You are on page 1of 11

Renewable and Sustainable Energy Reviews 82 (2018) 1565–1575

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews


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

Independent models for estimation of daily global solar radiation: A review MARK
and a case study

Muhammed A. Hassan , A. Khalil, S. Kaseb, M.A. Kassem
Mechanical Power Department, Faculty of Engineering, Cairo University, Giza, Egypt

A R T I C L E I N F O A BS T RAC T

Keywords: Independent models are models that estimate daily horizontal global solar radiation without need for prior
Solar energy records of any solar or meteorological parameters such as sunshine duration, cloud cover or temperature. These
Global radiation models are based on the day number, which means that they are very simple, enabling the user to estimate
Independent model global radiation using only a pocket calculator. Though, they are relatively newer and much fewer as compared
Regression analysis
to other dependent models such as sunshine-based models. In this study, a first comprehensive review of all
Empirical models
Day number
such models is provided. Besides, daily radiation data integrated from highly accurate minutely irradiance data
measured at Cairo (Egypt) is used to recalibrate and compare independent models statistically. All models are
trained and cross-validated using a detailed novel approach. The results show that, despite their simplicity,
independent models have very good estimations of solar irradiation with coefficients of determination no less
than 82%.

1. Introduction The oldest method to estimate daily horizontal global solar irradia-
tion was introduced by Angstrom [3] in 1924, in which the ratio
Estimating incident solar radiation on earth's surface using em- between actual global horizontal irradiation (H) and clear day global
pirical models is an old approach. This is mainly due to the lack of solar irradiation (Hc) was correlated linearly to the sunshine fraction (SF).
radiation records at most of locations around the world [1]. Solar The sunshine fraction is defined as actual to theoretical sunshine
radiations records are much scarcer than those of other common duration ratio. The actual sunshine duration (S) can be either
meteorological parameters such as temperature and relative humidity, measured or calculated from direct normal irradiance data, while the
due to many technical and financial issues such as the expensive costs, theoretical sunshine duration (So) is calculated as So = (2/15) cos−1(-
need for highly skilled labor, and requirements of periodical main- tan φ tan δ), where φ and δ are the latitude and declination angles,
tenance, cleaning and calibration of the solar sensors [2]. accordingly. However, the definition and estimation of clear day global
Models of solar radiation can be categorized based on different irradiation are also challenging. Thus, the original correlation was
criteria. For instance, model output (global, beam or diffuse radiation), modified later by Prescott [4], by replacing the clear day global
model input(s) (meteorological data, climatological data or other radiation irradiation term with horizontal extraterrestrial irradiation (Ho), which
components), time scale (daily, monthly average daily, hourly, monthly can be determined mathematically, to form the well-known Angstrom-
average hourly or even minutely basis), spatial coverage (site-dependent Prescott equation:
or global model), methodology (stochastic or time-series modeling), H = Ho(a + bSF ), (1)
approach (physical, semi-physical or empirical), inclination of surface
(horizontal, tilted or tracking surfaces), and type of sky (all sky or clear sky where a and b are the well-known Angstrom coefficients. The model
conditions) [2]. Other classifications may also be considered, including the was refitted later for different locations and it was frequently reported
algorithm used (statistical analysis or machine-learning algorithms), time as a site-dependent model since the Angstrom coefficients vary
coverage (all-year or seasonal), interaction of time-scales (modeling solar significantly according to the location and type of climate [5]. Other
components from others in the same time scale, or from lower time- modified forms of the model have been suggested for sake of better
resolution records). In the next paragraphs, we introduce a concise review estimations, for instance: quadratic [6], cubic [7], logarithmic [8],
of empirical models of daily global horizontal solar radiation. linear-logarithmic [9], exponential [10], linear-exponential [11], ex-


Correspondence to: Mechanical Power Department, Faculty of Engineering, Cairo University, P.O. Box:12613, Giza, Egypt.
E-mail addresses: mhd.zidan17@eng.cu.edu.eg, mhd.zidan17@gmail.com (M.A. Hassan), khalila2@asme.org, adelkhk@yahoo.com (A. Khalil), kaseb@cu.edu.eg (S. Kaseb),
mahmoudkassem2@gmail.com (M.A. Kassem).

http://dx.doi.org/10.1016/j.rser.2017.07.002
Received 21 December 2016; Received in revised form 29 April 2017; Accepted 1 July 2017
Available online 08 July 2017
1364-0321/ © 2017 Elsevier Ltd. All rights reserved.
M.A. Hassan et al. Renewable and Sustainable Energy Reviews 82 (2018) 1565–1575

Nomenclature R2 Coefficient of determination


MBE Mean bias error, MJ/m2.day
H Global horizontal solar irradiation, MJ/m2.day RMSE Root mean square error, MJ/m2.day
Ho Extraterrestrial horizontal solar irradiation, MJ/m2.day MPE Mean percentage error, %
δ Declination angle, degrees NDRMSE Non-dimensional root mean square error, %
N Day number r Correlation coefficient
I Global horizontal irradiance, W/m2
Kt Daily global clearness index

ponent [11] and simple power forms [12]. Other modified models solar radiation have been developed to replace sunshine duration and
included geographical parameters such as latitude angle or was cloud cover by air temperature. The first temperature-based model was
developed in seasonal forms [13], but they are less common. These suggested by Hargreaves and Samani [22], by estimating global
models are usually called: sunshine-based models. It worth noting that radiation using only maximum daily temperature difference (δT) and
the equation is frequently reported in daily and monthly average daily latitude:
time scales, where using the equation for modeling global radiation in a
H = Ho(a δT ), (3)
different time scale may lead to under- or over-estimation of the actual
global radiation values. The inequality of the two time scales was where δT = Tmax - Tmin, and Tmax and Tmin are the maximum and
demonstrated analytically by Muneer in [1]. minimum daily temperatures, respectively, in °C. The empirical
Another category of empirical global radiation models, called: coefficient (a) was initially set to 0.17 for arid and semi-arid regions,
cloud-based models, was developed by correlating daily global clear- while modified later by Hargreaves [23] to 0.16 and 0.17 for interior
ness index (Kt = H/Ho) to another meteorological parameter: cloud and coastal regions, accordingly. Samani [24] suggested a modified
cover (C). Black [14] firstly introduced the method by fitting a form of the original equation by replacing the empirical coefficient with
quadratic correlation between the two variables: a second order polynomial function of δT. Chen et al. [25] and Hunt
et al. [26] added other empirical coefficients for better estimations of
H = Ho(a + bC + cC 2 ). (2)
the model. Another important model in the temperature-based cate-
Here, C is the mean total cloud cover (octal). Paltridge and Proctor gory is the well-known Bristow and Campbell model [27], in which the
[15] proposed a more complex set of equations that estimates the global irradiation is evaluated based on another temperature difference
instantaneous global irradiance and total daily global irradiation using term (ΔT) and three empirical coefficients (a-c):
solar zenith angle, day length and cloud factor (a parameter that varies H = Ho(a(1 − exp(− b ∆T c ))). (4)
from zero during clear sky days to 1 during overcast days). It can be
determined using the cloud cover and the number of cloudy days in Here, ΔT is defined as ∆Tj = Tmax j −0.5(Tmin j + Tmin j+1), where j is a
each month. The authors also provided an equation to convert cloud day index (j for current day and j + 1 for the following day). The
cover records to cloud factor. Badescu [16] modified the original regression coefficients are mainly dependent on the environment type
Black's equation by refitting it into different order polynomials. Other (e.g. humid, arid,.. etc.), with typical values of: a = 0.700, b = 0.004–
modified or refitted models are available in literature [17–19]. 0.001 and c = 2.400 [20,27]. The equation was modified later by
However, this category of models is less common and usually less Goodin et al. [28], by inserting another Ho term in the exponential as a
accurate than the former one [13]. scaling factor. Simpler models that replace sunshine fraction in the
Despite their strong relationship with global radiation, sunshine Angstrom-Prescott equation with temperature terms:
duration and cloud cover records are less available if compared to more H = Ho(a + b(Temperature Term )) were also suggested. The temperature
common meteorological parameters such as temperature and relative term could be the daily average temperature (T) [29] or the tempera-
humidity [20,21]. In addition, their measurements encounter more ture ratio (Tratio, minimum to maximum daily temperature ratio) [30].
inherent sources of error. To deal with this situation, simpler models of An important multivariate linear equation of this type was suggested by

