rabl’s [13], and Ruth
and Chant’s [5] models follow in that order. As a result,
Gopinathan and Soler’s model can be used with
confidence to calculate the diffuse fraction in daily
scale for Egypt. The analytical form of Gopinathan and
Soler’s model [15] is given by
K D ¼ 0:7345 þ 0:0219:K T  0:5665:S=S 0 . (1)

Corresponding author. Tel.: +20 25560645; fax: +20 25548020. 2. In hourly scale it is recommended that the correlations
E-mail address: hamdy_elminir@hotmail.com (H.K. Elminir). developed by Erbs et al. [14] be used when no recorded

0360-5442/$ - see front matter r 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.energy.2006.10.010
 ARTICLE IN PRESS
1514 H.K. Elminir et al. / Energy 32 (2007) 1513–1523

Nomenclature S the sunshine duration (hour)

SO the maximum possible number of day light
DH the hourly value of the diffuse solar radiation (hour duration)
incident upon a horizontal surface (MJ/m2) tj the target output of the node j
Dn the day of the year, e.g. 1. Jan ¼ 1, 2. Jan ¼ 2, wji(t) the connection weight between neuron j and
etc. neuron i at iteration t
EO eccentricity correction factor of the Earth’s x the net input of the weighted values of the
orbit individual inputs (p1, p2,y, pj)
f(x) a non-linear sigmoid transfer-function yj the output of the node j
GH the hourly value of the global solar radiation
incident upon a horizontal surface (MJ/m2) Greek symbols
GOD the daily value of the extraterrestrial irradiation
(MJ/m2/day) G day angle in radians (shown in Eq. (16))
GOH the hourly value of the extraterrestrial irradia- a the momentum factor
tion (MJ/m2) Z the gain factor, which control the rate of
ISC solar constant, it is taken to be 1367 W/m2 convergence
KD diffuse fraction (DH/GH) d solar declination (degrees)
KT clearness index (GH/GOH) dk the error term of the output layer
PM10 the particulate matter of diameter less than f the latitude of the location (degrees)
10 mm oS hour angle at sunset (degrees)

