Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Solar Energy Vol. 71, No. 2, pp.

95–107, 2001
 2001 Elsevier Science Ltd
Pergamon PII: S 0 0 3 8 – 0 9 2 X ( 0 1 ) 0 0 0 0 8 – 1 All rights reserved. Printed in Great Britain
0038-092X / 01 / $ - see front matter
www.elsevier.com / locate / solener

¨ EQUATION PARAMETER ESTIMATION BY UNRESTRICTED


ANGSTROM
METHOD

ZEKAI~ ŞEN†
~
Istanbul Technical University, Civil Engineering Faculty, Hydraulics Division, Maslak 80626, ~Istanbul,
Turkey

Received 21 February 2000; revised version accepted 19 January 2001

Communicated by RICHARD PEREZ

Abstract—Global solar irradiation is directly related to the sunshine duration record through a linear model
which was first proposed by Angstrom ¨ in 1924. Generally, the parameter estimations of this model are
achieved by using the regression technique based on the least squares method. This technique imposes
procedural restrictions on the parameter estimations leading to under-estimations for clear sky condition and
over-estimations in the case of overcast sky conditions. These restrictions include linearity, normality, means
of conditional distribution, homoscedasticity, autocorrelation and lack of measurement error. In this paper, an
alternative unrestricted method (UM) is proposed for preserving the means and variances of the global
irradiation and the sunshine duration data. In the restrictive regression approach (Angstrom ¨ equation) the
cross-correlation coefficient represents only linear relationships. By not considering this coefficient in the UM,
some nonlinearities in the solar irradiation–sunshine duration relationship are taken into account. Especially
when the scatter diagram of solar irradiation versus sunshine duration does not show any distinguishable
pattern such as a straight line or a curve, then the use of UM is recommended for parameter estimations. The
UM is contrasted with the Angstrom ¨ regression method for 27 solar irradiation stations in Turkey.  2001
Elsevier Science Ltd. All rights reserved.

1. INTRODUCTION where a and b are the model parameters, and


H /H0 and S /S0 are the ratios of meteorological
The amount of global solar irradiation is one of
irradiation and sunshine variables at a location,
the primary variables for determining solar energy
respectively. These ratios vary between zero and
production in a region. Depending on the latitude,
one. From a mathematical point of view a is the
altitude and many meteorological factors, solar
intercept and b is the slope parameter of a straight
irradiation intensities are measured only at a
line expression. Physically, if the sky is complete-
limited number of sites all over the world. The
ly clear, then the meteorological and astronomical
measurement of solar irradiation requires the use
sunshine durations become equal to each other
of pyranometers, however the operation and
(S 5 S0 ) and consequently, Eq. (1) yields a 1 b 5 1.
maintenance of these instruments are not within
For completely overcast situations, the meteoro-
the budget limitations of many local meteorologi-
logical sunshine duration, S, is equal to zero and
cal stations. Therefore, objective interpolations of
hence a is the percentage of irradiation under this
measurements are sought for unmeasured sites
circumstance. Further detailed physical explana-
using global irradiation estimations determined
tions of these parameters are presented by Hin-
from sunshine duration data. Many stations in the
richsen (1994). Many articles have been pub-
world measure the sunshine duration, and it is the
lished for the efficient estimation of model param-
most widely available factor for the solar irradia-
eters from available data (Prescott, 1940; Sabbagh
tion estimations. The relation between the global
et al., 1977; Rietveld, 1978; Gopinathan, 1988;
solar irradiation and the sunshine duration was
Lewis, 1989; Akinoeglu ˇ and Ecevit, 1990; Wahab,
first demonstrated by Angstrom ¨ (1924) using a
1993; Hinrichsen, 1994; Şahin and Şen, 1998).
simple linear model:
In practice, the estimation of model parameters
H S
]5a1b ]
H0 S0 S D (1)
is achieved most often by the least squares
method and regression technique using procedural
assumptions and restrictions in the parameter
estimations. Such restrictions however are un-

Tel.: 190-212-285-3442; fax: 190-212-285-3129; e-mail: necessary because procedural restrictions might
zsen@itu.edu.tr lead to unreliable biases in the parameter estima-