Fig. 1. Graphical representation of the relation between global horizontal radiation and: (a) day number, (b) declination angle, and (c) horizontal extraterrestrial radiation.

1566
M.A. Hassan et al. Renewable and Sustainable Energy Reviews 82 (2018) 1565–1575

Li et al. [31], in which the daily maximum and minimum temperatures of measurements) at 68 locations in Turkey using a trigonometric
were used: function of the day number. Li et al. [45] developed a relatively
H = Ho(a + bTmax + cTmin ) complex sine-cosine form of the correlation between global irradiation
(5)
and day number. They validated the performance of the model by
A hybrid method is frequently adopted, in which different meteor- comparing its results with measurements at nine major stations in
ological variables are used as model's inputs. These models are usually China. They also compared its performance to that of the models
called meteorological parameters-based models, and the commonly proposed by Kaplanis and Kaplani [43], Bulut and Buyukalaca [44],
used meteorological variables are sunshine duration, cloud cover, and the sine-wave model by Al-Salaymeh [42]. The new model was
temperature, daily average pressure (P), water vapor pressure (Pv), found more accurate at the considered locations. Zang et al. [46]
wind speed (W), relative humidity (RH) and precipitable water vapor suggested a close, but simpler, sine-cosine model for estimation of
[13]. Perhaps the most important models in this category are those global radiation over 35 stations in six climatic zones in China to be
proposed by Abdalla [32], who has correlated the global clearness used for generation of typical meteorological years. Their results
index with different meteorological variables using a set of multivariate showed that the new sine-cosine model performs better than the
linear equations: cosine model by Kaplanis and Kaplani [43] and the trigonometric
H = Ho (a + bSF + cδT ), (6) model by Bulut and Buyukalaca [44]. Finally, Quej et al. [47] refitted
the previous models by Bulut and Buyukalaca [44], Kaplanis and
H = Ho (a + bSF + cδT + d RH ), (7) Kaplani [43] and Li et al. [45] for Yucatan Peninsula in Mexico. They
also refitted the sine wave model by Al-Salaymeh [42] but failed to refit
H = Ho (a + bSF + cδT + d RH + e(P / Pv )), (8)
his Gaussian model. Consequently, they modified it by making use of
where a–e are the regression coefficients. There are a numerous summation of two Gaussian function, which resulted in very good
number of such models in the literature with different types of inputs predictions. Based on the available measured data, they found that
and different degrees of complexity [13,33]. In most cases, these their new modified Gaussian model performs better than all other
models provided better estimations of global radiation. However, models.
complex models are not usually the best choice for model users, since In addition to the previously mentioned studies that were dedicated
some of these variables may not be recorded at the considered location. to developing independent global radiation models, some other studies
Besides, adding more inputs does not always improve the model have been carried out to calibrate former models for other locations
estimations [13], and whenever it does, the improvements are not and compare their performance statistically. Khorasanizadeh and
always significant [34]. Thus, these models must be subjected to a Mohammadi [48] used long-term global solar radiation records to test
careful cross-validation procedure to estimate the optimum number six day number-based models (trigonometric model by Bulut and
and type of inputs. Buyukalaca [44], cosine model by Kaplanis and Kaplani [43], sine-
In order to simplify global radiation models, especially for locations cosine model by Li et al. [45], and the Gaussian, 4th order polynomial
where no solar or meteorological variables are recorded, independent and sine-wave models by Al-Salaymeh [42]) at four cities in Iran
models have been recently developed. These models make use of the (Bandarabass, Isfahan, Kerman and Tabass). The sine-cosine wave
semi-linear relation between global irradiation and extraterrestrial model was found better in three cities, while the 4th order polynomial
horizontal irradiation (or declination angle). Alternatively, they can model showed best results at the fourth city. They also assessed the
be developed based on the sinusoidal behavior of global radiation performance of independent models by comparing their estimations
throughout the year, as a function of the day number (Fig. 1). Since with the predictions of some existing monthly averaged models and
these models have no meteorological variables that vary (with random found that the newly refitted independent models are superior.
components) from day to day, they have rough estimations of global Khorasanizadeh et al. [49] carried out a study to compare the accuracy
radiation. Nevertheless, they are frequently reported with good or very of day number-based models with other more common categories,
good statistical measures. There are many reviews on other dependent namely: sunshine- and temperature-based models. The city of Birjand
models [20,21,35–39]. However, there are no comprehensive reviews (Iran) was considered as a case study. Models by Kaplanis and Kaplani
on independent models. An extensive review of such models is [43], Li et al. [45] and Zang et al. [46] have been refitted for the new
provided below. location, validated using long-term measured radiation data, and
Despite their simplicity, independent models are relatively newer compared to the other dependent models. The results revealed the
than other types of global irradiation models. In 1999, Ertekin and superiority of day number-based models as compared to temperature-
Yaldiz [40] proposed a simple-linear model correlating global radiation based models. Sunshine-based models, however, were slightly more
to declination angle. The model was tested using global radiation accurate. The authors concluded that day number-based models are
records of six-years at Antalya, Turkey. In the same year, Togrul and qualified as proper alternatives to sunshine-based models due to their
Onut [41] suggested another close model by replacing the declination simplicity. In a recent study, Hassan et al. [50] used a dataset of more
angle with its sine term. Additionally, they proposed another simple- than 20 years of daily global solar radiation records to refit and validate
linear equation between global irradiation and daily horizontal extra- different day number-based models for ten cities in Egypt. The
terrestrial radiation. The two models were tested at Elazig (Turkey) validated models are those of Bulut and Buyukalaca [44], Kaplanis
using solar records of one-year duration. Al-Salaymeh [42] started and Kaplani [43], Li et al. [45] and the four models proposed by Al-
another method by correlating global irradiation directly to the day Salaymeh [42]. They found that the sine-cosine wave model by Li et al.
number. This is a simpler method since it excludes the prior calcula- [45] and the 4th order polynomial model by Al-Salaymeh [42] have the
tions of declination angle or extraterrestrial radiation. The author best performance measures.
proposed four different forms of the correlation between the two Machine learning algorithms have been adopted recently in many
variables, namely: sinusoidal, 4th order polynomial, Lorentzian and studies in order to refine solar radiation estimations. Despite being
Gaussian forms. The four correlations were tested using measured data widely used to develop sunshine-, temperature- and meteorological-
at Amman, Jordan. It was found that the first three forms provide based models [51–56], very few studies are available for independent
excellent estimations of global radiation at the considered location. models. An adaptive neuro-fuzzy inference system (ANFIS) model was
Kaplanis and Kaplani [43] proposed a cosine-wave equation between developed by Mohammadi et al. [57] using long-term measured data at
global irradiation and day number, which was fitted using solar data of Tabass, Iran. Their results assessed the capability of that algorithm in
six years at six locations in Greece. Bulut and Buyukalaca [44] modeling global solar irradiation. In addition, the authors revealed the
simulated long-term measured radiation data (from 10 up to 19 years superiority of the developed ANFIS model as compared to six