data are available. These correlations are given as significant effect on the quality of the data sets together
8 with the shade ring corrections. The training data sets have
> 1:0  0:09:K T ; 0pK T p0:22;
>
> been used to train the neural network according to
< 0:9511  0:1604:K T þ 4:388:K 2
T different structures as follows:
KD ¼
>
> 16:638:K 3
T þ 12:336:K 4
T ; 0:22oK T p0:8
>
:
0:165; K T 40:8: 1. For hourly scale, the neural was trained with patterns
(2) from 1999 to 2001 obtained at Aswan (231580 N,
321470 E, 192 m) in the south semi-desert zone, and
verified using the year of 2002. To further evaluate the
Contrary to the mathematical routines mentioned above
prediction capability, the developed model was tested
artificial novel network (ANN) models are able to learn the
using the patterns obtained during 2003 at the rural area
key information patterns within a multi-dimensional
of South-Valley (261120 N, 321450 E, 96 m) and Cairo
information domain. In addition, they are particularly
(301050 N, 321170 E, 36 m) in the middle plane area of
suitable to solve the problem of identification in the
Egypt. A brief description of the measuring instruments
presence of noisy data [20,21], and some promising results
and the climate of the selected sites are given in [19].
have been reported [22–27]. For those reasons, the problem
Since these locations were not included as a part of the
we address here is to compare the performance of
training set, the high accuracy obtained demonstrated
Gopinathan and Soler [15] and Erbs et al. [14] with that
the ability of the neural networks model to generalize
of the ANN model, which can be based on KT and S/S0.
and to produce accurate estimates.
2. For daily scale, not all operating stations were used. The
2. Data preprocessing main criterion for the selection of the stations was the
availability of the sunshine duration measurements
In any model development process, familiarity with the regarding the period covered. Bearing in mind the
available data is of the utmost importance. The quality of above, the neural was trained with the patterns of
the hourly measurements was ensured by performing Aswan from three years long (1999–2001) and verified
several criteria in order to identify wrong data. First, days using the data obtained during 2002.
that were mostly overcast were rejected primarily because it
is difficult to verify pyranometer data on such days. Next,
hourly data known to be wrong have been omitted. Finally, 3. Mathematical basis of neural network approach
hourly data that violated physical limits, including diffuse
fraction greater than 1.0, were excluded. It is note worthy We believe that an understanding of the neural network
here that due to cosine response problems the present is essential to the satisfactory and successful use of this
analysis is limited to hourly values with solar zenith angles model. So we present here a brief review of the ANN model
less than 801. Applications of these criteria do have a utilized in this study.
 ARTICLE IN PRESS
H.K. Elminir et al. / Energy 32 (2007) 1513–1523
3.1. Description of the ANN model
The neural network may be used to solve problems
which require an approach similar to human reasoning.
The fundamental processing element of it is a neuron,
which can process a local memory and can carry out
localized information. Each neuron computes a weighted
sum of the inputs it receives and adding it with a bias (b) to
form the net input (x). The bias is included in the neurons
to allow the activation function to be offset from zero,
x ¼ w1;1 :p1 þ w1;2 :p2 þ    þ w1;j :pj þ b. (3)
The net input (x) is then passed to the subsequent layer
through a non-linear sigmoid function to form its own
output (yj),
yj ¼ 1=ð1 þ ex Þ. (4)
Afterward, the output yj was compared with the target
output tj using an error function of the form
dk ¼ ðtj  yj Þyj ð1  yj Þ. (5)
For the neuron in the hidden layer, the error term is given
by the following equation:
X 3.2. ANN’s architecture proposed to estimate diffuse
dj ¼ yj ð1  yj Þ dk wk , (6)
k
where dk is the error term of the output layer, and wk is the
weight between the hidden layer and output layer. The
error was then propagated backward from the output layer
to the input layer to update the weight of each connection
as follows:
wji ðt þ 1Þ ¼ wji ðtÞ þ Zdj yj þ aðwji ðtÞ  wji ðt  1ÞÞ. (7)
Fig. 1. ANN’s architecture proposed to estimate diffuse fraction.
1515
Choosing a small learning rate Z leads to slow rate of
convergence, and too large Z leads to oscillation. The term
a is called momentum factor and determines the effect of
past weight changes on the current direction of movement.
Both of these constant terms are specified at the start of the
training cycle and determine the speed and stability of the
network [28]. The process was repeated for each input
pattern until the error was reduced to a threshold value.
The final frozen weights were then used during the testing
session. For a more detailed description of the ANN
model, the reader is referred to [29–31].
Two or more of the neurons shown earlier can be
combined in a layer, and a particular network could
contain one or more such layers. In the recurrent neural
network, multiple neurons are inter-connected to organize
the network. On the other hand, the feed forward network
that is being used in this study has a hierarchical structure
that consists of a number of layers without interconnection
between neurons in each layer, and the signals can only
propagate from the input layer to the output layer through
the hidden layer (see Fig. 1).
fraction
The procedure used here is similar to the one shown in
[27]. In this work, a perceptron neural network technique is
applied to estimate hourly values of the diffuse fraction,
using as input the global solar radiation and other
meteorological parameters, like long-wave atmospheric
emission, air temperature, relative humidity and atmo-
spheric pressure. The neural network was developed based
 ARTICLE IN PRESS
1516 H.K. Elminir et al. / Energy 32 (2007) 1513–1523