95
96 Z. Şen

tions. One critical assumption for the success of ]] ]]] ]] ]]


the regression equation is that the variables con-
sidered over certain time intervals are distributed
H
H0S D œ
a 5 ] 2 r hs
VarsH /H0d S
]]] ]] ]
VarsS /S0d S0
S D (3)

normally. As the time interval becomes smaller where r hs is the cross-correlation coefficient be-
the deviations from the Gaussian (normal) dis- tween global solar irradiation and sunshine dura-
tribution become greater. For example, the rela- tion data, Var() is the variance of the argument,
tive frequency distribution of daily solar irradia- and the overbars indicate arithmetic averages
tion or sunshine duration is more skewed com- during a basic time interval. Most often in solar
pared to the monthly or annual distribution. The engineering, the time interval is taken as a month
second restriction is that the application of the or a day and in rare cases as a season or a year.
regression technique to Eq. (1) imposes linearity As a result of the classical regression technique,
by the cross correlation coefficient between the the variance of predictand, given the value of
solar irradiation and sunshine duration data. How- predictor is
ever, the averages and variances of the solar ]] ]] ]]
irradiation and sunshine duration data play a
predominant role in many calculations and the
Var ] FS DYS D G
H
H0
S S
] 5 ] 5s1 2 r 2rsdVar ] .
S0 S0
H
H0S D
conservation of these parameters becomes more (4)
important than the cross correlation coefficient in
any prediction model. In Gordon and Reddy This expression provides the mathematical
(1988) it is stated that a simple functional form basis for interpreting r 2rs as the proportion of
for the stationary relative frequency distribution ]]
variability in (H /H0 ) that can be explained by
for daily solar irradiation requires a knowledge of ]]
knowing (S /S0 ). From Eq. (4), after arrange-
the mean and variance only. Furthermore, exami- ments, one can obtain
nation of confidence limits, namely how one can
ascertain the uncertainty range requires knowl-
edge about the mean and variance through the 2
]] ]]
F ]]
Y
VarsH /H0d 2Var sH /H0d sS /S0d 5 S /S0 G
r 5 ]]]]]]]]]]]]
rs ]] .
Chebyschev’s inequality (Benjamin and Cornell, VarsH /H0d
1970). Unfortunately, in almost any estimate of
(5)
solar irradiation by means of computer software,
the parameter estimations are achieved without If, in this expression the second term in the
caring about the theoretical restrictions in the numerator is equal to zero, then the regression
regression approach. This is a very common coefficient will be equal to one. This is tanta-
]]
practice in the use of the Angstrom ¨ equation. A mount to saying that by knowing S /S0 , there is no
]]
critical interpretation of the Angstrom ¨ equation variability in (H /H0 ). Similarly, if it is assumed
]] ]] ]]
has been presented by Gueymard et al. (1995). that Var[(H /H0 ) /(S /S0 ) 5 (S /S0 )] 5Var(H /H0 )
The main purpose of this paper is to present a then the regression coefficient will be zero. This
]]
simple procedural approach to Angstrom ¨ equation means that by knowing (S /S0 ), the variability in
]] 2
parameter estimations without any restrictive as- (H /H0 ) does not change. In this manner, r rs can
sumption imposition. The procedure is based on be interpreted as the proportion of variability in
]] ]]
the preservation of only the variances and aver- (H /H0 ) that is explained by knowing (S /S0 ). In all
ages of the global solar irradiation and sunshine these restrictive interpretations, one should keep
duration data. in mind that the cross-correlation coefficient is
defined for joint Gaussian (normal) distribution of
the global solar irradiation and sunshine duration
2. RESTRICTED MODEL (REGRESSION LINE) data. The requirement of normality is not valid,
especially if the period for taking averages is less
The application of the regression technique to than 1 year. Since, daily or monthly data are used
Eq. (1) for estimating the model parameters from in most practical applications, it is over-simplifi-
the available data leads to (Benjamin and Cornell, cation to expect marginal or joint distributions to
1970) abide with Gaussian (normal) function. There are
]]] ]] six restrictive assumptions in the regression equa-
VarsH /H0d tion parameter estimations such as used in the
b 5 r hs
œ
]]] ]]
VarsS /S0d
(2)
Angstrom¨ equation that should be taken into
consideration prior to any application (Johnston,
and 1980). They are as follows:
¨ equation parameter estimation by unrestricted method
Angstrom 97