1567
M.A. Hassan et al. Renewable and Sustainable Energy Reviews 82 (2018) 1565–1575

recalibrated empirical models. Gani et al. [58] adopted the neural A close model was suggested by Togrul and Onut [41]:
network auto-regressive model with exogenous inputs (NN-ARX) for (11)
H = a + bsinδ.
predicting irradiation values using only day number as a sole input.
Seven Iranian cities have been considered as case studies. They also Model: M3
developed ANFIS models for the different locations for comparison. They also developed a new method by correlating global radiation
Their results proved the accuracy and reliability of predicting daily to corresponding extraterrestrial radiation using the following simple-
global solar radiation using only day number. The superiority of NN- linear equation [41]:
ARX model over ANFIS was indicated. Hassan et al. [59] investigated H = a + bHo. (12)
the potential of four machine-learning algorithms, namely: multi-layer
perceptron neural networks, support vector machines, ANFIS and Daily horizontal extraterrestrial radiation (Ho) is the integration of
decision trees, in modeling daily global irradiation. The authors instantaneous horizontal extraterrestrial irradiance over the daylight
developed different models of global irradiation for Cairo (Egypt) period [60]:
based on those machine-learning algorithms and other empirical
24*3600*Ic ⎛ ⎛ πω ⎞ ⎞
equations and categorized the developed models in four groups: Ho = *⎜cos φ cos δ sin ωs + ⎜ s ⎟sin φ sin δ ⎟ .
sunshine-, temperature-, meteorological parameters- and day num- π ⎝ ⎝ 180 ⎠ ⎠ (13)
ber-based models. Among their results, the authors found that The solar constant (Ic) is equal to 1367 W/m [2], while ωs is the
2
temperature and day number-based models are reliable in modeling sunset hour angles, respectively:
solar radiation at locations where no sunshine records are available. In
contrast to other groups of models, the proposed decision tree ωs = cos−1(−tan φ tan δ ). (14)
algorithms were found more accurate in modeling global radiation
Model: M4
independently.
Al-Salaymeh [42] proposed four different equations correlating
Based on this careful review of available literature on empirical
global radiation and day number. The first one is a sine-wave equation:
modeling of daily horizontal global radiation, independent models
seem undervalued considering their simplicity and very good estima- ⎛ 2π ⎞
tions. The next sections provide a comparative study of all previous H = a + b sin ⎜ N + d ⎟ ,
⎝ c ⎠ (15)
empirical independent models using solar radiation data measured at
Cairo (Egypt) as a case study. All models are refitted for the new where a, b, c, and d are the amplitude, the phase shift, the wavelength,
location and cross-validated using a novel detailed technique of data and the intercept.
portioning in order to determine the true performance of each model Model: M5
and choose the most suitable model for quantifying global radiation at The second equation is a 4th order polynomial equation:
the considered location.
H = a + bN + cN 2 + d N 3 + eN 4, (16)

2. Materials and methods where a–e are the regression parameters.


Model: M6
2.1. Independent models The third equation has a Lorentzian structure:
a
H= .
In this section, we present all empirical independent models that N−b 2
1+( ) (17)
have been suggested in literature chronologically. c

Model: M1 Here, a, b and c are the amplitude, the center, and the width.
The first model found was a simple linear model, proposed by Model: M7
Ertekin and Yaldiz [40], between global radiation (H) and declination The fourth function has a Gaussian form:
angle (δ):
⎛ 2⎞
1⎛N − b⎞ ⎟
H = a + bδ, (9) H = a exp ⎜⎜ − ⎜ ⎟ ⎟,
⎝ 2⎝ c ⎠ ⎠ (18)
where a and b are the regression parameters. The declination angle
is calculated as a function of the day number (N) as follows: with parameters a, b and c standing for the amplitude, the center, and
⎛ 360 ⎛ ⎞⎞ the width.
δ = 23.45sin⎜ ⎜N +284⎟⎟ . Model: M8
⎝ 365 ⎝ ⎠⎠ (10)
Kaplanis and Kaplani [43] suggested a different cosine-wave
Model: M2 equation based on day number, with three regression coefficients (a-c):

Table 1
Original regression coefficients and reported accuracy of independent models.

