Renewable and Sustainable Energy Reviews 70 (2017) 314–329

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews

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

A critical review of the models used to estimate solar radiation MARK

⁎
Jianyuan Zhang, Li Zhao , Shuai Deng, Weicong Xu, Ying Zhang
Key Laboratory of Efficient Utilization of Low and Medium Grade Energy (Tianjin University), Ministry of Education of China, Tianjin 300072, China

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

Keywords: Solar radiation data is critical to the design and operation of solar energy utilization systems, so a large number
Solar radiation of models have been proposed and developed to estimate solar radiation in the past ten years. However, the
Estimation performances of such models are controversial in different studies, and there is a lack of systematic comparison
Empirical model among them. In addition, few studies pay attention to the time scales and practicability of the models. This
Artificial neural network
paper focuses on solving these questions through a critical literature review and the authors believe it can
benefit researchers to perform further investigations about solar radiation. This paper reviews and compares the
models from the points of view of time scale and estimation type for the first time. Furthermore, a large amount
of data about the evaluation metrics (root mean square error and mean absolute percentage error) from
different studies is summarized to clarify the performances of proposed models. The questions arising from the
processing of source data are also carefully examined. This paper has presented a novel method to compare the
estimation models and has provided a detailed analysis on available models. The results indicate that the
sunshine duration fraction models and artificial neural networks have similar performances when used to
estimate monthly average daily global radiation and daily global radiation, while more work is needed to study
the estimation method on smaller time intervals and the mechanisms of atmospheric attenuation for solar
radiation.

1. Introduction
Solar energy is increasingly attractive in the 21st century since the
environmental problems caused by burning fossil fuels are becoming
severe. During the winter of 2015, the weather condition of smog has
covered most of the cities located in the middle and east of China, and
then caused great damage to human health and life. Another important
reason for valuing solar energy is the excessive consumption of the
fossil fuels with limited reserve. The shortage of fossil fuels is a
worldwide long-term challenge that needs us striving for renewable
and sustainable energy resources.
Solar radiation data is indispensable for designing and assessing
solar energy utilization technologies. Practically measured data is the
most accurate but not always readily available, which is mainly due to
the initial investment and maintenance cost of the measuring instru-
ments and relevant recorders. For instance, there are 756 meteorolo-
gical stations altogether in China, among which only 122 of them have
records of global solar radiation [1]. Consequently, estimating solar
radiation data by correlating it with other easily measured meteor-
ological parameters, such as sunshine duration, ambient temperature,
cloud cover, humidity, etc., is an alternative method to obtain desired
solar radiation data when there is no record of measurement. Many
models have been proposed and developed based on this viewpoint.
Some of them are concise mathematical formulas, which are conve-
nient for engineering uses and called empirical models. Some of them
are artificial intelligence techniques, such as artificial neural network,
support vector machine, genetic programming, etc. Meanwhile, there
are also many models utilizing satellite data and atmospheric char-
acteristics to estimate solar radiation data, such as Rayleigh scattering,
aerosol extinction, ozone absorption, etc.
In reality, the transient solar irradiance at any location, with the
unit of W/m2, keeps changing throughout the daytime, which mainly
attributes to the movement of the earth and the chaotic effect of the
atmosphere. There is barely any effective way can accurately estimate
or predicate transient solar irradiance except practical measurement.
The measurements performed in most of the official meteorological
stations are for horizontal surface on the time scale of an hour with the
unit of MJ/(m2hour), or a day with the unit of MJ/(m2day). These two
quantities represent accumulative energy during different time spans
instead of transient intensity.
In 1924, Angstrom [2] related monthly average daily global
radiation to average clear-sky daily global radiation at the location in
question and the sunshine duration fraction by using an empirical
correlation. Since the uncertainty in the definition of a clear day, Page

⁎
Corresponding author.
E-mail address: jons@tju.edu.cn (L. Zhao).

http://dx.doi.org/10.1016/j.rser.2016.11.124
Received 5 July 2016; Received in revised form 24 September 2016; Accepted 2 November 2016
Available online 26 November 2016
1364-0321/ © 2016 Elsevier Ltd. All rights reserved.
J. Zhang et al.
and others [3] have modified the correlation by replacing the clear-sky
radiation with average daily extraterrestrial radiation on horizontal
surface. In the last decades, Angstrom-Page model has become the
most typical empirical model in solar engineering and is cited vastly by
peers and researchers. The parameter of sunshine duration fraction in
this model is extremely important in estimating solar radiation. Many
subsequent studies have added different expressions of it, including
quadratic, cubic, square root, logarithmic, exponential, power, etc., to
Angstrom-Page model to create new empirical models, as shown in
[4,5]. With the development of computer technology, researchers
gradually utilize artificial intelligence techniques, statistical methods
and satellite-based methods to estimate solar radiation. The incipient
research field of solar meteorology has grown considerably by aggre-
gating diverse areas of knowledge. Currently, more research papers
about estimation of solar radiation keep emerging in the literature each
year. Accordingly, making comparisons between them and giving
recommendations are urgently necessary, which are exactly the main
purposes of this study.
There are already some review papers [4–11] existing in the
literature pool. Kadir [4] listed 60 empirical models developed to
estimate monthly average daily global solar radiation, in which many of
them had same mathematical expressions just with different regressive
constants. However, according to the results of many existing studies,
these constants totally depend on the locations in question. Fariba et al.
[5] did a similar work to [4], collecting 78 empirical models, in which
they were classified into four groups of sunshine-based models, cloud-
based models, temperature-based models and other meteorological
parameters-based models. Then they respectively selected several
models from each group to make a case study for Yazd, Iran, in the
end a sunshine-based model with exponential expression gave the best
performance. Amit et al. [7] reviewed a large number of studies that
had applied ANN (artificial neural network) to estimate solar radiation
on horizontal surface, and three studies that had applied ANN to
forecast solar radiation. However, they did not clearly consider the time
scales that the estimations were made on as well as the types of the
solar radiation. They concluded that ANN models were more accurate
than empirical models. Rajesh et al. [9] reviewed many studies using
linear empirical model and ANN models to estimate monthly average
daily global radiation. They also pointed out that the ANN models were
better than regression models. In addition, [10,12] and many research
papers also concluded that ANN models were better than empirical
models.
After examining many review and research papers, we have found
several noteworthy questions on estimation of solar radiation. In
addition, the author and his team have just finished building a
demonstration project of 200 kW solar power plant, and such questions
have been verified during the processes of design and operation. Some
of above questions are presented here. 1) Few studies pay attention to
the time scale of estimation when they compare the accuracy of models.
Is the time scale monthly average daily, daily or hourly? It has a great
impact on the performances of models and is vital for practical
applications. 2) Judgments about which model is the best neglect the
intrinsic quality of estimations. Such as in the estimation of monthly
average daily global radiation, the monthly average data inherently is
very coarse and uncertain to describe the solar radiation distribution,
thus, slight improvement of accuracy does not prove the advantage of
new developed models, especially when the new developed models are
much more complicated than conventional models. 3) Too much work
has been spent in collecting the expressions of existing models instead
of the values of evaluation metrics. It provides readers with numerous
models but cannot show how accurate they are. 4) This paper calls the
models' outputs as "estimation" rather than "prediction" or "forecast".
The reason is, these models use already measured parameters, such as
sunshine duration and air temperature, to recreate the solar radiation
condition in the same period. However, the ability of "prediction" or
"forecast" is to estimate the solar radiation condition in the coming
315
Renewable and Sustainable Energy Reviews 70 (2017) 314–329
periods, such as 3 h ahead and 12 h ahead. There are also many papers
[13–15] aiming at forecasting solar radiation condition of coming
period. Section 5 has briefly introduced two of them to distinguish the
difference between the terms of "estimation" and "prediction", but the
detailed investigations are out of this paper. This paper specifically
focuses on these aforementioned questions. In addition, we have
carefully examined the data reduction processes and the rationality
of the values of statistical indicators in existing studies. For example,
''monthly average'' commonly means that average the data over a
month of a year, but some researchers average the data over a month
first, and then they average the previous obtained data over many years
for the same month. Despite the latter can result in lower values of
estimation errors, the superiority and usability of the model are
questionable.
The papers reviewed in this study are all from influential journals
with high impact factors. The estimation is only for global solar
radiation on horizontal surface, which is the most basic data in the
field of measurement about solar radiation. It is noteworthy that the
daily radiation and monthly average daily radiation are not very precise
records for solar radiation. They can reflect the overall resource
situation about solar energy, but cannot give the detailed change of
the intensity. They are helpful to select the promising locations for
building solar power plants, but are not practical to control an actual
solar energy concentrating system. Two familiar statistical indicators
MAPE (mean absolute percentage error) and RMSE (root mean square
error) have been selected to compare different models, because these
two indicators appear most in existing studies. Besides, the former can
reflect the relative error while the latter can reflect the absolute error.
The combination of them can reasonably assess the performances of
models.
The rest of this paper is organized as follows. Section 2 first gives
the computations of several common parameters that are often used in
the models. Then the evaluation metrics and the method of reviewing
and comparing existing studies are explained. Section 3 and Section 4
respectively summarize the models used to estimate monthly average
daily global radiation and daily global radiation. Section 5 first
describes two studies about the forecast of solar radiation, and then
summarizes the models used to estimate hourly global radiation.
Section 6 briefly summarizes the comparison results about different
types of models and discusses the research gaps in this field.
2. Fundamental considerations
2.1. Basic parameters
The parameters of sunshine duration fraction, hourly and daily
extraterrestrial radiation on horizontal surface are important for the
estimation of solar radiation. It is necessary to figure out them before
building a model. Sunshine duration fraction is the ratio of actual
sunshine duration to maximum possible sunshine duration. Hourly
extraterrestrial radiation is the solar radiation intercepted by horizon-
tal surface during an hour without the atmosphere, and daily extra-
terrestrial radiation has similar definition.
The maximum possible sunshine duration S0 is
2
S0 = cos−1 (− tan ϕ⋅ tan δ )
15 (1)
⎛ 284 + n⎞
δ = 23.45 sin ⎜360 ⎟
⎝ 365 ⎠ (2)
where ϕ is the latitude, δ is the declination [16], n is the ordinal
number of the day in solar calendar which can be a non-integer.
Hourly extraterrestrial solar radiation on horizontal surface is
J. Zhang et al. Renewable and Sustainable Energy Reviews 70 (2017) 314–329