on both feature determination and pattern selection et al. [14] model. The generalization of the model was
techniques. Our procedure has an advantage, as it can tested by the correlation coefficient (r). The entries in Fig. 2
predict the KD values in hourly and daily scales based on indicate that the average value of the r was 0.96 and 0.87
KT and S/S0. On the contrary, the procedure of Soares for ANN and Erbs et al. [14] models, respectively. More
et al. [27] needs long-wave atmospheric emission parameter specifically, 96% of the variation in the diffuse fraction can
which is not always available. be explained by the input parameters using ANN
After trying a number of different configurations of technique. Inspecting further the results given in Fig. 2, it
hidden neurons, it was found that 40 neurons in the hidden is obvious that the average value of the standard deviation
layer yield the best results. The output layer has one node (SD) for the results estimated by ANN model does not
for the diffuse fraction. In hourly scale, the input layer has exceed 0.07 with a minimum value up to 0.03. In general,
five individual inputs which are the normalized values of large SD values imply significant errors in estimating the
the month of the year, the day of the month, the hour of diffuse fraction. The frequency distribution of the differ-
the day, the hourly value of the global radiation, and the ence between observed data and those estimated by the
hourly value of the extraterrestrial irradiation. In daily ANN model shows that 99.7% of all hourly values were
scale, the input layer has three inputs which are the daily estimated with a deviation of less than 0.2 MJ/m2. These
value of the global radiation, the daily value of the results were partly mimicked by Erbs et al. model [14],
extraterrestrial irradiation, and the sunshine fraction when 71.7% of the values were estimated with a deviation
(S/S0). After several adjustments to the network parameters, of less than 0.2 MJ/m2 and over 90% of these deviations
the network converged to a threshold of 0.0004 for Z ¼ 0:01 lay within 0.4 MJ/m2. These results indicate that the ANN
and a ¼ 0:5. These values for Z and a were found to based estimation technique for diffuse fraction is more
accelerate the convergence of the back-propagation learning suitable to predict the diffuse fraction in hourly scale than
algorithm without overtraining. It is important to note that the regression model proposed by Erbs et al. [14].
the test session took only 2 s to get the output when Consequently, it is reasonable to describe the ANN
introduced by the final weights of the training session. outputs in terms of first order polynomials relating KD
with KT and S/S0 using multi-regression techniques
4. Results of the diffuse fraction forecasting provided by the widely used MINITAB program
version 14.
4.1. Performance of the models used in hourly scale Cairo Station:
8
> 0:793  2:198:K T ; 0pK T p0:22;
For hourly scale, the performance of the candidate >
>
< 1:341  9:566:K T þ 32:2:K T2
models is compared in the left panel of Fig. 2. The
KD ¼
frequency distributions of the difference between predicted >
> 47:909:K 3T þ 25:419:K 4T ; 0:22oK T p0:8;
>
:
and measured values were also introduced in the right 0:131; K T 40:8:
panel. This figure shows the scatter plot diagrams between
the observed values and those predicted through both of (8)
the candidate models of Erbs et al. [14] and the ANN Aswan station:
technique. Fig. 2 can provide answers to the following 8
> 0:653  1:728:K T ; 0pK T p0:22;
questions: >
>
< 0:724  1:821:K T þ 8:221:K 2
T
KD ¼
(i) Are the observed values and those predicted through >
> 16:370:K 3T þ 9:845:K 4T ; 0:22oK T p0:8;
>
:
the candidate models linearly related? 0:217; K T 40:8:
(ii) Does the variation in the predicted values depend on
the change in the observed values? (9)
(iii) Are there any outliers? South Valley station:
8
> 0:8526  1:778:K T ; 0pK T p0:22;
An outlier is defined as a data point that emanates from the >
>
< 0:814  1:106:K T þ 0:366:K 2
different models than do the rest of the data. In our case, T
KD ¼
we have decided to keep the outliers because the tools we >
> 0:997:K 3
T þ 1:221:K 4
T ; 0:22oK T p0:8
>
:
use may become robust with suitable implementation. 0:213; K T 40:8:
According to Fig. 2 the results obtained by ANN model are
generally satisfactory compared to the results of Erbs et al. (10)
[14]. The standard error (SE) between the observed and the If care is taken in considering the corresponding regional
results estimated by ANN model is around 4.18%, while climatic differences these correlations can be generalized
the method of Erbs et al. [14] follows in that order (i.e., and transferred to other sites. Eqs. (8)–(10) were then used
SEffi5.59%). Fig. 2 reveals also that, the ANN model to compute KD for the same three stations. Next, the
showed fairly higher determination coefficient R2. The computed values were compared with the observed data by
range of R2 was 0.95–0.89, while it was 0.87–0.65 for Erbs calculating the MBE, SE, r and RMSE, defined in the
 ARTICLE IN PRESS
H.K. Elminir et al. / Energy 32 (2007) 1513–1523 1517