1. Linearity: regression technique fits a straight- significant difference between any of these
line trend through a scatter of data points and a variances then the data has homoscedasticity.
correlation analysis test for the ‘goodness-of- 5. Independence: the crux of this assumption is
fit’ of this line. Clearly, if the trend cannot be that the value of each observation on the
represented by a straight-line, regression analy- independent variable (sunshine duration in the
sis will not portray it accurately. The unre- Angstrom¨ equation) is independent of all the
stricted model described in the next section values of all others, so that one cannot predict
does not require such a restriction, since it is the value of (S /S0 ) at time, say i, if one knows
concerned with the variances and arithmetic (S /S0 ) value at time, i 2 1. There are two
averages only. In the case of a definite trend interpretations as to the importance of this
the use of cross-correlation is necessary and it assumption: one is substantively logical and
brings the linearity restriction by definition. the other is statistically logical (Johnston,
2. Normality: it is widely assumed that use of the 1980). The statistical interpretation of indepen-
linear regression model requires that the vari- dence as a special case of autocorrelation
ables have normal distributions. The require- relates to the linearity assumption.
ment is not that the raw data be normally 6. Lack of measurement error: this assumption
distributed but that the conditional distribution requires that both global solar irradiation and
of the residuals should be normally distributed. sunshine duration measurements are without
If the conditional distribution is normal, then it error. If this is not the case, and the magnitude
is almost certain that the distributions of global of the error is not known, then the coefficients
solar irradiation and sunshine duration are also of the regression equation may be biased to an
normally distributed. Thus, it is necessary to extent that cannot be estimated.
test if the data are normally distributed in order If these six assumptions do not apply then the
to inquire as to whether a necessary prereq- regression coefficient estimations may be under
uisite for normal conditional distribution ex- suspicion. The application of the unrestricted
]] ]]
ists. The spatial arrays of monthly (H /H0 ) data model does not require that the (H /H0 ) versus
]]
are generally not normally distributed. In fact, (S /S0 ) scatter diagram should have a distinguish-
from February to October, the distributions are able pattern as a straight line or curve. Akinoglu ˇ
significantly positively skewed, suggesting that and Ecevit (1990) argue that some curvature
a few stations with especially large values of appears in the sunshine–irradiation relationship.
]]
(H /H0 ) produced a monthly mean that is This means inclusion of an extra non-linear term
greater than the mode. July, August, November in such a model. For example, a quadratic third
]]
and December all displayed spatial distribution term on the right hand side of Eq. (1) as (S /S0 )2 .
with significant negative kurtosis levels, sug- Even in this case, however, the classical regres-
gesting that a linear continuous spatial gradient sion approach is used with the same set of
existed with little to no central tendency restrictive assumptions as mentioned above. The
(Balling and Cerveny, 1983). amount of the global bias is not known and
3. Means of conditional distributions: for every therefore, bias correction procedures cannot be
value of sunshine duration, the mean of the defined and applied. Hence, the parameter esti-
differences between the measured and pre- mates of the Angstrom ¨ equation in solar engineer-
dicted global solar irradiation values obtained ing remain under suspicion. In order to avoid
by Eq. (1) must be zero. If it is not, the these restrictive assumptions and rather than
coefficients of the regression equation (a and applying the procedural regression analysis to
b) are biased estimates. Furthermore, the im- data with a set of restrictive assumptions, it may
plication of major departure from this assump- be preferable to try and preserve only the arith-
tion is that the trend in the scatter diagram is metic averages and variances of the global solar
not in fact linear. irradiation and the sunshine duration data. After
4. Homoscedasticity: this means equal variances all, the arithmetic averages and variances are the
in the conditional distributions and it is an most significant statistical parameters in any
important assumption. If it is not satisfied then design work (Gordon and Reddy, 1988).
the regression equation coefficients (a and b)
may be severely biased. In order to test for
homoscedasticity the data must be subdivided
3. UNRESTRICTED MODEL
into three or more groups and the variance of
each group must be calculated. If there is The unrestricted parameter estimations require
98 Z. Şen