I0 =
12 × 3600Gsc
(1 + 0.033 cos
360n
) where the subscript “c” indicates calculated values, and the subscript
π 365
π (ω2 − ω1) “m” indicates measured values.
× [cos ϕ⋅ cos δ⋅(sin ω 2 − sin ω1) + 180
sin ϕ⋅ sin δ ] (3)
2.3. Methods
where ω1 and ω2 are the limit hour angle of an hour, in which ω2 is the
larger, all in degrees; Gsc is the solar constant, 1367 W/m2, [16]; I0 is
Through investigating the literature, the author has found that most
with the unit of J/(m2hour).
of related papers aim at estimating global solar radiation. It is largely
Daily extraterrestrial solar radiation on horizontal surface is
affected by the mode of recording solar radiation data. Commonly,
24 × 3600Gsc 360n meteorological stations only measure the global radiation and diffuse
H0 = π
(1 + 0.033 cos 365
)
πωs radiation on the time spans of an hour or a day, while the difference
× (cos ϕ⋅ cos δ⋅ sin ωs + sin ϕ⋅ sin δ ) (4)
180 between them is the beam radiation. Beam radiation refers to the solar
radiation received from the sun without having been scattered, whose
ωs = cos−1 (− tan ϕ⋅ tan δ ) (5)
direction is the same as the relative geometric position between the sun
where ωs is the sunset hour angle, positive, with the unit of degree; H0 and the earth. Diffuse radiation has no particular direction due to the
is with the unit of J/(m2day). scattering of atmosphere. Consequently, beam radiation is much more
I0 and H0 are obtained by integrating the product of normal difficult to measure because its direction keeps changing, so the
transient extraterrestrial irradiance and the incident angle of beam measuring equipment has to track the sun continuously. The pyran-
radiation to horizontal surface. In this process, the hour angle and the ometer measuring global solar radiation is modified to measure diffuse
real clock time are connected by the following correlation, radiation by adding a shading ring to eliminate the beam radiation.
Many studies have related the estimation of diffuse radiation with the
2π dω
= ratio of actual global radiation to corresponding extraterrestrial radia-
24 × 3600 dτ (6)
tion by using polynomials.
where ω is the hour angle, τ is the real time. The models reviewed in this paper are applicable to estimate global
Note that in these calculations, it has to convert the local clock time solar radiation on horizontal surface at ground level rather than tilted
we commonly use into the solar time. Solar time system bases on the surface. Their estimations are only for global radiation on bigger time
apparent angular motion of the sun across the sky, in which when the spans, such as an hour and a day, rather than a few minutes or seconds.
sun crosses the meridian of the observer (directly overhead) it is solar Actually, there is barely any effective way can accurately estimate
noon. The correction in minutes is given as [16] transient solar irradiance except practical measurement. Due to the
Tsolar − Tloc = 4(Lst − Lloc ) + E (7) random change of solar radiation every day, many studies have
averaged the daily and hourly radiation data over a month to catch
⎛ 0.000075 + 0.001868 cos B − 0.032077 sin B ⎞ the trend. The term of “monthly average daily global solar radiation” is
E = 229.2 ⎜ ⎟
⎝− 0.014615 cos 2B − 0.04089 sin 2B ⎠ (8) the mean value of the global solar radiation intercepted by horizontal
surface during a day over a month, and other similar terms have
360 analogous definitions unless new criteria are given.
B= (n − 1)
365 (9) The time span of the estimation has great impact on the perfor-
mance and utility of the model. From the viewpoint of data quality, a
where Tsolar and Tloc are the solar time and local clock time,
model proposed to estimate hourly radiation is more powerful and
respectively, in minutes; Lst and Lloc are the standard meridian for
useful than a model proposed to estimate daily radiation. We should
local time zone (for China, it is east longitude 120°) and the longitude
not just compare the values of MAPE and RMSE while neglect the
of the location, in degrees west; E takes into account the perturbation
effectiveness of the obtained output. Time span is the most important
of the earth's rotation. It is noteworthy that the maximum absolute
signage to classify the models. Section 3, Section 4 and Section 5 have
value of E throughout a year is less than twenty minutes.
respectively summarized the models used to estimate monthly average
daily global radiation, daily global radiation and hourly global radia-
2.2. Evaluation metrics
tion. In these sections, according to the modeling principles, the
models are divided into a few categories, such as empirical models
The accuracy of a model determines its value for use. Two familiar
and artificial neural networks. Then, to distinguish the differences
statistical indicators MAPE (mean absolute percentage error) and
between similar models clearly, they are categorized again, such as
RMSE (root mean square error) have been selected to compare
sunshine duration fraction models and modified sunshine duration
different models. MAPE represents the mean absolute percentage
fraction models. There is a brief summary in each group. Section 6 has
deviation between the estimated and measured values. RMSE can
summarized the conclusions and discussed the research gaps in this
provide the information about the short-term performance, by compar-
field.
ing the actual deviation one by one between the estimated and
In the process of classifying existing studies, the data reduction
measured values. Most of already published articles use them to
processes of them have been carefully examined, because sometimes
evaluate the models, at least one of them. Thus, they provide a
the actual processes are contradictory with the claims. For example,
benchmark to compare identical models and different models from
some studies claim that the estimation is for daily radiation, but
different studies. Meanwhile, the MAPE reflects the mean relative error
actually, it is monthly average daily radiation or long-term daily
of the estimation while the RMSE reflects the mean absolute error to
radiation, which makes the research results untenable. In addition, it
some degree, so the combination of them is reasonable. The smaller
is very difficult to find two or more different studies that have tested the
these two indicators are, the better the performance of the model is.
same models with same source data and evaluation metrics. Generally,
The mathematical expressions of them are
it is easier to find existing studies that have investigated identical
100%
N
yi, c − yi, m models or similar models for different locations. Consequently, the
MAPE = ∑ evaluation metrics may be different from each other. Another note-
N i =1
yi, m (10) worthy question is, many studies [17–24] have investigated different
N types of models in the same article, so it has to separate them and
∑i =1 (yi, c − yi, m )2 classify them into corresponding groups. This study has specially
RMSE =
N (11) focused on these questions, of course, the processing procedure is very