Fig. 2. Scatter plots of the observed and predicted hourly diffuse fraction. Frequency distribution of the differences between observed and predicted values
were introduced in the right panel.

appendix. To test the symmetry of the computed and
observed data the two widely used statistical indicators
called; skewness and kurtosis were introduced. The
skewness for a normal distribution is zero, and any
symmetric data should have skewness near zero. Negative
values for the skewness indicate data that are skewed left
and positive values for the skewness indicate data that are
skewed right. The kurtosis is a measure of whether the data
are peaked or flat relative to a normal distribution. Positive
kurtosis indicates a peaked distribution and negative
kurtosis indicates a flat distribution. The mentioned
indicators are shown in Fig. 3. Once again, this analysis
 ARTICLE IN PRESS
1518 H.K. Elminir et al. / Energy 32 (2007) 1513–1523
has been preformed over independent data, not used in the
training session.
A study of the results in Fig. 3 shows that the MBE for
Cairo and South Valley are both positive in some cases and
negative in others suggesting that none of the equations
over or under-estimates KD values consistently. The
tendency for overestimation was found only in Aswan.
Even in this case MBE, never exceeds 0.18 MJ/m2. More
specifically, for Cairo site, the mean value of the MBE is
ranged between 0.16 and 0.13 MJ/m2, whereas the other
two sites follow with relatively lower values i.e., for South
Valley, it ranged between 0.14 and 0.09 MJ/m2 and it was
0.02 MJ/m2 for Aswan. As expected, the RMSE obtained
for Cairo (the site with the largest air pollution potential)
was large, but remain in a domain of errors for which these
relations can be applied with good accuracy. The air
pollution causes a decrease in atmospheric transmittance
during most of the year, thus influencing both measured
hours of sunshine and KD. It can be observed that for all
cities, the errors are low in summer (ffi0.24 MJ/m2) and
then increased in the winter (ffi0.36 MJ/m2). For the whole
year, the average values are around 0.35, 0.23 and
0.27 MJ/m2, for Cairo, Aswan and South Valley, respec-
tively. It is also evident that, the SE obtained for Cairo was
large; its mean value is 2.4% compared with 1.9% for the
other two sites. In general, Eqs. (8)–(10) provide a tighter
fit for all sites, although the r appears to be lowest in Cairo.
Fig. 3 shows this detail well. It can be easily seen from
Fig. 3 that, the r for Cairo does not perform so well as
that for the other two sites; its mean value is 0.94 compared
with 0.99 and 0.98 for Aswan and South Valley,
respectively.
As we noted, data observations from one population can
have the same r and SE as data observations from another
population. The two population distributions, however,
can have different shapes. For this reason, additional
estimators namely skewness and kurtosis have been
reported in Fig. 3. The entries of Fig. 3 show that the
observed values and those derived by Eqs. (8)–(10) have
approximately the same elongation and the same kurtosis.
Based on the results obtained, one can say with confidence
that Eqs. (8)–(10) can be used to calculate the hourly values
of diffuse fraction, at least for the Egyptian environment.
4.2. Performance of the models used in daily scale
The discussion in the subsequent part will be focused
only on Aswan site. The main reason for the selection of
Aswan was the availability of the sunshine duration
measurements regarding the covered period. Inspecting
the entries in Fig. 4, it may be observed that the ANN
model yields the best results. A large deviation (up to
0.7 MJ/m2/day) between the measured values and those
predicted by means of Gopinathan and Soler model [15]
has been found. For the ANN model, the deviation for all
months was around 70.1 MJ/m2/day. To be more precise,
Fig. 5 shows that 99.18% of the ANN outputs were
estimated within a deviation of less than 70.1 MJ/m2/day
and 0.82% of these deviations lay within the range from
0.15–0.35 MJ/m2/day. In the case of Gopinathan and
Soler model [15], these values are about 78% and
3%, respectively. It has also been found that 19% of
these deviations lay within the range from 0.15 to
0.65 MJ/m2/day. This finding leads to the necessity of
providing the relative mean bias and root mean square
errors (RMSEs) for both methodologies.
In Fig. 6 we can appreciate clear differences in the
statistical results obtained by the two methods. The MBE
varies between 0.14 and 0.04 MJ/m2/day for the ANN
model. However, it is around 0.16 and 0.22 MJ/m2/day
when observed values are compared with the results of
Gopinathan and Soler model [15]. The RMSE associated
with proposed method of Gopinathan and Soler [15] are
consistently higher and more variable than those obtained
using the ANN model. For the ANN model, the annual
value of the RMSE is about 0.59 MJ/m2/day and it is about
1.14 MJ/m2/day for Gopinathan and Soler model [15]. It is
clear from the entries in Fig. 6 that the model proposed by
Gopinathan and Soler [15] estimate the KD values with SE
in the range of 17.76–4.84%, while the ANN model has the
range of 5.96–2.51%. Moreover, careful examination of
Fig. 6 shows relatively large errors during winter and
spring months. The hot wind known as Khamasin explains
why both methods yield large errors during spring.
Khamasin wind is usually observed during March and
occasionally in April and May. The wind stream carries
large amounts of sand from the nearby desert, thus
influencing the measured KD values. In winter, the high
atmospheric loading for PM10, CO and SO2 was frequently
observed. This reason could explain the high levels in error
during winter. From these findings the conclusion is drawn
that the ANN model is an acceptable tool of the
measurement. An attempt was also made to describe the
ANN outputs in terms of first order polynomials relating
KD with KT and S/S0 using multi-regression techniques.
The analytical form of this polynomial is given by
K D ¼ 0:7980  0:7475:K T  0:0702:S=S 0 . (11)
The performance of the previous correlation employing the
entire data is given in Table 1. It is clear that Eq. (11) was
more accurate and its errors were lower than those
recorded for Gopinathan and Soler model [15]. Over the
entire period, it has been observed that the annual value of
the RMSE was found to be 0.36 MJ/m2/day, which
corresponds to 8.7% in terms of the SE. The annual mean
RMSE found for Eq. (1) derived by Gopinathan and Soler
[15] was 1.14 MJ/m2/day, which corresponds to 10.82% in
terms of the SE. An analysis of the entries in Table 1 shows
a relatively small change in MBE, both correlations tended
to underestimate the KD values. As can be seen, Eqs. (1)
and (11) exhibited low MBE values, implying that they all
have a good long-term representation of the physical
 ARTICLE IN PRESS
H.K. Elminir et al. / Energy 32 (2007) 1513–1523 1519