Model a b c d e f g Accuracy City

M1 13.580 0.333 – – – – – r = 0.984 Antalya


M2 4.089 6.459 – – – – – r = 0.975 Elazig
M3 −1.320 0.675 – – – – – r = 0.974 Elazig
M4 5250.650 2782.090 375.430 4.783 – – – r = 0.964 Amman
M5 8220.000 183.400 119.800 – – – – r = 0.961 Amman
M6 7905.650 183.210 107.620 – – – – r = 0.960 Amman
M7 2706.825 −15.554 0.836 −0.004 0.000 – – r = 0.964 Amman
M8 15.140 10.326 9.405 – – – – r = 0.996 Athens
M9 21.410 2.570 – – – – – r = 0.840 Istanbul
M10 16.440 −7.889 0.909 8.360 0.519 1.946 7.915 r = 0.972 Minqin
M11 15.976 1.934 0.975 −3.374 0.883 – – r = 0.826 Sanya
M12 10.269 7.633 109.048 41.191 5.355 226.989 43.932 R2 = 0.863 Calakmul

1568
M.A. Hassan et al. Renewable and Sustainable Energy Reviews 82 (2018) 1565–1575

⎛ 2π ⎞ heliometer is used for measuring direct normal irradiance. The three


H = a + b cos ⎜ N + c⎟ .
⎝ 364 ⎠ (19) sensors are mounted on a Kipp & Zonen-SOLYS2 sun tracker with
shading ball assembly. Detailed specifications of the two sensors are
Model: M9 provided in Table 2. The sunshine duration is computed from recorded
Bulut and Buyukalaca [44] proposed a trigonometric form of the pyrheliometer's data instead of being measured directly since it is
underlying function between incident radiation and day number: defined by the World Meteorological Organization (WMO) as the time
⎛ π ⎞ 1.5 during which the beam normal irradiance exceeds the threshold of
H = a + (b − a ) sin ⎜ (N +5)⎟ . 120 W/m2 [60]. Alongside with the radiation sensors, the station is
⎝ 365 ⎠ (20)
equipped a CS215 temperature and relative humidity sensor, an NRG
Model: M10 #40 C anemometer, an NRG #200 P wind direction vane, and a Setra-
Li et al. [45] developed a sine-cosine form of the relation between 278 barometric pressure sensor. A National Instruments data acquisi-
global radiation and day number, with seven regression parameters (a- tion system is used to monitor, record and integrate all measured
g): variables.
The data is recorded in minutely time-resolution. This high resolu-
⎛ 2πc ⎞ ⎛ 2πf ⎞
H = a + b sin ⎜ N + d ⎟ + e cos ⎜ N + g⎟ . tion enables the user to detect any suspicious data, where atmospheric
⎝ 365 ⎠ ⎝ 365 ⎠ (21) parameters exceed their physical limits. It addition, the user can
Model: M11 monitor small variations of variables and take instant actions, e.g.
Zang, H. et al. [46] suggested a close, but simpler model (fewer cleaning the sensors. Besides, when the user needs to estimate hourly
regression parameters): or daily total radiation, the high resolution of collected data results in
more accurate integrated values. In the solar station described above,
⎛ 2πc ⎞ ⎛ 2πe ⎞ solar radiation data has been recorded from October 2012 to December
H = a + b sin ⎜ N ⎟ + d cos ⎜ N ⎟.
⎝ 365 ⎠ ⎝ 365 ⎠ (22) 2016, with one major interruption (April 2013 – December 2013).
Model: M12
Finally, Quej, V. et al. [47] modified the Gaussian model of Al-
2.3. Data filtering and smoothing
Salaymeh for Yucatan Peninsula in Mexico by making use of sum of two
Gaussian functions:
In the majority of studies dedicated to modeling daily global solar
⎛ 2⎞ ⎛ 2⎞ radiation, the dataset is usually available in daily time-resolution from
1⎛ N − c⎞ ⎟ 1 ⎛N − f ⎞ ⎟
H = a + b exp ⎜⎜ − ⎜ ⎟ ⎟ + e exp ⎜⎜ − ⎜ ⎟ ⎟. meteorological administrations. In this study, however, the solar
⎝ 2⎝ d ⎠ ⎠ ⎝ 2⎝ g ⎠ ⎠ (23) irradiance (in W/m2) is measured in minutely time-resolution in order
The original regression coefficients, as well as the reported accuracy to be integrated to daily time-resolution (in MJ/m2.day). Therefore, a
of all previously discussed models, are listed in Table 1. In case if detailed procedure is described below for filtering and smoothing
regression coefficients are reported for different locations, only one measured data.
location is selected to be listed in the table. The correlation coefficient While ground measurements of solar radiation are the best method
(r) is defined as: for the knowledge of incident radiation, errors can be expected as in
n
any type of measurements. Solar radiation measurements, in particu-
∑i =1 (Hc − Hc )(Hm − Hm ) lar, are more likely to have errors than measurements of other
r=
n n
[∑i =1 (Hc − Hc )2][∑i =1 (Hm − Hm )2 ] meteorological parameters. Usually, questionable values appear even
(24)
when sensor cleaning, maintenance, and calibration procedures are
th
where n is the number of observations. Hc and Hm are the i calculated performed periodically [1,61]. In order to make sure that all measured
and measured values of global irradiation, respectively. While Hc and data points are reliable, a hybrid procedure is adopted based on the
Hm are the means of calculated and measured values, all in MJ/m2.day. restrictions suggested by Jacovides et al. [62] Reindl et al. [63], and
Scharmer and Greif [64]. Firstly, data points of night or blocked
2.2. Measured data radiation periods are excluded: I (global irradiance) ≥ 0. Periods of low
sun elevations are also excluded due to the cosine response: α (solar
The solar radiation dataset used in this study has been collected at altitude angle) ≥ 5°, and I ≥ 5 W/m2. A 10% allowance is considered for
Faculty of Engineering, Cairo University (El-Sheikh Zayed annex, shading ring correction: Id /I (diffuse fraction, kd: diffuse to global
latitude: 30.04°N, longitude: 31.01°E). A solar-meteorological station irradiance ratio) ≤ 1.1. Global irradiance values have to be within the
is operated since 2012 to measure radiation and atmospheric para- range of the expected clear-sky extreme values by considering the
meters. As for solar radiation sensors, the station is equipped with two influence of the atmospheric layer: I/Io,h (global clearness index, kt:
Kipp & Zonen-CMP21 pyranometers for measuring horizontal global global to extraterrestrial horizontal irradiance ratio) ≤ 1.2. Here, the
and diffuse irradiance components, while a Kipp & Zonen-CHP1 pyr- value of the clearness index may exceed 1 for minutely data due to

Table 2
Specifications of CMP21 pyranometer and CHP1 pyrheliometer.

CMP21 pyranometer CHP1 pyrheliometer

ISO classification Secondary standard First class


Response time (95%) <5s 5s
Zero offsets due to ambient temperature change (5 K/h) ± 2 W/m2 ± 1 W/m2
Non-stability (change/year in percentage of full scale) ± 0.5% ± 0.5%
Non-linearity ± 0.2% (at 500 W/m2 within 100–1000 W/m2) ± 0.2% (0–1000 W/m2)
Temperature dependence of sensitivity ± 1% (−20 °C to + 50 °C) ± 0.5% (−20 °C to + 50 °C)
Expected daily uncertainty (in clean state) ± 2% ± 1%
Full opening view angle 180° 5° ± 0.2°
Operating temperature −40 °C to + 80 °C −40 °C to + 80 °C
Spectral range (50% points) 310–2800 nm 200–4000 nm