316
J. Zhang et al.
Fig. 1. The classification of reviewed models..
troublesome, but it is beneficial for solar engineering. Fig. 1 shows the
detailed classification about the models in this paper.
3. Monthly average daily global solar radiation
Daily global solar radiation, with the unit of MJ/(m2day), is the
solar energy intercepted by unit area of horizontal surface during a day.
Unit transformation can change it into average irradiance with the unit
of W/m2, but the latter may confuse readers, because it seems that the
data collection procedure is continuous and carried out many times.
Monthly average daily radiation is the average value of daily radiation
over a month. That is to say, the change of solar radiation in a year is
represented by twelve values. It is not precise record of solar radiation.
It can reflect the potential of solar energy at a site of interest and is
useful for sketchy design work, but is not practical for detailed system
control. Accordingly, slight improvement of accuracy does not prove
the advantage of new developed models, especially when the new
models are much more complicated than conventional models. For a
place with monthly average daily global radiation varying around
18 MJ/(m2day), such as Jeddah, Saudi Arabia [17], a model with the
RMSE of 1 MJ/(m2day) roughly has the same effectiveness as a model
with the RMSE of 2 MJ/(m2day). This question is very prominent if the
models have applied similar or same principles, and there is no in-
depth theoretical analysis about them.
Another issue that needs to be noted is the process of data
reduction in existing studies. Some studies first average the daily data
over a certain month, and then they average the already obtained data
over several years for the same month, so the condition of solar
radiation at a site in several years is represented by twelve values.
Obviously, this type of data is much vaguer than monthly average daily
data, thus, here we nominate it as "long-term monthly average daily
data".
3.1. Empirical models
An empirical model correlates solar radiation with other easily
measurable parameters, such as sunshine duration, temperature and
humidity, by applying concise mathematical functions. Due to its
simplicity and strong operability, the empirical model is much more
convenient for engineering applications. After investigating a large
number of studies, the author has found that sunshine duration
317
Renewable and Sustainable Energy Reviews 70 (2017) 314–329
fraction has the biggest influence on the accuracy, followed by ambient
temperature. Based on the previous findings, these empirical models
are divided into three groups. The first group is SDF (sunshine
duration fraction) models, in which only the sunshine duration fraction
is employed to estimate solar radiation. The second group is MSDF
(modified sunshine duration fraction) models, in which some other
parameters are added to SDF models, such as temperature, humidity,
cloud cover, altitude, etc. The third group is NSD (non-sunshine
duration) models, in which the models do not use the sunshine
duration but other parameters as inputs, mainly the ambient tempera-
ture.
3.1.1. Sunshine duration fraction models
The first SDF model was proposed by Angstrom [2], which related
the clear sky index with sunshine duration fraction by using linear
equation. The mathematical expression is
G S
=a+b
Gc S0 (12)
where G is the actual monthly average daily global radiation, Gc is the
average clear-day radiation, S is the actual monthly average sunshine
duration, S0 is the monthly average of maximum possible sunshine
duration, a, b are empirical constants.
With available measured data, the method of linear regression is
applicable to determine the empirical constants. Due to the uncertainty
in the definition of a clear day, Page and others [3] have modified the
correlation by replacing the clear-day radiation with extraterrestrial
radiation on horizontal surface, which is given as
G S
=a+b
G0 S0 (13)
where G0 is the monthly average daily extraterrestrial radiation on
horizontal surface.
After that, many researchers have done a large number of mod-
ifications on this model to improve the accuracy. Several typical models
from [25–32] are given below,
G / G0 = a + b (S / S0 ) + c (S /S0 )2 (14)
G / G0 = a + b (S / S0 ) + c (S /S0 )2 + d (S /S0 )3 (15)
G / G0 = a + b log(S / S0 ) (16)
J. Zhang et al. Renewable and Sustainable Energy Reviews 70 (2017) 314–329

G / G0 = a + b (S / S0 ) + c log(S / S0 ) (17) was used to test the models. They commented that in comparison to
linear model, quadratic and cubic models did not significantly improve
G / G0 = a + b exp(S / S0 ) (18) the accuracy, but the linear model required less computational work.
The conclusion was consistent with [25]. The RMSE of the linear model
G / G0 = a + b (S / S0 ) + c exp(S / S0 ) (19)
for the four locations ranged from 0.43 to 0.8 MJ/(m2day). Li et al.
G / G0 = a + b (S / S0 )c (20) [27] evaluated 8 SDF models for 4 stations in Tibet, China, which were
given as Eq. (13) to Eq. (20). 11 years of data was used for calibration
S S S S and 4 years of data was used for validation. They commented that
G / G0 = a + b cos(c ) + d sin(c ) + e cos(2c ) + f sin(2c )
S0 S0 S0 S0 (21) linear model was reasonably accurate in practice and complex mod-
ifications were not necessary. For the four stations, the RMSE of the
where a, b, c, d, e, f are empirical constants.
eight models ranged from 0.7 to 1.33 MJ/(m2day), and the RMSE of
Of course, there are many models with different mathematical
linear model ranged from 0.72 to 1.26 MJ/(m2day). Behrang et al. [28]
expressions, see in [4,5,8,33], but the basic principles of them are the
investigated 11 models for 17 cities of Iran by using particle swarm
same, thus they are not listed here to condense this paper. The models
optimization technique to determine empirical coefficients instead of
given above can clearly show the development about SDF model: more
least square method. Eight of them were given above and the other
functions are introduced to improve the accuracy, while the basic
three models were similar to Eq. (21). The results clearly showed that,
principle is, through existing data to obtain empirical coefficients and
for an identical location the eleven models had extremely similar
then apply them for locations with no measured solar radiation data.
accuracy, but for different locations they changed largely. For instance,
However, a basic difficulty lies in empirical models: the empirical
for the city of Bandarabass, the MAPE of the eleven models ranged
coefficients have location dependence, which has been confirmed by
from 6.13% to 6.19%. However, for the city of Zanjan, the MAPE
many researchers, see in [4,5,8,33]. Location dependence refers to that
ranged from 15.28% to 16.76%. Hacer et al. [29] studied 5 SDF models
the coefficients derived from one location may be not suitable for
to estimate long-term monthly average daily radiation for 7 meteor-
another location. It is easy to explain, two different locations may have
ological stations in Turkey. Overall speaking, the accuracies of them
different latitudes, altitudes, topographical characteristics and so on,
were similar for the same location, which was consistent with [25–28].
while the condition of solar radiation is heavily affected by these
For the station of Bitlis, the MAPE of them ranged from 3.55% to
factors. Before using derived coefficients for another place, we have to
4.16%. However, the corresponding RMSE values were every small,
ensure the treatment is reasonable, such as the two locations have
ranging from 0.019 to 0.023 MJ/(m2day), which was largely different
similar meteorological conditions. In many studies, the empirical
from other investigations. Yao et al. [33] compared 89 existing monthly
coefficients are the same for a year, while there are also some studies
average daily radiation models and 19 existing daily global radiation
deriving coefficients for different months of a year respectively, see in
models for Shanghai, China, in which the empirical coefficients were
[34,35]. Obviously, this manipulation as well as the addition of
the same as original papers instead of new fitting coefficients for
complex functions all aim at improving the accuracy. Accordingly,
Shanghai. Thus, many models had same mathematical expressions just
there is a trade-off between the accuracy and the complex computa-
with different coefficients. We do not approve of this comparison
tional work.
because the empirical coefficients are location-dependent, and uncon-
Several typical studies are described here and the results of them
ditional utilizations are not appropriate. After that, they also derived
are summarized in Table 1. Kadir [25] investigated 7 SDF models for
new fitting coefficients for 5 SDF models for Shanghai. The RMSE of
Turkey, including the models of linear (Eq. (13)), quadratic (Eq. (14)),
them ranged from 1.54 to 1.77 MJ/(m2day) and the MAPE ranged
logarithmic (Eq. (16), Eq. (17)), exponential (Eq. (18), Eq. (19)) and
from 5.65% to 7.0%. Kasra et al. [36] tested 4 SDF models for Isfahan,
power (Eq. (20) with a=0). From the figures and text information, it
Iran. 9 years of data was used for training and 4 years of data was used
seems that the estimation was for long-term monthly average daily
for testing. The RMSE of them ranged from 1.10 to 1.18 MJ/(m2day)
global solar radiation. The measured data from 18 locations in Turkey
and the MAPE ranged from 4.15% to 4.19%. Manzano et al. [37]
for 24 years was used for training and validation. The empirical
evaluated the linear Angstrom–Prescott model for 25 stations in Spain.
coefficients were separately derived for each location. The perfor-
More than 10 years of data was used for calibration and 2 years of data
mances of the seven models had only slight difference for the same
was used for validation. The RMSE ranged from 0.36 to 0.80 MJ/
location, but they changed largely for different locations. Except three
(m2day) except four stations whose RMSE equaled 1.43, 1.34, 1.17,
locations (3.65, 2.45, 2.19 MJ/(m2day)), the RMSE of the linear model
1.09, respectively. Ahmet et al. [30] investigated linear, quadratic and
for the rest of locations ranged from 0.31 to 1.79 MJ/(m2day), and the
cubic empirical models for 4 cities of Turkey. Park et al. [31] studied
corresponding MAPE ranged from 1.99% to 12.58%. Katiyar et al. [26]
linear empirical model for 22 stations in South Korea. Chelbi et al. [32]
compared the linear, quadratic and cubic models to estimate long-term
investigated 5 empirical models for 4 stations in Tunisia.
monthly average daily radiation for India. 5 years of data from 4 cities

Table 1
The comparison of SDF models for monthly average daily radiation.