2 Calculated 3 Calculated
Observed
Observed 2
1.5 Aswan
Skewness

kurtosis
1 0
-1
0.5
-2
Aswan
0 -3

2 3 Calculated
Calculated
Observed 2 South Valley
1.5 Observed
Skewness

kurtosis
1 0
-1
0.5
-2
South Valley
0 -3

2 3
Calculated Calculated
2 Observed
1.5 Observed
Cairo
Skewness

1
kurtosis
1 0
-1
0.5
-2
Cairo -3
0

16 1
Aswan
Standard Error (%)

Correlation Coef.

12 0.9

8 0.8 Cairo
Aswan
4 0.7 South Valley
Cairo
0 0.6

16 2 Aswan
South Valley
Standard Error (%)

South Valley
RMSE (MJ/m2)

1.6
12
Cairo
1.2
8
0.8
4 0.4

0 0

16 0.6 Cairo
Cairo
Standard Error (%)

0.4
MBE (MJ/m2)

12
0.2
8
0
4
-0.2
S. Valley
0 -0.4
J F M A M J J A S O N D J F M A M J J A S O N D
Month Month

Fig. 3. The statistical indicators used to compare between observed and calculated data.
 ARTICLE IN PRESS
1520 H.K. Elminir et al. / Energy 32 (2007) 1513–1523

0.2
ANN's Model

KD (Est.) - KD (Obs.)
0.1

-0.1

-0.2

0.4 Gopinathan and Solar's Model

KD (Est.)- KD (Obs.)

0.2
0
-0.2
-0.4
30 60 90 120 150 180 210 240 270 300 330 360
-0.6 The day of the year
-0.8

Fig. 4. Scatter diagram of the differences between observed and those calculated by the candidate models.