two simultaneous equations since there are two ]] ]]] ]] ]]


parameters to be determined. The average and the
variance of both sides in Eq. (1) lead to, without
H
S Dœ
H0
VarsH /H0d S
a9 5 ] 2 ]]] ]] ] .
VarsS /S0d S0
S D (9)

any procedural restrictive assumptions, the fol- Physically, variations in the solar irradiation data
lowing equations are always smaller than the sunshine duration data
]] ]] ]] ]]
and consequently, Var(S /S0 ) $ Var(H /H0 ). For
S D
H
] 5 a9 1 b9 ]
H0 S D
S
S0
(6) Eq. (8), this means that always 0 # b9 # 1. Fur-
thermore, Eq. (8) is a special case of Eq. (2)
and when r hs 5 1. The same is valid between Eqs. (3)
]] ]] and (9). In fact, from these explanations, it is
H
S D
H0
S
Var ] 5 b 9 2 Var ]
S0S D (7) clear that all of the bias effect from the restrictive
assumptions are represented globally in r hs which
where for distinction, the unrestrictive model does not appear in the unrestrictive model param-
parameters are shown as a9 and b9, respectively. eter estimations.
These two equations are the basis for the con- The second term in Eq. (9) is always smaller
servation of the arithmetic mean and variances of than the first one and hence a9 is always positive.
global solar irradiation and sunshine duration The following relationships are valid between the
data. The basic Angstrom ¨ equation remains un- restrictive and unrestrictive model parameters:
changed whether the restrictive or unrestrictive b
model is used. Eq. (6) implies that in both models b9 5 ] (10)
r hs
the centroid, i.e. averages of the solar irradiation
and sunshine duration data, is equally preserved and
(Fig. 1). Furthermore, another implication from ]]
this statement is that both models yield close
estimations around the centroid. The deviations
a
S
a9 5 ] 1 1 2 ]
r hs
1
r hs DS D
H
] .
H0
(11)
between the two model estimations appear at solar
These theoretical relationships between the pa-
irradiation and sunshine duration data values
rameters of the two models imply the following
away from the arithmetic averages. The simulta-
points. Since b and b9 are the slopes of the
neous solution of Eqs. (6) and (7) yields parame-
straight lines, the slope of the restricted
ter estimates as
¨ equation is larger than the unrestricted
(Angstrom)
]]] ]]
VarsH /H0d approach (b9 . b) according to Eq. (10) since
œ
b9 5 ]]] ]]
VarsS /S0d
(8) always 0 # r hs # 1 for global solar irradiation and
sunshine duration data scatter on a Cartesian
and coordinate system (see Fig. 1). As mentioned

Fig. 1. Centroid of solar energy data.