Reference Year Models Estimation period Location number RMSE MAPE

Zhou et al. [19] 2005 1 long-term 69 0.72 to 1.59 3.64% to 11.32%

El-Sebaii et al. [17,18] 2009 2 long-term 1.331 0.04, 0.05 —
Kadir [25] 2009 7 long-term 18 0.31 to 1.79 1.99% to 12.58%
Katiyar et al. [26] 2010 3 long-term 4 0.43 to 0.8 —
Li et al. [27] 2011 8 monthly 4 0.7 to 1.33 —
Behrang et al. [28] 2011 11 monthly 17 — 6.13% to 6.19%, 15.28% to 16.76%
Hacer et al. [29] 2012 5 long-term 7 0.019 to 0.023 3.55% to 4.16%
Fariba et al. [5] 2013 1/78 monthly 1 0.54 —
Yao et al. [33] 2014 5 monthly 1 1.54 to 1.77 5.67% to 7.0%
Kasra et al. [36] 2015 4 monthly 1 1.10 to 1.18 4.15% to 4.19%
Manzano et al. [37] 2015 linear monthly 25 0.36 to 0.80 —
Jamshid et al. [12] 2015 linear monthly 2 2.05, 1.14 —
Shahaboddin et al. [20] 2015 2 monthly 1 1.39, 1.30 5.17%, 5.04%