1569
M.A. Hassan et al. Renewable and Sustainable Energy Reviews 82 (2018) 1565–1575

sunlight reflecting off clouds and into the sensor. However, as the time sample are saved, and the regression coefficient considered for this
resolution decreases (e.g. hourly or daily means), the number of data sample are those of the fold that has the closest RMSE to the average
points with values near or greater than 1 significantly decreases [65]. RMSE of all folds. At this point, the model has been cross-validated
The normal irradiance must be less than its corresponding extrater- using a single random sample. The performance of the model is highly
restrial component: In ≤ Io. Measured data must lie within the expected dependent on the order of observations in the data matrix, especially
kd-kt envelop: Id/Io,h ≤ 0.8, Id/I ≥ 0.9 for kt < 0.2, and Id/I ≤ 0.8 for kt when the data points are serially correlated [69]. In order to get a more
> 0.6. Another set of restrictions is considered for integrated or honest performance indicators of model, the whole procedure de-
averaged daily values. Upper and lower limits of daily mean clearness
index (Kt) are considered as 1 and 0.015, respectively. Daily total
sunshine duration is also checked such as to be less than or equal to its
corresponding astronomical value (SF < 1).
Missing data is caused by equipment failure, cleaning and main-
tenance routines, equipment being offline, or due to discarding
recorded values when applying the quality control procedure.
Different methods have been used for estimating missing values such
as Average Nearest Observation (ANO) [66]. The ANO is a very simple
method since it replaces missing values with the mean of nearest
previous and following observations. However, it generates different
estimations depending on the marching direction (i.e. forward or
backward). Therefore, it has a poor performance when applied to time
series data that has weak autocorrelation and/or strong daily season-
ality [67]. The ANO method, however, is the basis of the advanced Two-
directional Exponential Smoothing method (TES) used here, which
estimates the missing observations based on the autocorrelations of the
time series to account for the fact that the missing values occur at non-
random times. Firstly, all missing observations are generated using the
ANO method in forward and backward directions. Then, the missing
values are estimated using Holt's linear trend method: Hi = 0.5
(Hi,forward + Hi,backward) [68].

2.4. Data portioning

The common approach in modeling solar radiation using empirical


models is to split measured data without randomization into two sub-
datasets for training and validating the models. For instance, if a
dataset of four years is available, three years of measurements may be
chosen for training stage while the fourth year is kept for model
validation. This approach, however, does not account for the annual
variations of solar radiation, especially when the dataset is too large,
i.e. covering a decade or more. Another issue is that choosing the
optimum model, in case of comparing different models, is subjective.
The performance of each model highly depends on which data points
have been chosen for validation, and which have been chosen for
training.
Another simple technique of portioning measured data and validat-
ing the models is suggested here based on the repeated K-fold cross-
validation method. Instead of splitting the data chronologically,
measured data are randomized using the software (MATLAB
R2016a®), and the only interference is choosing the number of folds
(K) and the number of samples (U) in the cross-validation process. As
shown in Fig. 2, solar radiation data is collected and the quality control
procedure, described above, is applied to raw data. Next, TES method
is used to handle missing data. Measured data is now used to construct
an n ×P + 1 matrix, where P is the number of inputs (the extra column
in the matrix is for the target output). Afterward, the rows (observa-
tions) of the matrix are randomized for the first sample, and the K-folds
cross-validation method is applied. Using 10 folds (K = 10), the
randomized data matrix is split into 10-equally sized sub-matrices.
The first nine sub-matrices are used to construct the training data
matrix, while the remaining one is kept for validation. Each model is
fitted using the training matrix and then validated using the validation
matrix [69]. The resulted regression coefficients and validation statis-
tical indicators are saved. After that, another sub-matrix is preserved
for validation, while the other nine are used for training. This process is
repeated until all data points are used for training and validating the Fig. 2. A flowchart of the procedure used for training and validation of independent
model. The average statistical indicators of all folds of this random models.

1570
M.A. Hassan et al. Renewable and Sustainable Energy Reviews 82 (2018) 1565–1575

n
Table 3 ∑i =1 (Hc − Hm )2
Recalibrated regression parameters of independent models. R2 = 1− n
∑i =1 (Hm − Hm )2 (29)
Model Model parameters
It worth noting that a conflict usually appears when comparing
a b c d e f g different models based on different statistical indicators, where a model
M1 20.507 0.357 – – – – – may have the best performance based on one statistical indicator (e.g.
M2 20.486 21.044 – – – – – MBE), while another model is the best based on another statistical
M3 −2.173 0.711 – – – – – indicator (e.g. RMSE) [2]. In such cases, the model with lowest RMSE
M4 19.969 −8.704 1.003 13.782 – – – is preferred, since the RMSE is usually better in showing the true
M5 11.843 0.027 0.002 0.000 0.000 – –
M6 28.906 172.492 143.213 – – – –
accuracy of the model.
M7 28.015 173.574 123.267 – – – – Table 3 shows the recalibrated regression parameters of all
M8 20.546 −8.422 6.454 – – – – considered models for the new location. Table 4, on the other hand,
M9 10.869 28.241 – – – – – provides the statistical measures of training and validation stages. It is
M10 20.297 −0.566 2.487 −5.641 −8.421 0.966 6.553
clear that all refitted models, except for models M11 and M12, have
M11 20.512 41.718 −0.005 −2.542 2.237 – –
M12 21.240 −9.269 5.906 −23.907 −3.331 51.486 −14.839 very good estimations of global solar radiation with all R2 values
greater than 82%, and all NDRMSE values less than 8.5%. The results
also suggest that the recalibrated 4th order polynomial model is the
scribed is repeated using different random order of observations. In best choice for modeling daily global radiation without need to
this study, the process is repeated 1000 times for each model (U = meteorological data. It has validation RMSE and R2 of 7.761% and
1000). For each sample, the regression coefficients and statistical 86.231%, respectively, which is very appealing regarding its simplicity.
indicators are saved. The reported statistical indicators are the average It also performs better in training stage. Its competitors are the
of corresponding statistical indicators of all samples, while the reported recalibrated sine-cosine model by Li et al. [45] and the two declination
regression coefficients are those of the sample that has the closest angle-based models. Models M11 and M12, on the other side, failed to
RMSE to the average RMSE of all samples. generalize to the new data even when single training/validation step is
performed. The original models (without recalibration) also showed
poor performance in the new location, suggesting that they have poor
3. Results and discussion generalization characteristics as compared to elder models, which have
been refitted before for different locations.
The previously discussed independent models are recalibrated here Fig. 3 shows scatter plots of measured daily global irradiation
for the new location and cross-validated according to the procedure against estimated values for all trained models. Data points are
provided in the previous section. The performance of these models is distinguished according to their random assignment (training or
evaluated based on different statistical measures, namely: Mean Bias validation). The shown plots are for the fold (k) with closest RMSE
Error (MBE), Mean Percentage Error (MPE), Root Mean Square Error to the average RMSE of all folds, in the sample (u) with closest RMSE
(RMSE), Non-Dimensional Root Mean Square Error (NDRMSE) and to the average RMSE of all samples. The failure of models M11 and
Coefficient of Determination (R 2 ), defined as: M12 in generalization to new data is shown visually. Fig. 4, on the
n
∑i =1 (Hc − Hm ) other hand, depicts the distribution of models’ residuals against
MBE = , MJ/m2.day, estimated radiation values. All models have almost random distribu-
n (25)
tions, except for models M11 and M12 whose residuals have apparent
n⎛ Hc − Hm ⎞ patterns.
1
MPE = ∑⎜ ⎟*100%, A time-series plot of measured and estimated values of global
i =1 ⎝
n Hm ⎠ (26) radiation is shown in Fig. 5 for all successfully validated models. Since
all models depend mainly on the day number, all curves have close
n
1 trends throughout the year with some significant deviations at the
RMSE = ∑ (Hc − Hm )2 ,MJ/m2.day,
n i =1 (27) beginning and the end of the year. Model M3 has a different behavior
due to the sine term of sunrise angle in the extraterrestrial radiation
n ⎛ Hc − Hm ⎞2 equation. The inset in the same figure is for model M5, where only 20
1
NDRMSE = ∑⎜ ⎟ *100%, random days in 2014 are shown. For these days, the uncertainty
i =1 ⎝
n Hm ⎠ (28) bounds of measurements, as well as the 95% confidence interval of the