¨ equation parameter estimation by unrestricted method
Angstrom 99

previously, two methods almost coincide practi- are used. For the implementation of the unre-
cally around the centroid (averages of global solar stricted method of Angstrom ¨ equation parameter
irradiation and sunshine duration data). This fur- estimations, 27 global solar radiation stations are
ther indicates that under the light of the previous selected from Turkey. The same data set has been
statement the unrestricted model over-estimates already studied by Şahin and Şen (1998) for the
compared to the Angstrom ¨ estimations for sun- statistical analysis of the Angstrom¨ coefficients.
shine duration data greater than the average value The areal distribution of these stations is shown in
and under-estimates the solar irradiation for sun- Fig. 2 and their location and record features are
shine duration data smaller than the average. On given in Table 1. These locations represent differ-
the other hand, Eq. (11) shows that a9 , a. ent climatological regions of Turkey (Kadioglu, ˇ
Furthermore, the summation of model parameters 1997; Kadiogluˇ et al., 1999).
is At each station 12 years of monthly global
]] irradiation data are available. In order to know
a1b
S
a9 1 b9 5 ]] 1 1 2 ] ] .
r hs
1
DS D
r hs
H
H0
(12) whether the global solar irradiation and sunshine
duration data are normally distributed, the fre-
quency distribution function of the ratios are
These last expressions indicate that the two plotted for months January, April, July and Oc-
approaches are completely equivalent to each tober in Figs. 3 and 4. These include all 27 station
other for r hs 5 1. The new model is essentially monthly averages over Turkey. It is obvious that
described by Eqs. (6), (10) and (11). Its applica- all the frequency distribution functions (FDF) are
tion supposes that the restricted model is first used not Gaussian (normal). This invalidates the re-
to obtain a9, b9 and r. If r is close to 1, then the quirement of normality assumption in the least
¨ equation coefficients estima-
classical Angstrom squares procedural technique. Hence, the
tion with restrictions is almost equivalent to a9 Angstrom¨ equation parameter estimations by the
and b9. Otherwise, the new model results should classical least squares technique remain in bias. A
be considered for application as in Eq. (6). comparison of frequency distribution functions
(FDFs) in Fig. 3 shows that although in January
]]
low (H /H0 ) values appear at high frequencies, in
4. DATA AND APPLICATION ]]
other months higher (H /H0 ) values appear at low
The State Meteorological Department measures frequencies. This is not the expected behaviour
the solar irradiation intensity at 54 stations by the from a normal FDF. Similar features are observ-
]]
use of actinometers. On the other hand, there are able for (S /S0 ) data in Fig. 4 in a more pro-
86 sunshine duration sites where pyrheliometers nounced manner. These FDFs are overwhelmingly

Fig. 2. Global irradiation stations in Turkey.


100 Z. Şen

Table 1. Station features squares technique. In such a situation minor


Station name Latitude Longitude Altitude differences between the estimates of two models
(8) (8) (m)
remain with the scatter diagram appearance. The
Adana 36.98 35.30 20 smaller the scatter around a straight-line, the
Adiyaman 37.75 38.28 678
Afyon 38.75 30.53 1034 smaller the difference between the two methods’
Amasya 40.65 35.85 412 parameter estimations. Furthermore, in Fig. 6
Anamur 36.10 32.83 5 seasonal frequency distribution functions are
Ankara 39.95 32.88 891
Antalya 36.88 30.70 51 shown for Istanbul solar energy data. It is appar-
Aydin 37.85 27.83 56 ent from these FDFs that although they are all
Balikesir 39.65 27.87 102 non-normal with deviations from the Gaussian
Bursa 40.18 29.07 100
Çanakkale 40.13 26.40 6 distribution, as expected, the deviations in the
Çankiri 40.60 33.62 751 monthly FDF are more than the seasonal dis-