318
J. Zhang et al.
Table 2
The comparison of MSDF models for monthly average daily radiation.
Reference Year Models Added Parameters
[19] 2005 6 φ, H
[17,18] 2009 3 Tave, R, cloud cover
[21] 2013 3 Tave, R
[5] 2013 1/78 ΔT
[20] 2015 2 Tave, Tmax, R
Where Tave is the average ambient temperature, Tmax is the maximum ambient temperature, Tmin is the minimum ambient temperature, ΔT is the difference between Tmax and Tmin,
H is the altitude, φ is the latitude, R is the relative humidity.
Table 1 has summarized the results of above studies. We can see
that many forms of SDF models are employed to estimate long-term
monthly average and monthly average daily global radiation for
different locations. The most important feature is, the performances
of different SDF models are similar for the same location, but they
change largely for different locations. This feature can also be verified
by comparing the results from different studies, such as [37] and [12].
Some conclusions about SDF models are given here. 1) The SDF
models can reasonably estimate the long-term monthly average and
monthly average daily global radiation. 2) The linear model and the
models with complex modifications have similar accuracy for the same
location, but the accuracy changes largely for different locations. 3) The
linear empirical model is often recommended for use because it
requires less computational work but has satisfied accuracy. 4) On
the whole, the RMSE values of the SDF models used to estimate
monthly average daily radiation are in the range of 0.36–2.05 MJ/
(m2day).
3.1.2. Modified sunshine duration fraction models
Some studies add other meteorological parameters, such as ambi-
ent temperature, relative humidity and cloud cover, to SDF models to
improve the accuracy of estimation. Sometimes they also get the
geographical parameters involved, such as longitude, latitude and
altitude. This paper nominates these models as MSDF (modified
sunshine duration fraction) models. Several typical MSDF models from
[17] are given here,
G S
= a + b ( ) + cTave
G0 S0
G S
= a + b ( ) + cRh
G0 S0
G S
= a + b ( ) + cCw
G0 S0
where Tave is the average ambient temperature, Rh is the relative
humidity, Cw is the cloud cover.
Zhou et al. [19] studied 6 MSDF models and 3 SDF models to
estimate long-term monthly average daily global solar radiation for 69
meteorological stations in China. The parameters added in modified
models were latitude and altitude. The coefficients were separately
derived for the 69 stations. Except four stations (2.23, 2.28, 1.91, 1.74),
the RMSE of the recommended cubic SDF model ranged from 0.72 to
1.59 MJ/(m2day), and the corresponding MAPE ranged from 3.64% to
11.32%. They also developed 9 general models with fixed coefficients
for overall stations. The RMSE of the six MSDF models ranged from
1.404 to 1.414 MJ/(m2day) and the MAPE ranged from 8.05% to
8.20%. The RMSE of the three SDF models ranged from 1.634 to
1.636 MJ/(m2day) and the MAPE ranged from 9.21% to 9.23%. El-
Sebaii et al. [17,18] applied 3 MSDF models, 3 SDF models and 3 NSD
models to estimate long-term monthly average daily global solar
radiation for Jeddah, Saudi. The parameters added in MSDF models
were temperature, relative humidity and cloud cover. 9 years of data
was employed to derive new empirical coefficients, and the RMSE of
the nine models in this stage ranged from 0.02 to 0.15 MJ/(m2day),
Renewable and Sustainable Energy Reviews 70 (2017) 314–329
Estimation period Location number RMSE MAPE
long-term 69 1.40 to 1.41 8.05% to 8.20%
long-term 1 0.02, 0.03, 0.08 —
long-term 6 0.51 to 0.60, 2.16 to 2.26 —
monthly 1 0.85 —
monthly 1 1.54, 1.55 5.92%, 6.06%
which somehow showed a great improvement compared with other
investigations. Khorasanizadeh et al. [21] evaluated 3 MSDF models, 5
SDF models and 3 NSD models to estimate long-term monthly average
daily global solar radiation for six major cities of Iran. The parameters
added in MSDF models were temperature and relative humidity. The
MSDF models had shown slight better performance for two cities, while
for other cities all the models had similar accuracy. Overall, the change
of the accuracy of the eleven models between different locations is
much bigger than that for an identical location. For the city of Isfahan,
the RMSE of all models ranged from 0.47 to 0.82 MJ/(m2day). But for
Karaj, the RMSE of ten models ranged from 2.16 to 2.51, and the other
one was 3.16 MJ/(m2day). Compared with SDF models, the MSDF
models did not show clear improvement and the non-sunshine models
did not show much degradation. Shahaboddin et al. [20] evaluated 2
MSDF models, 2 SDF models and 1 NSD model for Shiraz, Iran. 7 years
of data was used for training and 3 years of data was used for testing.
The RMSE of the five models ranged from 1.30 to 1.55 MJ/(m2day)
and the MAPE ranged from 5.04% to 6.80%. Antonio et al. [38]
established an linear equation to correlate solar radiation with the
product of sunshine duration and daily temperature variation by
analyzing the energy balance between soil layer and adjacent atmo-
sphere layer.
Table 2 has summarized the results of above studies. We can see
that few studies use MSDF models to estimate monthly average daily
global radiation, while most of them are for long-term monthly average
daily radiation. Compared with SDF models, the addition of other
parameters does not improve the performance. Of course, this com-
parison is not sufficient due to the scarcity of studies. As we already
(22)
know, the parameter of sunshine duration has strong relation with the
solar radiation condition, while the difference between the maximum
(23) and minimum ambient temperature in a day has less influence. Thus,
the addition of other less influential parameters may degenerate the
performances of SDF models. Some conclusions about MSDF models
(24) are given here. 1) MSDF models are often used to estimate long-term
monthly average daily radiation instead of monthly average daily
radiation. 2) Comparing with SDF models, the addition of more
parameters does not improve the performance.
3.1.3. Non-sunshine duration models
The parameter of sunshine duration is often used to estimate solar
radiation. When the record data of sunshine duration is not available,
the models without utilizing sunshine duration are expected. These
models are termed NSD models (non-sunshine duration models) and
this section will discuss their performances. Some typical models are
given here,
G
= a Tmax − Tmin
G0 (25)
G
= 1 − exp(−bΔT c )
G0 (26)
G T
= a + b min
G0 Tmax (27)
319
J. Zhang et al.
G T R explanation about the question. In terms of configuration, an ANN
= a + b min ×
G0 Tmax 100 (28)
where Tmax is the maximum ambient temperature, Tmin is the
minimum ambient temperature, ΔT is the difference between Tmax
and Tmin, R is the relative humidity.
Adaramola [39] studied the Angstrom-Page model and 6 NSD
models to estimate long-term monthly average daily global solar
radiation for Akure, Nigeria. Ambient temperature, relative humidity
and precipitation were used in NSD models. The RMSE of the linear
model was 0.93 MJ/(m2day), while the RMSE of other models ranged
from 1.00 to 1.61 MJ/(m2day). The MAPE of the linear model was
3.62%, while the RMSE of other models ranged from 4.78% to 8.25%.
Iranna et al. [40] established 16 NSD models to estimate monthly
average daily clearness. Relative humidity, wind speed, moisture,
longitude, altitude and five other temperature related parameters were
used as inputs. The data from 875 stations around the world was
collected to test the models. Fariba et al. [5] tested the Hargreaves
model (Eq. (25)) and a cloud-based model for Yazd, Iran. 16 years of
data was used to develop empirical constants and 5 years of data was
used to validate the models. The RMSE of the Hargreaves model was
0.71 MJ/(m2day), while the RMSE of the cloud-based model was
1.12 MJ/(m2day) (Table 3).
There are few studies using NSD models to estimate monthly
average daily radiation. The Eq. (25) was tested by several studies
[5,12,20] and the performance was acceptable.
3.2. Artificial neural networks
Before introducing the ANN (artificial neural network), some
details about the method of comparing ANN models must be given
first. An ANN model is much more complex than an empirical model in
terms of the computational work. There are many variables in an ANN
model, including the input, the number of layers and neurons, training
algorithm, transfer function, etc. Any modification of them can create a
new ANN model, so it is necessary to develop appropriate rules to
compare these ANN models. The first thing should be considered is the
output of an ANN model. The comparison has no meaning if the output
is different, and that is also why we classify the sections as monthly
average daily radiation, daily radiation and hourly radiation. The
second thing should be considered is the input of an ANN model. In
the process of estimation, researchers can choose different ways to
achieve the same purpose. The selection of method is human-deter-
mined, but the source input data is physically required. More input
parameter means that the model needs more prerequisites to run, and
the data collection process will also become more complex.
Consequently, the following comparisons have focused on the accuracy,
inputs and outputs of models.
Artificial neural network is a numerical modeling technique that
develops most in recent years. It is inspired by biological neural system
and is capable of processing non-linear relationship, data sorting,
pattern detection, optimization, clustering and simulation. It is called
"black box" modeling technique because it does not present physical
Table 3
The comparison of NSD models for monthly average daily radiation.
Reference Year Models Input parameters Estimation period
[17,18] 2009 3 ΔT, Tave, R, cloud cover long-term
[39] 2012 6 ΔT, Tave, Tmax/Tmin, long-term
R, precipitation
[21] 2013 3 ΔT, Tave, R long-term
[5] 2013 2 ΔT, cloud cover monthly
[20] 2015 1 ΔT monthly
[12] 2015 1 ΔT monthly
Where Tave is the average ambient temperature, Tmax is the maximum ambient temperature, Tmin is the minimum ambient temperature, ΔT is the difference between Tmax and Tmin,
R is the relative humidity.
320
Renewable and Sustainable Energy Reviews 70 (2017) 314–329
model usually contains input layer, hidden layer and output layer. In
terms of the process of data manipulation, it mainly consists of two
stages: training section and testing section. In training section, the
ANN finishes learning and storing the pattern information of existing
database. In testing section, the ANN recalls the information to
produce output data based on particular input database. The calcula-
tion units of ANN are interconnected neurons in the layers. The left
side of Fig. 2 shows the computational rule of a neuron, whose
mathematical expression is
n
y = g ( ∑ wi xi )
i =1 (29)
where xi is the input parameter, wi is the weight for xi, g(·) is the
transfer function. The transfer function can be linear, sigmoid and
hyperbolic. The right side of Fig. 2 shows a multiplayer perceptron
network with three layers, whose mathematical expression is
Y = f (X , W ) (30)
where X represents the input vector, Y represents the output vector, W
represents the vector of connection weights, f(·) denotes the functional
relationship between input vector and output vector. In the training
stage, with available database, the ANN keeps adjusting the weights in
order to minimize the difference between outputs and observations.
Then the determined weights can be used for estimation with new
inputs. In an ANN developed to estimate solar radiation, the inputs are
usually meteorological and geographical parameters, while the outputs
are the quantities of solar radiation on different time spans. There is a
detailed description in [41].
Senkal et al. [42] studied ANN model for 12 cities in Turkey. The
inputs were latitude, longitude, altitude, month, mean diffuse radiation
and mean beam radiation. 5 months of data from 9 stations was used to
train the ANN, while 5 months of data from other 3 stations was used
for testing. They also proposed a satellite-based method to estimate the
monthly average daily radiation. The RMSE of ANN and satellite-based
method in training stage were 2.32 and 2.75 MJ/(m2day) respectively,
and in testing stage were 3.94 and 5.37 MJ/(m2day) respectively. From
previous findings of empirical models, it is easy to know that this
ANN's performance is not good, which may be due to that the source
data is not enough. Another reason may be that the stations used for
testing are not the stations used for training. Senkal [43] built an ANN
model using latitude, longitude, altitude, two types of surface emissiv-
ity and land surface temperature as inputs. The last three parameters
were obtained from satellite data. One year of data from 10 stations
was used to train the ANN and one year of data from other 9 stations
was used for testing. The RMSE in training and testing stage were 0.16
and 0.32 MJ/(m2day), respectively. Clearly, the performance becomes
much better than [42], but the second ANN model needs satellite data
as source data. Qin et al. [41] built a feed-forward ANN with single
hidden layer, which used 6 remote sensing products as inputs,
including difference of land surface temperature between daytime
and nighttime, mean land surface temperature, monthly precipitation,
Location number RMSE MAPE
1 0.02, 0.11, 0.15 —
1 1.00 to 1.61 4.78% to 8.25%
6 0.70 to 0.82, 2.39 to 3.16 —
1 0.71, 1.12 —
1 1.50 6.82%
2 2.43, 1.28 —
J. Zhang et al.
Fig. 2. Schematic diagram of a neuron unit (left side) and an artificial neural network (right side)..
enhanced vegetation index, number of days and air pressure ratio. They
selected Levenberg-Marquardt algorithm coupled with Bayesian reg-
ulation to train the ANN, and used hyperbolic tangent sigmoid function
as the transfer function in hidden layer. There were 12 neurons in the
hidden layer. 7 years of data from 22 stations on Tibetan Plateau was
used to train the ANN and 5 years of data from other 12 stations was
used for testing. The RMSE in testing stage was 1.40 MJ/(m2day) and
the relative RMSE is 8.47%. Ozgoren et al. [44] developed a feed-
forward ANN for Turkey. They applied the method of stepwise multi-
nonlinear regression to determine the most suitable independent
variables for the input layer. Ten variables were selected, including
sunshine duration, month of the year, cloudiness, soil temperature,
mean atmospheric temperature, altitude, wind speed, maximum atmo-
spheric temperature, minimum atmospheric temperature and latitude.
There were 10 neurons in the hidden layer. Levenberg-Marquardt
optimization algorithm was used to train the ANN. 7 years of data from
27 stations was used for training while 7 years of data from other 4
stations was used for testing. The MAPE in training and testing stage
were 4.29% and 5.34% respectively. Mohamed [45] used particle
swarm optimization to train the ANN developed for Saudi Arabia.
The inputs were month of the year, latitude, longitude, altitude and
sunshine duration. However, the estimation was for long-term monthly
average daily global solar radiation. The data from 31 stations was used
to train the ANN while the data from other 10 stations was used for
testing. The average MAPE is 8.85%. Yadav et al. [46,47] applied the
software of Waikato Environment for Knowledge Analysis to find the
most influencing input parameters for estimation. They identified the
average temperature, maximum temperature, minimum temperature,
altitude and sunshine duration as the most relevant input variables,
while latitude and longitude were the least influencing variables.
However, the estimation was also for long-term monthly average daily
global solar radiation. An ANN with 10 neurons in hidden layer
achieved the best performance, and the maximum MAPE was 6.89%.
Jiang et al. [48] added a hard-ridge penalty to RBF (radial basis
function) to reduce the number of nodes in the hidden layer. They
applied Pearson correlation coefficients and Apriori association rules to
select the relevant input variables. 12 parameters were selected,
including average total opaque sky cover, opaque sky cover, precipita-
tion, broadband aerosol optical depth, maximum temperature, mini-
mum temperature, average temperature, daylight temperature, relative
humidity, heating degree days, cooling degree days and wind speed. 11
years of data from 4 sites in the United States was used to train the
ANN while 2 years of data was used for testing. The CS-hard-ridge-RBF
model achieved the best performance, and the MAPE of the four sites
ranged from 5.44% to 7.64% and the RMSE ranged from 1.01 to
1.42 MJ/(m2day). Shahaboddin et al. [20] applied the algorithm of
ELM (extreme learning machine) to train a single-layer feed-forward
neural network, and compared it with SVM (support vector machine),
GP (genetic programming) and ANN (artificial neural network). The
inputs were three combinations of the parameters of sunshine duration
fraction, air temperature difference, average air temperature and
321
Renewable and Sustainable Energy Reviews 70 (2017) 314–329
relative humidity. 7 years of data from Shiraz, Iran was used to train
the models while 3 years of data was used for testing. The models with
four inputs achieved the best performances. The RMSE of ELM was
0.59 MJ/(m2day) and the MAPE was 2.25%. The RMSE of the other
three models ranged from 0.86 to 0.93 MJ/(m2day), and the MAPE
ranged from 3.38% to 4.15%. Lanre et al. [49] studied ANFIS (adaptive
neuro-fuzzy inference system) for Iseyin, Nigeria. ANFIS was a hybrid
intelligent system that merges the learning power of ANN with the
knowledge representation of fuzzy logic. The inputs were maximum
temperature, minimum temperature and sunshine duration. 15 years
of data was used to train the model while 6 years of data was used for
testing. The RMSE in training and testing stages were 1.09 and
1.76 MJ/(m2day), respectively.
Table 4 has summarized the results of above studies. The study of
[42] has poor performance even it has used the mean diffuse and beam
radiation as inputs instead of the sunshine duration and temperature.
They apply the parameter of surface emissivity as input in [43] and the
performances become much better. However, the satellite data is not
usually available in many locations. As to other studies, the perfor-
mances of them do not change largely each other. They do not show
clear improvement than SDF empirical models, but the suitability to
untrained stations is good. Some conclusions about the ANN models
are given here. 1) ANN models can reasonably estimate the monthly
average daily radiation. 2) Many new parameters are introduced as
inputs to the ANN models to estimate solar radiation, and many
optimization algorithms are applied to filter the input parameters. 3)
Researchers keep modifying the ANN models to improve the accuracy,
but the effect is not clear when compared with empirical models. 4) On
the whole, the RMSE values of the ANN models are in the range of
0.32–2.05 MJ/(m2day).
3.3. Other techniques
Support vector machine is a well-known machine learning approach
that has been applied in variety of fields such as computing, hydrology
and environmental researches. It shows superior performance to
neural network and other conventional statistical models.
Mohammadi et al. [22] studied a hybrid approach that combined
SVM (support vector machine) with WT (wavelet transform algorithm)
for an coastal city Bandar Abass, Iran. They compared the hybrid
model with ANN, GP (genetic programming) and ARMA (autoregres-
sive moving average). 10 years of data was used to train the models
while 4 years of data was used to test. The four models achieved the
best performance by applying the parameters of sunshine duration
fraction, difference between maximum and minimum ambient tem-
peratures, relative humidity, water vapor pressure, average ambient
temperature and extraterrestrial global solar radiation as inputs. The
RMSE of the SVM-WT model was 0.66 MJ/(m2day), while the RMSE
of ANN, GP and ARMA were 1.81, 1.79 and 1.83 MJ/(m2day),
respectively. The MAPE of the WVM-WT model was 3.26%, while the
MAPE of ANN, GP and ARMA were 8.68%, 8.55% and 8.76%,
J. Zhang et al. Renewable and Sustainable Energy Reviews 70 (2017) 314–329

Table 4
The comparison of ANN models for monthly average daily radiation.