Table 4
Training and validation statistical measures of re-fitted models. Best values are shown in bold font.

Model Training Validation

MBE RMSE R2 MPE % MBE RMSE NDRMSE % R2

M1 0.000 2.347 0.862 0.002 0.000 2.330 7.845 0.860


M2 0.000 2.345 0.862 0.002 0.000 2.328 7.837 0.860
M3 0.000 2.374 0.859 0.000 0.000 2.355 7.927 0.856
M4 0.000 2.488 0.832 −0.006 −0.002 2.473 8.328 0.828
M5 0.000 2.318 0.866 0.001 0.000 2.305 7.761 0.862
M6 0.008 2.519 0.828 0.039 0.011 2.507 8.442 0.825
M7 −0.028 2.435 0.852 −0.094 −0.028 2.419 8.143 0.849
M8 0.000 2.345 0.862 0.001 0.000 2.336 7.863 0.859
M9 0.000 2.403 0.855 0.000 0.000 2.387 8.037 0.853
M10 0.000 2.319 0.866 −0.002 −0.001 2.312 7.784 0.861
M11 0.053 5.189 0.266 0.125 0.036 5.176 17.422 0.255
M12 0.000 5.668 0.201 0.005 0.000 5.651 19.024 0.188

1571
M.A. Hassan et al. Renewable and Sustainable Energy Reviews 82 (2018) 1565–1575

Fig. 3. Scatter plots of estimated and measured radiation values.

Fig. 4. Scatter plots of models’ residuals.

model, are displayed. Good agreement between the two intervals can be and the non-linear meteorological parameters-based model by Ododo
noticed. et al. [72]. All dependent models have been trained and validated using
For further assessment of the prediction performance, the inde- the same procedure of independent models. Table 5 provides a list of
pendent models are compared to five different dependent models, those dependent models, alongside with their corresponding refitted
namely: the 3rd order polynomial sunshine-based model by Zabara [7], parameters and statistical measures. The provided results demonstrate
the linear temperature-based model by Chen et al. [25], the non-linear the importance of sunshine duration in estimating the global solar
meteorological parameters-based model by Allen [70], the multivariate irradiation. The two models with the sunshine fraction is an input
linear meteorological parameters-based model by El-Sebaii et al. [71], (models: D1 and D5) have superior performance as compared to other

1572
M.A. Hassan et al. Renewable and Sustainable Energy Reviews 82 (2018) 1565–1575

Fig. 5. Time-series plot of models’ estimations in comparison with measured radiation values.

Table 5
Statistical results of refitted dependent models. Best values are shown in bold font.

Model Ref. Equation Refitted parameters Training Validation

a b c d e MBE RMSE R2 MBE MPE % RMSE NDRMSE % R2

D1 [7] H = Ho(a + b SF + c SF2 + d SF3) 0.437 −0.125 0.858 −0.451 – −0.036 1.362 0.954 −0.120 −0.036 1.364 4.593 0.952
D2 [25] H = Ho(a + b δT0.5) 0.345 0.080 – – – 0.008 2.162 0.883 0.022 0.006 2.158 7.265 0.879
D3 [70] H = Ho(a(P/Po)0.5 δT0.5) 0.172 – – – – −0.259 2.462 0.848 −0.871 −0.259 2.427 8.172 0.845
D4 [71] H = Ho(a + bT + c RH) 0.727 0.001 −0.002 – – 0.041 2.311 0.866 0.142 0.042 2.298 7.737 0.863
D5 [72] H = Ho(a + b SF + cTmax + d RH + e 0.579 0.228 −0.005 −0.001 0.005 0.006 1.324 0.956 0.021 0.006 1.322 4.452 0.955
Tmax SF)

Fig. 6. A graphical comparison between models D2, D5 and M5: (a) statistical comparison plot using validation dataset, (b) one-year time-series plot.

dependent and dependent models in both training and validation the best independent model (model M5) is shown in Fig. 6. Fig. 6a
stages. On the other side, when the sunshine fraction is excluded shows a statistical comparison using validation dataset (validation data
(models: D1-D3), the performance of the model drops significantly to points are different for each model due to the randomization proce-
the accuracy level of independent models. This can be also noticed dure), while Fig. 6b shows a one-year time-series plot of the three
when comparing the statistical measures of models D1 and D5, where models. While model D5 is capable of capturing the variation of global
adding temperature and relative humidity terms slightly improved the radiation during overcast days, the independent model provides rough
predictions. This suggests that complicating the models by adding estimations of global radiation behavior throughout the year.
some meteorological parameters such as temperature, mean sea-level Finally, Fig. 7 demonstrates the importance of cross-validating the
pressure or relative humidity is not worthy, and the independent models using random samples. Histograms of validation RMSE are
models are preferred in locations where sunshine duration is not displayed for three models: M4, M5, and M10. The histograms show
recorded. A graphical comparison between model D2 (temperature- the distribution of all RMSEs of the 1000 samples, while the vertical
based model), model D5 (meteorological parameters-based model) and lines indicate the average of these RMSEs (as given in Table 4). Here,

1573
M.A. Hassan et al. Renewable and Sustainable Energy Reviews 82 (2018) 1565–1575

Fig. 7. Histograms of validation RMSE: models M4, M5 and M10.