45
ANN's Model
40
Gopinathan and Solar's Model

35
Frequency of occurrences (%)

30

25

20

15

10

0
0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 -0.1 -0.1 -0.2 -0.2 -0.3 -0.3 -0.4 -0.4 -0.5 -0.5 -0.6 -0.6 -0.7
KD (Est.) - KD (Obs.)

Fig. 5. The cumulative frequency of the differences between observed and predicted values.

problem. From the relatively reasonable magnitude of linear regression models was also done. According to the
error of the previous results, it can be concluded that above results, the following findings were obtained:
Eq. (11) can be recommended to forecast the diffuse
fraction in daily scale for Aswan site. 1. For the models used in hourly scale,
(i) The SE of the ANN was about 4.2%, while that by
the Erbs et al. [14] was about 5.6%. Comparing the
5. Summary and concluding remarks RMSE for both models, it can be seen that the ANN
exhibits the best results and furthermore, it is more
The use of ANN technique in modeling diffuse fraction rapid in the forecasting phase.
has been reported. A comparison between the perfor- (ii) The frequency distribution of the difference between
mances of the proposed ANN model with that of two measured data and those estimated by the ANN
 ARTICLE IN PRESS
H.K. Elminir et al. / Energy 32 (2007) 1513–1523 1521

0.25 MBE (ANN)

0.2 MBE (Gop.)
0.15
MBE (MJ/m2/day)

0.1
0.06
0
-0.05
-0.1
-0.15
-0.2

0.9 RMSE (ANN) 20

0.8 RMSE (Gop) 18
SE (ANN)
16
RMSE (MJ/m2/day)

0.7
14 SE (Gop)
0.6
12

SE (%)
0.5
10
0.4
8
0.3
6
0.2 4
0.1 2
0 0
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Fig. 6. Monthly values of the statistical indicators calculated for both methods.

Table 1
Summary of the error analysis

Eq. (11) Gopinathan and Soler model

RMSE (MJ/m2/day) SE (%) MBE (MJ/m2/day) RMSE (MJ/m2/day) SE (%) MBE (MJ/m2/day)

January 0.5635 8.6783 0.0297 0.7389 17.075 0.1327

February 0.5301 4.2167 0.0999 0.5830 17.887 0.1102
March 0.1291 1.8470 0.0641 0.3989 3.6971 0.0717
April 0.3640 2.5154 0.1484 0.7365 12.051 0.1345
May 0.2722 3.0162 0.1483 0.8034 4.5871 0.1443
June 0.1114 1.8656 0.0717 0.4967 3.0300 0.0907
July 0.1364 0.5661 0.0007 0.0837 6.8576 0.0150
August 0.2983 2.8670 0.0552 0.6082 9.0168 0.1092
September 0.4104 2.1813 0.3333 0.5370 5.2601 0.0980
October 0.5472 2.1527 0.3830 0.2835 8.5047 0.2188
November 0.1809 0.9741 0.0928 0.1861 7.2148 0.1602
December 0.2250 3.9877 0.0510 0.2260 8.1275 0.0629

shows that 99.7% of all values were estimated with a
deviation of less than 0.2 MJ/m2. These results were
partly mimicked by Erbs et al. model [14], when
71.7% of the values were estimated with a deviation
of less than 0.2 MJ/m2 and over 90% of these
deviations lay within 0.4 MJ/m2.
2. For the models used in daily scale,
(i) A large deviation (up to 0.7 MJ/m2/day) between
the measured values and those predicted by means
of Gopinathan and Soler model [15] has been
found. For the ANN model, the deviation for all
months was around 70.1 MJ/m2/day.
(ii) The MBE varies between 0.14 and 0.04 MJ/m2/
day for the ANN model. However, it is around
0.16 and 0.22 MJ/m2/day when observed values
are compared with the results of Gopinathan and
Soler model [15].
(iii) The model proposed by Gopinathan and Soler [15]
estimate the KD values with SE in the range of
4.84–17.76%, while the ANN model has the range
of 2.51–5.96%.
These findings show that the neural network is more
suitable to predict diffuse fraction than the proposed
regression models at least for the Egyptian sites. An
attempt was also done to describe the ANN outputs in
terms of first order polynomials relating KD with KT and
S/S0. The analytical forms of these polynomials are given
 ARTICLE IN PRESS
1522 H.K. Elminir et al. / Energy 32 (2007) 1513–1523