Diyarbakır 37.92 40.20 677 tributions. It is obvious that as the averaging time
Erzincan 39.73 39.50 1218
Eskişehir 39.77 30.52 789 period increases the skewness coefficient de-
Istanbul 40.97 29.08 39 creases and the frequency distribution becomes
Isparta 37.70 30.55 997
~Izmir closer to a normal (Gaussian) case. Parameter
38.40 27.17 25
Kars 40.60 43.08 1775 estimations according to restricted and unre-
Kastamonu 41.37 33.77 800 stricted models are given in Table 2.
Kayseri 38.72 35.48 1093 For both the restricted and unrestricted method
Kirşehir 38.13 34.17 985
Konya 37.87 32.50 1031 parameter estimates the frequency distribution
Malatya 38.35 38.30 898 functions over Turkey are presented in Fig. 7.
Mersin 36.82 34.60 5 This figure indicates that both methods yield
Trabzon 41.10 39.72 30
Van 38.47 43.35 1671 FDFs as close to normal function. It is obvious
that the unrestricted method function provides
over-estimates compared with the regression re-
positively skewed. In natural events, the smaller sults for high values of sunshine duration data that
the averaging time period (smaller than a year are greater than the average sunshine duration.
such as hours, days, weeks, months, seasons) the This point may be a reasonable answer to the
more the FDF deviates from normality. Unfor- statement by Gueymard et al. (1995) that for clear
tunately, in all the previous studies where the sky conditions the Angstrom ¨ equation under-esti-
Angstrom¨ equation parameter estimates are ob- mates insulation.
tained by classical least squares procedure, the Finally, in Fig. 8 the two Angstrom ¨ straight
regression coefficient (hence cross-correlation co- lines obtained separately from the unrestricted and
efficient) is valid for normal distribution functions classical regression approaches are presented for
only as explained in the previous section. Since Ankara, Adana and Istanbul stations. It is noticed
solar energy variables are not normally distributed that both straight lines pass through the centroid
]] ]]
at 27 stations in Turkey, one cannot be very (H /H0 ), (S /S0 ) of the scatter diagram, as was
confident about the physical implication of the suggested theoretically in Fig. 1.
parameter estimates except their entrance into the
calculation procedures mechanically in the re-
5. CONCLUSIONS
stricted model. Similar non-normal FDFs are valid
at individual solar irradiation measurement sta- An unrestrictive method of the Angstrom¨ equa-
tions. For instance, Istanbul, Ankara, Adana and tion parameter estimations has been presented in
Diyarbakır locations’ monthly global solar irradia- this paper. The basis of the methodology is the
tion and sunshine duration FDFs are shown in preservation of the averages and the variances of
Fig. 5. Among these FDFs, except at Ankara and global solar irradiation and sunshine duration data
probably Diyarbakır, solar energy variables have at a pyranometer site. It has been observed that in
non-normal distributions. The Ankara station pa- the classical regression technique requirements of
rameter estimations are expected to be closer to normality in the frequency distribution function
each other than those from other sites whether and of linearity and the use of the cross-correla-
restricted or unrestricted models are used. In other tion coefficient are embedded unnecessarily in the
words, in the case of near normal FDFs both parameter estimations. Assumptions in the restric-
methods are expected to yield almost the same tive model cause over-estimations in the solar
results. Normality in the FDFs implies validity of irradiance amounts as suggested by Angstrom ¨ for
the cross correlation coefficient in the least small (smaller than the arithmetic average) sun-
Angstrom
¨ equation parameter estimation by unrestricted method
Fig. 3. Areal frequency distribution functions of H /H0 in (a) January; (b) April; (c) July; (d) December.