the performances of models M5 and M10 are appealing since they show [4] Prescott JA. Evaporation from a water surface in relation to solar radiation. Trans R
Soc South Aust 1940;64: p. 114–25.
fewer variations in their accuracies with change of input data. The two [5] Myers DR. Solar radiation: practical modeling for renewable energy applications,
distributions are close to a normal one. Also, their RMSEs are much 1st ed.. New York: CRC Press; 2013.
lower than that of model M4, which suffers from multimodal nature (its [6] Ogelman H, Ecevit A, Tasdemiroglu E. A new method for estimating solar radiation
from bright sunshine data. Sol Energy 1984;33:619–25.
accuracy in estimating global irradiation is highly affected by the [7] Zabara K. Estimation of the global solar radiation in Greece. Sol Wind Technol
training and validation entries). The most important issue here is that 1986;3:267–72.
there is an overlap in the distributions of the three models, which [8] Ampratwum B, Dorvlo AS. Estimation of solar radiation from the number of
sunshine hours. Appl Energy 1999;63:161–7.
means that choosing the best model from a set of different models [9] Newland FJ. A study of solar radiation models for the coastal region of South China.
based on chronological portioning of data is somehow subjective. Sol Energy 1988;31:227–35.
Model M5, which has the best generalization abilities among all models [10] Almorox J, Benito M, Hontoria C. Estimation of monthly Angstrom–Prescott
equation coefficients from measured daily data in Toledo, Spain. Renew Energy
considered here, may result in a greater RMSE than other models if a
2005;30:931–6.
single sample is used. Adopting the procedure described here leads to a [11] Bakirci K. Correlations for estimation of daily global solar radiation with hours of
more honest judgment of the true performance of validated models. bright sunshine in Turkey. Energy 2009;34:485–501.
[12] Sen Z. Solar energy fundamentals and modeling techniques: atmosphere environ-
ment, climate change and renewable energy, 1st ed.. London: Springer; 2008.
4. Conclusions [13] Besharat F, Dehghan AA, Faghih AR. Empirical models for estimating global solar
radiation: a review and case study. Renew Sustain Energy Rev 2013;21:798–821.
In this paper, a comprehensive review of all studies concerned with [14] Black JN. The distribution of solar radiation over the earth's surface. Arch Met
Geophys Bioklim 1956;7:165–89.
modeling horizontal global radiation using independent models is [15] Paltridge GW, Proctor D. Monthly mean solar radiation statistics for Australia. Sol
provided. All empirical independent models have been recalibrated Energy 1976;18:235–43.
and cross-validated using a detailed novel method. The results assess [16] Badescu V. Correlations to estimate monthly mean daily solar global irradiation:
application to Romania. Energy 1999;24:883–93.
the accuracy of independent models considering their simplicity, with [17] Reddy SJ. An empirical method for estimating sunshine from total cloud amount.
coefficient of determination ranging from 82.5% up to 86.2%. When Sol Energy 1974;15:281–5.
compared to other dependent models, the independent models were [18] Supit I, Van Kappel RR. A simple method to estimate global radiation. Sol Energy
1998;63:147–60.
less accurate than sunshine-based models due to the strong relation
[19] Ehnberg JSG, Bollen MHJ. Simulation of global solar radiation based on cloud
between sunshine duration and global irradiation. However, they were observations. Sol Energy 2005;78:157–62.
at the same accuracy level of temperature- and meteorological para- [20] Almorox J, Hontoria C, Benito M. Models for obtaining daily global solar radiation
with measured air temperature data in Madrid (Spain). Appl Energy
meters-based models (with sunshine fraction is not an input). This
2011;88:1703–9.
suggests that they are more suitable for modeling global radiation at [21] Benghanem M. Solar radiation estimated from measured air temperature: a review
locations where no sunshine records are obtainable. The importance of and proposed new model. Int J Renew Energy Technol 2013;4:191–211.
adopting the provided procedure of randomizing and portioning the [22] Hargreaves GH, Samani ZA. Estimating potential evapotranspiration. J Irrig Drain
Eng 1982;108:225–30.
data and cross-validating the models is demonstrated, to be used in [23] Hargreaves GH. Simplified coefficients for estimating monthly solar radiation in
studies dealing with modeling solar radiation empirically. North America and Europe. Departmental Paper, Dept of Biol and Irrig Engrg,
Utah State University, Logan; 1994.
[24] Samani Z. Estimating solar radiation and evapotranspiration using minimum
References climatological data (Hargreaves-Samani equation). J Irrig Drain Eng
2000;126:265–7.
[1] Muneer T. Solar radiation and daylight models, 2nd ed.. New York: Elsevier; 2004. [25] Chen R, Kang E, Lu S, Yang J, Ji X, Zhang Z, Zhang J. New methods to estimate
[2] Gueymard CA, Myers DR. Validation and ranking methodologies for solar radiation global radiation based on meteorological data in China. Energy Convers Manag
models. In: Badescu V, editor. Modeling solar radiation at the earth's surface1st ed.. 2006;47:2991–8.
New York: Springer; 2008. p. 479–509. [26] Hunt LA, Kuchar L, Swanton CJ. Estimation of solar radiation for use in crop
[3] Angstrom A. Solar and terrestrial radiation. Report to the international commission modelling. Agric Meteorol 1998;91:293–300.
for solar research on actinometric investigations of solar and atmospheric radiation. [27] Bristow KL, Campbell GS. On the relationship between incoming solar radiation
Q J R Meteorol Soc 1924;50: p. 121–5. and daily maximum and minimum temperature. Agric Meteorol 1984;31:159–66.
[28] Goodin DG, Hutchinson L, Vanderlip R, Knapp M. Estimating solar irradiance for