Reference Year Input parameters Output RMSE MAPE

Senkal et al. [42] 2009 N,φ, L, H, mean diffuse radiation, mean beam radiation monthly 3.94 —
Senkal [43] 2010 φ, L, H, two types of surface emissivity, land surface temperature monthly 0.32 —
Qin et al. [41] 2011 two types of land surface temperature, precipitation, enhanced vegetation index, number of days monthly 1.4 (8.47%) —
and air pressure ratio
Ozgoren et al. [44] 2012 S, N, cloudiness, soil temperature, etc., 10 parameters monthly — 5.34%
Mohamed [45] 2012 S, N, φ, L, H long-term — 8.85%
Yadav et al. [46,47] 2014 S, Tave, Tmax, Tmin, H long-term — 6.89%
Jiang et al. [48] 2015 average total opaque sky cover, opaque sky cover, precipitation, etc., 12 parameters monthly 1.01 to 1.42 5.44% to 7.64%
Mohammadi et al. [22] 2015 S/ S0, G0, Tave, ΔT, humidity, water vapor pressure monthly 1.81 8.68%
Mohammadi et al. [23] 2015 S, Tmax, Tmin monthly 2.05 13.43%
Shahaboddin et al. [20] 2015 ELM-ANN: S/ S0, ΔT, Tave, humidity monthly 0.59 2.25%
ANN: S/ S0, ΔT, Tave, humidity monthly 0.93 4.15%
Lanre et al.[49] 2015 S, Tmax, Tmin monthly 1.76 —

Where N is the ordinal number of the month of interest in a year, G0 is the extraterrestrial radiation on horizontal surface, S is the actual sunshine duration, S0 is the maximum possible
sunshine duration, Tave is the average ambient temperature, Tmax is the maximum ambient temperature, Tmin is the minimum ambient temperature, ΔT is the difference between
Tmax and Tmin, L is the longitude, H is the altitude, φ is the latitude.

Table 5 2 stations in China. 20 years of data was used to train the models while
The comparison of other techniques for monthly average daily radiation. 10 years of data was used for testing. The data used to train the models
was only the monthly average daily radiation data. The best perfor-
Methods Reference RMSE MAPE
mance achieved by the models was: for Urumqi, RMSE equals
Artificial neural network Mohammadi et al. [22] 1.81 8.68% 1.44 MJ/(m2day); for Beijing, RMSE equals 1.56 MJ/(m2day).
Mohammadi et al. [23] 2.05 13.43% Ayodele et al. [51] applied an probability function to represent the
distribution of clearness index in a year. The coefficients were
Support vector machine Mohammadi et al. [22] 0.66 3.26%
Mohammadi et al. [23] 1.87 11.52%
determined by using 7 years daily global solar radiation data. After
Mohammadi et al. [36] 0.45 1.41% that, they inversely transformed the probability function to get the
Shahaboddin et al. 0.86 3.4% estimation of clearness index. There was no other input for the model,
[20] and from its theory, we surmised that the estimation is for long-term
monthly average global solar radiation. Except the month of October,
Genetic programming Mohammadi et al. [22] 1.79 8.55%
Mohammadi et al. [23] 1.95 13.21% the performances of other months were excellent. The average RMSE
Shahaboddin et al. 0.86 3.38% was 0.29 MJ/(m2day) and the average MAPE was 2.71%. They also
[20] evaluated 3 empirical models and derived the coefficients respectively
for each month. Overall, the performances of these four models were
Autoregressive moving Mohammadi et al. [22] 1.83 8.76%
average
similar, such as for January, the RMSE of them ranged from 0.213 to
Sun et al. [50] 1.44, 1.56 — 0.221 MJ/(m2day).
Ayodele et al. [51] 0.29, long- 2.71% Janjai et al. [52] investigated a satellite-based model for 5 stations
term in Cambodian and 4 stations in Thailand. They utilized the visible
channel data from three geostationary satellites to get the earth-
Satellite-based method Janjai et al. [52] 1.13 —
Senkal et al. [42] 2.75, 5.37 — atmospheric reflectivity, which was then related to absorption and
scattering coefficients of various atmospheric constituents. The RMSE
of this model was 1.13 MJ/(m2day).
respectively. In [23], they investigated another hybrid approach which Table 5 has summarized the results of above studies. We can see
combined the SVW with FFA (firefly algorithm) for three sites of that many new approaches are introduced to estimate monthly average
Nigeria. The inputs were sunshine duration, maximum temperature daily radiation in recent years. Mohammadi et al. [22,23] have
and minimum temperature. They also compared the SVW-FFA model elaborately compared the methods of artificial neural network, support
with GP model and ANN model. 17 years of data was used to train the vector machine, genetic programming and autoregressive moving
models and 4 years of data was used for testing. The average RMSE of average. It seems that the support vector machine is better but the
SVW-FFA, ANN and GP in testing stage were 1.87, 2.05 and 1.95 MJ/ performance is not stable, while other techniques have similar accu-
(m2day), respectively. The corresponding MAPE were 11.52%, 13.43% racy. On the whole, considering the essential error of the estimations,
and 13.21%, respectively. In [36], they studied SVR for Isfahan, Iran. the results shown in Table 5 are consistent, and these new approaches
Two kernel functions of radial basis function and polynomial basis and the ANN models are at the same level of accuracy.
function were tested. The inputs were sunshine duration and maximum
possible sunshine duration. The SVR-rbf showed better performance,
while its RMSE and MAPE were 0.45 MJ/(m2day) and 1.41%, respec- 4. Daily global solar radiation
tively.
Autoregressive moving model assumes the values of estimation as a Daily global solar radiation is much more accurate than monthly
random sequence that only changes with time, which is also called time average daily radiation, because it has recorded the change of solar
series model. When this model is used to estimate solar radiation, it radiation day by day. However, it must note that some studies have
means that the actual meteorological condition will not be considered. replaced the daily radiation with long-term daily radiation in the
Sun et al. [50] evaluated the effectiveness of ARMA (autoregressive process of data reduction. The difference between them is, the latter
moving average) model coupled with various generalized autoregres- is the average value of the recorded daily radiation data for a particular
sive conditional heteroskedasticity processes to estimate solar radia- day. For example, there are seven years of data available for the day of
tion. They analyzed 6 different GARCH approaches with the data from March 1, while the average of the seven values is the long-term daily
radiation, so it cannot reflect the condition of an actual day. Also, if use