by Eqs. (8)–(11). If care is taken in considering the the samplingpffiffiffiffidistribution

ffi and may be estimated by the
corresponding regional climatic differences these correla- formula:s= N where s is the SD of the population
tions can be generalized and transferred to other sites. distribution. The symmetry of the results is evaluated
using the following relationship:
Appendix A P 3
j ðyj  ȳÞ
skewness ¼ , (22)
The hourly and daily extraterrestrial solar irradiations in ðN  1Þ:S 3
units of MJ/m2 were determined using the following
P
equations: j ðyj  ȳÞ4
 po   kurtosis ¼  3, (23)
24 S ðN  1Þ:S 4
G OD ¼ :I SC :E O : cos f: cos d: sin oS  : cos oS ,
p 180
P 2
P 2
(12) N j xj ð j xj Þ
S¼ . (24)
NðN  1Þ
G OH ¼ E O :I SC :ðcos f: cos d: cos oi þ sin f: sin dÞ, (13)
where the meanings of the different symbols are as given in
the nomenclature list. The above (i.e., Eq. (12)) yields References
extraterrestrial radiation for 1 h centered around the hour
angle oi. Appropriate formula for EO, d and oS for each [1] Mosalam Shaltout MA, Hassan AH, Ghitas AE. Optimum tilt angles
day of the year are calculated using of PV systems for electric power generation at Cairo. In: Proceedings
of 11th ECPV solar energy conference, Montreux, Switzerland; 12–16
d ¼ 23:45 sinð360ðð284 þ Dn Þ=365ÞÞ, (14) October 1992. p. 1516–18.
! [2] Mosalam Shaltout MA, El-Hadad A, Fadly M, Hassan AH,
1:000110 þ 0:034221: cos G þ 0:001280: sin G Mahrous A. Determination of suitable types of solar cells for optimal
EO ¼ , outdoor performance in desert climate. Renewable Energy
þ0:000719: cos 2G þ 0:00077: sin 2G 2000;19:71–4.
[3] Dave JV, Norman B. Importance of the diffuse sky radiation in
(15)
evaluation of the performance of a solar cell. Sol Energy
1976;18:215–23.
G ¼ 2pðDn  1Þ=365, (16) [4] Okeke CE, Aunforom AC. Effect of diffuse solar radiation on the
efficiency of an array of silicon solar cells. Sol Energy 1987;5:105–13.
oS ¼ cos1 ð tan f: tan dÞ. (17) [5] Riordan C, Hulstrom R, Cannon T, Myers D. Solar radiation
research for photovoltaic applications. Sol Cells 1991;30:489–500.
The maximum possible number of day light hours is
[6] Aladas-Arboledas L, Batlles FJ, Olmo F. Solar radiation resource
evaluated using the following relationship: assessment by means of silicon cells. Sol Energy 1995;5(3):183–92.
2 [7] Hirata Y, Tani T. Output variation of photovoltaic modules with
S0 ¼ :cos1 ð tan f tan dÞ. (18) environmental factors. Part one: the effect of spectral solar radiation
15 on photovoltaic module output. Sol Energy 1995;55(6):463–8.
The performance of the models was assessed on the basis of [8] Antonio P, Angelo S, Luciano R. Effect of solar irradiation
two characteristics: the mean bias error (MBE) as an conditions on the outdoor performance of photovoltaic modules.
Opt Commun 1998;153:153–63.
indicator of the systematic error and the RMSE which
[9] Liu BYH, Jordan RC. The interrelationship and characteristic
gives an idea of the magnitude of the non-systematic error. distribution of direct, diffuse and total solar radiation. Sol Energy
Additional indicator called correlation coefficient (r) was 1960;4(3):1–19.
used to test the linear relation between predicted and [10] Page JK. The estimation of monthly mean values of daily total
observed data. The mathematical formulas of the previous shortwave radiation on vertical and inclined surfaces from sunshine
records for latitude 401N–401S. Proceedings of UN conference new
indicators are given by
sources energy, vol. 4, Paper No. S98, 1961. p. 378–90.
1 X [11] Ruth DW, Chant RE. The relationship of diffuse radiation to total