101
102
Z. Şen
Fig. 4. Frequency distribution functions of S /S0 (a) January; (b) April; (c) July; (d) December.
Angstrom
¨ equation parameter estimation by unrestricted method
Fig. 5. Monthly frequency distribution functions of H /H0 and S /S0 for (a) Istanbul; (b) Ankara; (c) Adana; (d) Diyarbakır.

103
104
Z. Şen
Fig. 6. Seasonal frequency distribution functions of H /H0 in Istanbul.
Angstrom
¨ equation parameter estimation by unrestricted method
Fig. 7. Comparison of areal parameter frequency distribution functions of (a) restricted; (b) unrestricted model parameter estimations.

105
106 Z. Şen

Fig. 8. Straight lines by restricted and unrestricted methods: (a) Ankara; (b) Adana; (c) Istanbul.
¨ equation parameter estimation by unrestricted method
Angstrom 107

Table 2. Regression and unrestricted method parameter esti- Var(.) variance of the argument
mations ]
X mean of X variable
Station name Restricted Unrestricted H global solar irradiation (W/ m 2 )
]
H average of global solar irradiation (W/ m 2 )
a b a9 b9
H0 astronomical global solar irradiation (W/ m 2 )
Adana 0.331 0.287 0.201 0.501 S sunshine duration (h)
Adiyaman 0.302 0.215 0.255 0.284 ]
S average of sunshine duration (h)
Afyon 0.395 0.277 0.364 0.336
Amasya 0.382 0.382 0.259 0.406 S0 astronomical sunshine duration (h)
Anamur 0.363 0.245 0.261 0.398 r hs cross correlation coefficient
Ankara 0.311 0.323 0.281 0.379
Antalya 0.332 0.382 0.217 0.546
Aydin 0.315 0.421 0.246 0.532
Balikesir 0.225 0.366 0.200 0.414
Bursa 0.270 0.333 0.207 0.462 REFERENCES
Çanakkale 0.314 0.328 0.266 0.410
Çankiri 0.354 0.320 0.318 0.396 Akinogluˇ B. G. and Ecevit A. (1990) Construction of a
Diyarbakir 0.234 0.479 0.165 0.606 quadratic model using modified Angstrom ¨ coefficients to
Erzincan 0.437 0.153 0.343 0.335 estimate global solar radiation. Solar Energy 45(2), 85–92.
Eskişehir 0.391 0.257 0.326 0.387 Angstrom ¨ A. (1924) Solar and terrestrial radiation. Q. J. R.
Istanbul 0.295 0.354 0.271 0.408 Met. Soc. 50, 121–125.
Isparta 0.361 0.160 0.298 0.263 Balling Jr. R. C. and Cerveny R. S. (1983) Spatial and
~Izmir 0.330 0.326 0.246 0.451 temporal variations in long-term normal percent possible
Kars 0.496 0.123 0.338 0.446 radiation levels in the United States. J. Clim. Appl.
Kastamonu 0.323 0.243 0.286 0.292 Meteorol. 22, 1726–1732.
Kayseri 0.364 0.229 0.297 0.357 Benjamin J. R. and Cornell C. A. (1970) In Probability
Kirşehir 0.426 0.196 0.365 0.302 Statistics and Decision for Civil Engineers, p. 684, McGraw
Konya 0.383 0.265 0.300 0.401 Hill, New York.
Malatya 0.307 0.366 0.259 0.448 Gopinathan K. K. (1988) A general formula for computing the
Mersin 0.328 0.395 0.361 0.454 coefficients of the correlation connecting global solar radia-
Trabzon 0.278 0.377 0.233 0.513 tion to sunshine duration. Solar Energy 41(6), 499–502.
Van 0.510 0.143 0.385 0.342 Gordon J. M. and Reddy T. A. (1988) Time series analysis of
daily horizontal solar radiation. Solar Energy 41, 215–226.
Gueymard C., Jidra P. and Eatrada-Cajigai V. (1995) A critical
look at recent interpretations of the Angstrom ¨ approach and
shine duration and under-estimations for large its future in global solar irradiation prediction. Solar Energy
sunshine duration values. Around the average 54, 357–363.
values solar irradiation and sunshine duration Hinrichsen K. (1994) The Angstrom ¨ formula with coefficients
having a physical meaning. Solar Energy 52, 491–495.
values are close to each other for both models, Johnston R. J. (1980) In Multivariate Statistical Analysis in
however, the unrestricted approach alleviates Geography, p. 280, Longman House, Essex.
these biased-estimation situations. Additionally, Kadiogluˇ M. (1997) Trends in surface air temperature data
over Turkey. Int. J. Climatol. 17, 511–520.
the suggested unrestricted model includes some Kadiogluˇ M., Sen Z. and Gultekin ¨ M. L. (1999) Spatial
features of nonlinearity in the solar energy data heating monthly degree-day features and climatological
scatter diagram by ignoring consideration of patterns in Turkey. Theor. Appl. Climatol. 64, 263–269.
Lewis G. (1989) The utility of the Angstrom ¨ type equation for
cross-correlation coefficient. The application of the estimation of global radiation. Solar Energy 45(5),
the methodology is evidenced by data from 297–299.
Turkey. A general conclusion is that only in the Prescott J. A. (1940) Evaporation from water surface in
cases of normal or nearly normal data both relation to solar radiation. Trans. R. Soc. Austr. 40, 114–
118.
restricted and unrestricted methods might be used Rietveld M. R. (1978) A new method for estimating the
with practical indifference. Otherwise, the unre- regression coefficients in the formula relating solar radiation
stricted method must be preferred. to sunshine. Agric. Met. 19, 243–252.
Sabbagh J. A., Sayigh A. A. M. and El-Salam E. M. A. (1977)
Estimation of the total solar radiation from meteorological
data. Solar Energy 19, 307–311.
Şahin A. and Şen Z. (1998) Statistical analysis of the
NOMENCLATURE Angstrom¨ formula coefficients and application for Turkey.
Solar Energy 62, 29–38.
a, b ¨
restricted (Angstrom) method coefficients Wahab A. M. (1993) New approach to estimate Angstrom ¨
a9, b9 unrestricted method coefficients coefficients. Solar Energy 51, 241–245.

You might also like