322
J. Zhang et al. Renewable and Sustainable Energy Reviews 70 (2017) 314–329

the long-term daily radiation data to describe a location, it means that
the solar radiation condition does not change between adjacent years.
These two quantities are intrinsically different when used for practice.
Of course, the daily global radiation is much more useful, but the
difficulty also increases, because it changes rapidly along with the
weather condition.
4.1. Long-term daily global radiation
This section will focus on the models developed to estimate long-
term daily global radiation. In fact, it is not difficult to estimate the
long-term daily global radiation, because it does not consider the
change of weather. In other words, for a particular location it is just a
series of fixed numbers. We just need to find a mathematical function
that matches the trend of the numbers, while the process of estimating
solar radiation also loses its significance. This type of estimation may
be helpful for the simulation study of the long-term performances of
solar energy utilization techniques.
Li et al. [53] built a model that combined sine and cosine functions
for 79 stations across China. More than 10 years of data was used to
determine the coefficients for each station. The only input for this
model was the number of the day of interest. The RMSE of all the
stations ranged from 1.03 to 1.83 MJ/(m2day), while the MAPE ranged
from 4.00% to 15.43%. Zang et al. [1] investigated a model similar to
[53] for 35 stations in China, but this model had reduced two
coefficients. The MAPE and RMSE for the 35 stations ranged from
4.33% to 16.22%, 1.10 to 1.88 MJ/(m2day), respectively.
Khorasanizadeh et al. [54,55] tested 6 empirical models for 4 cities
of Iran, and all the models just needed the day number as input. Four
models depended on sine and cosine function, one model depended on
exponential function, and the last one depended on polynomial
function. The empirical coefficients were separately derived for the
four stations. For all the four stations, the RMSE of the six models
ranged from 0.72 to 1.26 MJ/(m2day), and the MAPE ranged from
3.38% to 5.72%. They [56] also tested the method of adaptive neuro-
fuzzy inference system for Tabass, Iran with the only input of day
number. The RMSE in testing period was 0.89 MJ/(m2day) while the
MAPE was 3.96%. Mghouchi et al. [57] showed different conception in
which they integrated the sum of direct and diffuse irradiance
throughout a day to obtain the daily solar radiation. They tested two
empirical models for 24 Moroccan cities and the only input for these
two models was the day number. The first model needed the day
number to determine the turbidity atmospheric factor while the second
needed the day number to select the proper empirical coefficients to
compute direct normal irradiance and diffuse irradiance. However,
there was no available measured solar radiation data for testing.
The most important feature about these models is, they do not need
more inputs just the number of the day of interest. They are simple but
cannot reasonably estimate the daily solar radiation for an actual day.
However, they provide an easy way to describe the long-term solar
radiation condition of a particular location.
stations were respectively divided into 9 thermal climate zones and 7
solar climate zones based on different criteria. The data from each zone
was used to obtain its general regression coefficients. In the first
classification, the RMSE of linear model ranged from 1.22 to 2.56 MJ/
(m2day) except two stations (3.07, 3.74 MJ/(m2day)). In the second
classification, the RMSE ranged from 1.26 to 2.61 MJ/(m2day) except
one station (3.78 MJ/(m2day)). They also investigated the ANN model
using the day number, latitude, longitude, altitude, daily mean
temperature and sunshine duration fraction as inputs. Liu et al. [24]
studied 2 SDF models, 3 MSDF models and 3 NSD models for Gaize,
Tibetan Plateau. The data of 1085 days was used for calibration and the
data of 701 days was used for validation. The RMSE of the first five
models ranged from 1.68 to 1.79 MJ/(m2day), while the RMSE of the
last three models ranged from 3.05 to 3.13 MJ/(m2day). They also
concluded that deriving coefficients for different seasons respectively
was unnecessary. Zhao et al. [59] studied the linear model for 9
stations in China, in which the air pollution index was introduced to
improve the performance. Three revised models were built and the
RMSE of them ranged from 1.72 to 2.55 MJ/(m2day) except the station
of Lanzhou 5.21–5.24 MJ/(m2day)). Chen et al. [60] studied 5 SDF
models for 3 stations in Liaoning province, China. 35 years of data
from each station was collected. 70% of the data was used to derive
empirical coefficients and 30% was used for testing. The empirical
coefficients were obtained respectively for each station. These five
models had similar accuracy for the same location. For the station of
Chaoyang, the RMSE of them ranged from 2.65 to 2.73 MJ/(m2day),
while for the other two locations, the RMSE ranged from 1.98 to 2.10,
2.25 to 2.37 MJ/(m2day), respectively. They also studied the support
vector machine with using different forms of sunshine duration as
inputs. Jamshid et al. [61] investigated 3 SDF models and 1 MSDF
model for two stations in Iran. More than 5 years of data was collected
and 70% of the data was used for calibration. For the station of
Zahedan, the RMSE of them ranged from 3.37 to 3.70 MJ/(m2day),
while for Bojnurd, the RMSE ranged from 2.14 to 2.91 MJ/(m2day).
They also studied the method of support vector regression, in which the
polynomial and radial basis functions were respectively applied as the
kernel function, and the parameters of sunshine duration, maximum
temperature, minimum temperature and relative humidity were se-
lected as inputs.
There are few MSDF models used to estimate daily global radiation,
see in [24,61], so they are discussed in this section together. In these
two studies, the differences between the RMSE values of the MSDF and
SDF models are very small, which are basically negligible.
Table 6 has summarized the results of above studies. The study of
[33] has not good performance because the MAPE values are very
large. They have developed four models to estimate daily radiation, and
Table 6
The comparison of SDF models for daily global radiation.
Reference Year Models Location RMSE MAPE
number

4.2. Sunshine duration fraction models Kevin et al. [58] 2008 linear 41 1.22 to 2.56, —
1.26 to 2.61
Behrang et al. 2010 6 1 — 10.03% to
Many studies apply the SDF models developed to estimate monthly
[70] 10.9%.
average daily radiation to estimate daily radiation, merely the compu- Li et al. [27] 2011 8 4 1.11 to 1.71 —
tations of variables become for an actual day instead of the average day Manzano et al. 2012 linear 25 1.11 to 1.73 —
of a month. The mathematical expressions of typical SDF models are [37]
already given in Section 3, so they are omitted here to avoid duplica- Liu et al. [24] 2012 5 1 1.68 to 1.79 —
Zhao et al. [59] 2013 linear 9 1.72 to 2.55 —
tion. Also, some papers have studied the empirical models of estimat- Chen et al. [60] 2013 5 3 1.98 to 2.73 —
ing monthly average daily radiation and daily radiation at the same Yao et al. [33] 2014 4 2 3.70 to 4.50, 44.78% to
time, which are already introduced in Section 3, so this section do not 8.06 88.04%
repeat them and just collect the results about the estimation of daily Kasra et al. [36] 2015 4 1 2.65 to 2.76 13.06% to
13.62%
radiation from them.
Jamshid et al. 2015 4 2 2.14 to 2.91, —
Kevin et al. [58] studied the linear Angstrom–Prescott model to [61] 3.37 to 3.7
estimate daily global solar radiation for 41 stations in China. Those

323
J. Zhang et al.
Table 7
The comparison of NSD models for daily global radiation.
Reference Year Models Input parameters
Fortin et al. [62] 2008 3 ΔT, N, φ
Liu et al. [64] 2009 16 ΔT, ΔTm, Tave, Tmax, Tmin, H, precipitation
Li et al. [65] 2010 12 Rainfall, dew point, fog
Tmax, Tmin
Liu et al. [24] 2012 3 ΔT, Tave, transformed rain fall data
Landeras et al. [63] 2012 3 ΔT, N, φ
Li et al. [66] 2013 4 Tmax, Tmin, ΔT, relative humidity
Almorox et al. [67] 2011 8 Tmax, Tmin, ΔT
Almorox [68] 2013 6 Tmax, Tmin, atmospheric pressure, saturation vapor pressure
Pan et al. [69] 2013 1 ΔT
Ali [91] 2011 1 G0, Tmax, Tmin
Where N is the number of the day of interest in a year, G0 is the extraterrestrial radiation on horizontal surface, Tave is the average ambient temperature, Tmax is the maximum ambient
temperature, Tmin is the minimum ambient temperature, ΔT is the difference between Tmax and Tmin, ΔTm is the monthly mean ΔT, φ is the latitude, H is the altitude.
the RMSE of the linear model is the largest, 8.06 MJ/(m2day), but the
reason is not clear. Except [33], the results of other studies are
consistent, in which the RMSE values fluctuate around 2 MJ/(
m2day). Compared with the SDF models used to estimate monthly
average daily radiation, the RMSE values become larger when they are
used to estimate daily radiation. On the whole, the RMSE of these
studies are in the range of 1.11–4.50 MJ/(m2day), in which most of
them are acceptable for engineering uses.
4.3. Non-sunshine duration models
NSD (non-sunshine duration) models do not use the sunshine
duration as input, which are helpful when the data of sunshine
duration is not available for the location of interest. After reviewing
many studies, the author has found the most popular parameter in
NSD models is the ambient temperature. The collection of ambient
temperature data is much easier than that of sunshine duration. The
ambient temperature data has been used in different forms, such as the
temperature difference between maximum and minimum, the ratio of
maximum to minimum, the ratios of maximum to average tempera-
ture. Other parameters such as relative humidity, cloud cover, pre-
cipitation and vapor pressure are sometimes involved.
Fortin et al. [62] studied 3 temperature-based NSD models for 11
stations in northeastern America. The first model just needed the
temperature difference as input. The last two models still needed the
number of the day and latitude to calculate the clear-day solar
radiation. These three models were then optimized by applying
standard unconstrained non-linear scheme, and 1385 observations
were used for training and 1137 observations were used for validation.
They treated the 11 stations as a whole because the stations presented
same meteorological condition. The optimized models showed slightly
better performance, whose RMSE ranged from 4.6 to 4.7 MJ/(m2day).
Landeras et al. [63] optimized the models with same algorithm, in
which the RMSE of optimized models ranged from 3.14 to 3.29 MJ/
(m2day), while the RMSE of raw models ranged from 3.21 to 3.94 MJ/
(m2day). Liu et al. [64] investigated 16 NSD models for 15 sites in
China, in which except three models the others only needed the air
temperature as source data. The models were divided into two groups:
Bristow-Campbell models with exponential forms, and Hargreaves
models with square root forms. 20 years of data was used to calibrate
the models and 10 years of data was used for testing. Overall, the best
performance obtained by this two groups was that RMSE equaled 3.91
and 3.90 MJ/(m2day), respectively. Li et al. [65] investigated 12 NSD
models for Chongqing, China. The parameters of rainfall, mean dew
point temperature and fog were involved in the first three models,
while the maximum temperature and minimum temperature were
added to the last nine models. Then 2921 days were used to calibrate
the models and 2552 days were used for testing. The RMSE of the first
three models ranged from 5.18 to 6.24 MJ/(m2day), while the RMSE of
324
Renewable and Sustainable Energy Reviews 70 (2017) 314–329
Location RMSE MAPE
11 4.60 to 4.70 —
15 3.91, 3.90 —
1 5.18 to 6.24 —
1 2.52 to 3.05 —
1 3.05 to 3.13 —
4 3.14 to 3.29 —
83 3.83 to 4.02 —
7 2.70 to 3.25 16.37% to 29.18%
1 2.05 to 2.14 9.27% to 9.84%
11 2.54 to 3.24 —
1 3.64 —
the last nine models ranged from 2.52 to 3.05 MJ/(m2day). They [66]
also studied 4 NSD models and 2 sunshine duration-based models for
83 stations of China. The stations were divided into five solar radiation
zones using the clearness-means clustering method, and in each zone,
the empirical coefficients were derived respectively. The results in-
dicated that the temperature-based models had poor performance. For
the first zone containing eight stations, the mean RMSE of NSD models
ranged from 3.83 to 4.02 MJ/(m2day), while the mean RMSE of
sunshine duration-based models ranged from 2.71 to 2.75 MJ/
(m2day). Almorox et al. [67] investigated 8 NSD models for 7 stations
located in Madrid, Spain, which mainly depended on the maximum
temperature and minimum temperature. The parameters of mean
temperature, altitude, latitude and the day of the year were also
involved in some models. Overall, the RMSE of the eight models
ranged from 2.70 to 3.25 MJ/(m2day) and the MAPE ranged from
16.37% to 29.18%. They [68] also evaluated 6 temperature-based
models for Cordoba, Argentina. The RMSE of the six models ranged
from 2.05 to 2.14 MJ/(m2day) and the MAPE ranged from 9.27% to
9.84%. It was noteworthy that they had compared the performance of
Hargreaves-Samani model (square root model) on seven time intervals,
from 1 d, 5 days to 25 days, 30 days. The RMSE of the model decreased
from 3.67 to 1.32 MJ/(m2day) along with the increment of the time
interval, and the MAPE decreased from 22.62% to 6.24%. Pan et al.
[69] studied the exponential temperature-based model for 11 meteor-
ological stations on Tibetan Plateau, which only needed the tempera-
ture difference as input. 35 years of data was used to calibrate the
model and 5 years of data was used for testing. The RMSE of the model
for all stations ranged from 2.54 to 3.24 MJ/(m2day).
Table 7 has summarized the results of above studies. More NSD
models are used to estimate daily radiation than that of monthly
average daily radiation. Comparing with the SDF models used to
estimate daily radiation, the estimation errors of NSD models are
larger. Comparing with the NSD models used to estimate monthly
average daily radiation, the errors are larger too. All the RMSE values
of NSD models exceed 2 MJ/(m2day), and the maximum reaches
6.24 MJ/(m2day). More importantly, these values are continuously
distributed in the range rather than that just one or two values are very
large. However, there is a reason for the large values of RMSE in [65],
because the three models use neither the sunshine duration nor the
ambient temperature. On the whole, the RMSE values of them are in
the range of 2.05–4.70 MJ/(m2day).
4.4. Artificial neural networks
Many studies have applied the artificial neural network to estimate
monthly average daily global radiation, and they are already explained
elaborately in Section 3.2. The ANNs used to estimate daily global
radiation have similar configurations, but their outputs are for daily
radiation and the inputs are measured parameters on an actual day
J. Zhang et al. Renewable and Sustainable Energy Reviews 70 (2017) 314–329