MBE ¼ : ðy  xj Þ; (19) radiation in Canada. Sol Energy 1976;18(2):153–4.
N j j
[12] Orgill JF, Holland KGT. Correlation equation for hourly diffuse
radiation on a horizontal surface. Sol Energy 1977;19:357–9.
(P )1=2
j ðyj  xj Þ2 [13] Collares-Pereira M, Rabl A. The average distribution of solar
RMSE ¼ , (20) radiation—correlations between diffuse and hemispherical and
N between daily and hourly insolation values. Sol Energy
1979;22(2):155–64.
P [14] Erbs DG, Klein SA, Duffie JA. Estimation of the diffuse radiation
j ðyj  ȳÞ:ðxj  x̄Þ
r¼nP P o0:5 , (21) fraction for hourly, daily and monthly average global radiation. Sol
½ j ðyj  ȳÞ2 :½ j ðxj  x̄Þ2  Energy 1982;28:293–302.
[15] Gopinathan KK, Soler A. Diffuse radiation models and monthly
where, N is the number of data points, xj is the average, daily diffuse data for a wide latitude range. Energy
1995;20(7):657–67.
observed data point, yj is the predicted data point, x̄ is [16] Trabea AA. Multiple linear correlation for diffuse radiation from
the observed mean value and ȳ is the predicted mean global solar radiation and sunshine data over Egypt. Renewable
value. The SE of a sample from a population is the SD of Energy 1999;17:411–20.
 ARTICLE IN PRESS
H.K. Elminir et al. / Energy 32 (2007) 1513–1523
[17] LeBaron B, Dirmhirn I. Strengths and limitations of the Liu and
Jordan model to determine diffuse from global irradiance. Sol Energy
1983;31:167–72.
[18] Soler A. Dependence on latitude of the relation between the diffuse
fraction of solar radiation and the ratio of global to extraterrestrial
radiation for monthly average daily values. Sol Energy 1990;44:
297–302.
[19] Elminir HK. Experimental and theoretical investigation of diffuse
solar radiation: data and models quality tested for Egyptian sites.
Energy 2007;32(1):73–82.
[20] Simpson PK. Artificial neural systems: foundations, paradigms,
applications and implementations, 1st ed. Elmsford, New York:
Pergamon Press; 1990.
[21] Hertz J, Krogh A, Palmer RG. Introduction to the theory of neural
computation. Canada: Addison-Wesley Publishing Company; 1991.
[22] Alawi S, Hinai H. An ANN based approach for predicting global
radiation in locations with no direct measurement instrumentation.
Renewable Energy 1998;14:199–204.
[23] Mohandes M, Rehman S, Halawani TO. Estimation of global solar
radiation using artificial neural networks. Renewable Energy 1998;
14:179–84.
1523
[24] Kalogirou SA. Applications of artificial neural networks in energy
systems: a review. Energy Convers Manage 1999;40:1073–87.
[25] Sfetsos A, Coonick H. Univariate and multivariate forecasting of
hourly solar radiation with artificial intelligence techniques. Sol
Energy 2000;68(2):169–78.
[26] Hontoria L, Aguilera J, Zufiria P. Generation of hourly irradiation
synthetic series using the neural network multilayer perceptron. Sol
Energy 2002;72:441–6.
[27] Soares J, Oliveira AP, Božnar MZ, Mlakar P, Escobedo JF, Machado
AJ. Modeling hourly diffuse solar-radiation in the city of São Paulo
using a neural-network technique. Appl Energy 2004;79:201–14.
[28] Kalogirou SA. Applications of artificial neural networks in energy
systems: a review. Energy Convers Manage 1999;40:1073–87.
[29] Kalogirou SA. Application of artificial neural networks for energy
systems. Appl Energy 2000;67:17–35.
[30] Kalogirou SA. Application of artificial neural networks in renewable
energy systems applications: a review. Renewable Sustainable Energy
Rev 2001;5:373–401.
[31] Kalogirou SA. Optimization of solar systems using artificial
neural-networks and genetic algorithms. Appl Energy 2004;77(4):
383–405.