1574
M.A. Hassan et al. Renewable and Sustainable Energy Reviews 82 (2018) 1565–1575

crop modeling using daily air temperature. Data Agron J 1999;91:845–51. [52] Piri J, Kisi O. Modelling solar radiation reached to the earth using ANFIS, NN-
[29] Kirmani S, Jamil M, Rizwan M. Empirical correlation of estimating global solar ARX, and empirical models (Case studies: Zahedan and Bojnurd stations). J Atmos
radiation using meteorological parameters. Int J Sustain Energy 2015;34:37–41. Sol-Terr Phys 2015;123:39–47.
[30] Adaramola MS. Estimating global solar radiation using common meteorological [53] Mohammadi K, Shamshirband S, Danesh AS. Temperature-based estimation of
data in Akure, Nigeria. Renew Energy 2012;47:38–44. global solar radiation using soft computing methodologies. Theor Appl Climatol
[31] Li M, Liu H, Guo P, Wu W. Estimation of daily solar radiation from routinely 2015;125:101–12.
observed meteorological data in Chongqing, China. Energy Convers Manag [54] Shamshirband S, Mohammadi K, Lip P, Petkovi D. A comparative evaluation for
2010;51:2575–9. identifying the suitability of extreme learning machine to predict horizontal global
[32] Abdalla YAG. New correlations of global solar radiation with meteorological solar radiation. Renew Sustain Energy Rev 2015;52:1031–42.
parameters for Bahrain. Int J Sol Energy 1994;16:111–20. [55] Ramedani Z, Omid M, Keyhani A, Shamshirband S, Khoshnevisan B. Potential of
[33] El-Metwally M. Simple new methods to estimate global solar radiation based on radial basis function based support vector regression for global solar radiation
meteorological data in Egypt. Atmos Res 2004;69:217–39. prediction. Renew Sustain Energy Rev 2014;39:1005–11.
[34] Mohammadi K, Khorasanizadeh H, Shamshirband S, Tong CW. Influence of [56] Yadav AK, Chandel SS. Solar radiation prediction using artificial neural network
introducing various meteorological parameters to the Angstrom–Prescott model for techniques: a review. Renew Sustain Energy Rev 2014;33:772–81.
estimation of global solar radiation. Environ Earth Sci 2015;75:1–12. [57] Mohammadi K, Shamshirband S, Wen C, Amjad K, Petkovic D. Potential of
[35] Al-Mostafa ZA, Maghrabi AH, Al-Shehri SM. Sunshine-based global radiation adaptive neuro-fuzzy system for prediction of daily global solar radiation by day of
models: a review and case study. Energy Convers Manag 2014;84:209–16. the year. Energy Convers Manag 2015;93:406–13.
[36] Yao W, Li Z, Wang Y, Jiang F, Hu L. Evaluation of global solar radiation models for [58] Gani A, Mohammadi K, Shamshirband S, Khorasanizadeh H, Seyed Danesh A, Piri
Shanghai, China. Energy Convers Manag 2014;84:597–612. J, Ismail Z, Zamani M. Day of the year-based prediction of horizontal global solar
[37] Almorox J, Hontoria C. Global solar radiation estimation using sunshine duration radiation by a neural network auto-regressive model. Theor Appl Climatol
in Spain. Energy Convers Manag 2004;45:1529–35. 2016;125:679–89.
[38] Almorox J, Bocco M, Willington E. Estimation of daily global solar radiation from [59] Hassan MA, Khalil A, Kaseb S, Kassem MA. Potential of four different machine-
measured temperatures at Canada de luque, Cordoba, Argentina. Renew Energy learning algorithms in modeling daily global solar radiation. Renew Energy
2013;60:382–7. 2017;111:52–62.
[39] Yacef R, Mellit A, Belaid S. New combined models for estimating daily global solar [60] Kalogirou SA. Solar energy engineering: processes and systems, 1st ed.. London:
radiation from measured air temperature in semi-arid climates: application in Elsevier Inc; 2009.
Ghardaia, Algeria. Energy Convers Manag 2014;79:606–15. [61] Moradi I. Quality control of global solar radiation using sunshine duration hours.
[40] Ertekin C, Yaldiz O. Estimation of monthly average daily global radiation on Energy 2009;34:1–6.
horizontal surface for Antalya (Turkey). Renew Energy 1999;17:84–91. [62] Jacovides CP, Tymvios FS, Assimakopoulos VD, Kaltsounides NA. Comparative
[41] Togrul IT, Onat E. A study for estimating solar radiation in Elazig using study of various correlations in estimating hourly diffuse fraction of global solar
geographical and meteorological data. Energy Convers Manag 1999;40:1577–84. radiation. Renew Energy 2006;31:2492–504.
[42] Al-Salaymeh A. Modelling of global daily solar radiation on horizontal surfaces for [63] Reindl DT, Beckman WA, Duffie JA. Diffuse fraction correlations. Sol Energy
Amman City. Emir J Eng Res 2006;11:49–56. 1990;45:1–7.
[43] Kaplanis S, Kaplani E. A model to predict expected mean and stochastic hourly [64] Scharmer K, Greif J, editors.European solar radiation atlas. Vol. 1: fundamentals
global solar radiation I(h;nj) values. Renew Energy 2007;32:1414–25. and maps. Paris: Les Presses de l’Ecole, Ecole des Mines; 2000.
[44] Bulut H, Buyukalaca O. Simple model for the generation of daily global solar- [65] Vignola F, Michalsky J, Stoffel T. Solar and infrared radiation measurements. New
radiation data in Turkey. Appl Energy 2007;84:477–91. York: CRC Press; 2012.
[45] Li H, Ma W, Lian Y, Wang X. Estimating daily global solar radiation by day of year [66] Gupta P, Srinivasan R. Missing data prediction and forecasting for water quantity
in China. Appl Energy 2010;87:3011–7. data. In: Proceedings of the International Conference on Modeling, Simulation and
[46] Zang H, Xu Q, Bian H. Generation of typical solar radiation data for different Control. Singapore: IACSIT Press; 2011.
climates of China. Energy 2012;38:236–48. [67] Huo J, Cox CD, Seaver WL, Robinson RB, Jiang Y. Application of two-directional
[47] Quej VH, Almorox J, Ibrakhimov M, Saito L. Estimating daily global solar radiation time series models to replace missing data. J Environ Eng 2010;136:435–43.
by day of the year in six cities located in Yucatan Peninsula, Mexico. J Clean Prod [68] Nzuza MB. Statistical modelling and estimation of solar radiation (M.Sc. Thesis).
2017;141:75–82. Durban, South Africa: University of Kwazulu-Natal; 2014.
[48] Khorasanizadeh H, Mohammadi K. Prediction of daily global solar radiation by day [69] Wilks DS. Statistical methods in the atmospheric sciences, 3rd ed.. Massachusetts:
of the year in four cities located in the sunny regions of Iran. Energy Convers Academic Press; 2011.
Manag 2013;76:385–92. [70] Allen R. Evaluation of procedures of estimating mean monthly solar radiation from
[49] Khorasanizadeh H, Mohammadi K, Jalilvand M. A statistical comparative study to air temperature. Technical Report. Rome:Food and Agriculture Organization of the
demonstrate the merit of day of the year-based models for estimation of horizontal United Nations (FAO); 1995.
global solar radiation. Energy Convers Manag 2014;87:37–47. [71] El-Sebaii AA, Al-Ghamdi AA, Al-Hazmi FS, Faidah A. Estimation of global solar
[50] Hassan GE, Youssef ME, Ali MA, Mohamed ZE, Shehata AI. Performance radiation on horizontal surfaces in Jeddah, Saudi Arabia. Energy Policy
assessment of different day-of-the-year-based models for estimating global solar 2009;37:3645–9.
radiation - Case study: Egypt. J Atmos Sol-Terr Phys 2016;149:69–80. [72] Ododo JC, Sulaiman AT, Aidan J, Yuguda MM, Ogbu FA. The importance of
[51] Olatomiwa L, Mekhilef S, Shamshirband S, Mohammadi K, Petkovic D, Sudheer C. maximum air temperature in the parameterisation of solar radiation in Nigeria.
A support vector machine-firefly algorithm-based model for global solar radiation Renew Energy 1995;6:751–63.
prediction. Sol Energy 2015;115:632–44.

1575

You might also like