Table 8
The comparison of ANN models for daily global radiation.

Reference Year Input parameters RMSE MAPE

Kevin et al. [58,72] 2008 N, S/S0, Tave,φ, L, H, 1.27 to 2.55, 1.24 to 2.63 —
Fortin et al. [62] 2008 N, G0, Tmax, Tmin 3.74 to 4.2 —
Shafiqur et al. [71] 2008 N, Tave, mean relative humidity — 4.49%
N, Tmax — 10.3%
N, Tave — 11.8%
Behrang et al. [70] 2010 N, S, Tave, relative humidity, evaporation, wind speed — 5.21 %to 6.86%, 5.56% to 6.53%
Lu et al. [74] 2011 satellite data, H, air mass 2.86(20.4%), 3.09 (21.9%) —
Ali [91] 2011 G0, Tmax, Tmin 2.53 (13.9%) to 2.55 ( 14.0%) —
Landeras et al. [63] 2012 N, G0, Tmax, Tmin, clear sky radiation 2.93 to 2.97 —
Alvaro et al. [73] 2013 satellite data from Meteosat 9, clear sky radiation 1.25 (6.74%) —
Zeynab et al. [77] 2014 N, G0, S, S0, Tmax, Tmin, clear-sky solar radiation 3.7 —
Mohammadi et al. [22] 2015 G0, S/S0, ΔT, Tave, humidity, water vapor pressure 1.84 9.31%

Where N is the number of the day of interest in a year, S is the actual sunshine duration, S0 is the maximum possible sunshine duration, Tave is the average ambient temperature, Tmax
is the maximum ambient temperature, Tmin is the minimum ambient temperature, ΔT is the difference between Tmax and Tmin, G0 is the extraterrestrial radiation on horizontal
surface, L is the longitude, H is the altitude, φ is the latitude.

instead of the average day of a month. Except two important studies, not show superiority. The RMSE values of ANN models are in the range
we no longer describe the others here. The reasonable and desired of 1.24–4.20 MJ/(m2day).
results from them are collected in Table 8.
Behrang et al. [70] studied MLP (multi-layer perceptron) network
and RBF (radial basis function) network for Dezful, Iran. They applied
4.5. Other techniques
six combinations of the parameters of day number, mean air tempera-
ture, relative humidity, sunshine duration, evaporation and wind speed
Tang et al. [75] evaluated a hybrid model improved by Yang and
as inputs. 1398 days were used to train the models and 214 days were
Koike [76] to estimate daily global solar radiation. The model
used for testing. The results clearly indicated that the parameter of
calculated the clear-sky global radiation and clear sky index respec-
sunshine duration could improve the accuracy of estimation. As to four
tively, and then the product of them was the actual global solar
MLPs containing sunshine duration as input, the MAPE in testing stage
radiation. The daily clear-sky global radiation was divided into two
ranged from 5.21% to 6.86%. As to five RBFs containing sunshine
parts: direct radiation and diffuse radiation, which were integrated
duration as input, the MAPE in testing stage ranged from 5.56% to
throughout a day with the consideration of detailed solar radiation
6.53%. They also studied five SDF models, three MSDF models and one
transfer processes, such as Rayleigh scattering, aerosol extinction,
NSD model for Dezful, Iran. The NSD model just used the relative
ozone absorption, etc. The clear sky index was related with sunshine
humidity as input. Two MSDF models used the relative humidity as
duration using a quadric equation. This model needed seven input
added parameters, the other one considered the relative humidity and
parameters: Angstrom turbidity, thickness of ozone layer, surface
temperature together. The NSD model and the two former MSDF
elevation, air temperature, relative humidity, air pressure and sunshine
models had poor performance. The MAPE were 22.88%, 16.29% and
duration. After a careful quality control, the measured radiation data
18.4%, respectively. The MAPE of other six models ranged from
from 1993 to 2000 for 97 meteorological stations over China was used
10.03% to 10.9%. Gorka et al. [63] compared ANN, GEP (gene
to validate the hybrid model. The average MBE and RMSE of the model
expression programming), ANFIS (adaptive neuro-fuzzy inference
were 0.7 and 2.0 MJ/(m2day) respectively. They also applied that
system) and three temperature-based empirical models for Alava,
model to estimate monthly average daily radiation, while the MBE and
Spain. 4420 observations from four stations were used to train the
RMSE were 0.7 and 1.3 MJ/(m2day), respectively. Zeynab et al.
models and 2855 observations were used for testing. These models
compared two SVR (support vector regression) models with FLR (fuzzy
applied five combinations of the parameters of extraterrestrial radia-
linear regression) in [77], and also compared them with ANN and
tion, maximum air temperature, minimum air temperature, day
ANFIS (adaptive neural fuzzy inference system) in [78]. Radial basis
number and clear sky radiation as inputs. The results proved that
function and polynomial function were separately selected as the kernel
using extraterrestrial radiation as the only input was not reasonable.
functions of two SVR models. Seven years of data from the only station
Using extraterrestrial radiation and maximum temperature as inputs
in Tehran, Iran was collected for training and testing. The inputs of
had shown better performance but not good enough. The RMSE of the
those models were: maximum sunshine duration, the number of the
optimized ANNs using other three combinations as inputs ranged from
day, maximum and minimum temperature, actual sunshine duration,
2.93 to 2.97 MJ/(m2day). The RMSE of the corresponding optimized
clear-sky solar radiation and extraterrestrial radiation. In testing
ANFIS ranged from 3.14 to 3.33 MJ/(m2day). The RMSE of the
phase, the RMSE of the SVR_rbf, SVR_poly, FLR, ANN and ANFIS
optimized GEP ranged from 3.31 to 3.49 MJ/(m2day).
models were 3.3, 3.4, 3.91, 3.7 and 3.8 MJ/(m2day), respectively.
Table 8 has summarized the results of above studies. The perfor-
Jamshid et al. [61] studied two SVR models for two stations in Iran.
mances of the ANN models in different studies are irregular. The
The inputs were sunshine duration, maximum and minimum tempera-
average daily global solar radiation used for testing in [71] is around
ture and relative humidity. In testing phase, the RMSE of SVR-rbf and
20 MJ/( m2day). Kevin et al. [58,72] achieve the best performance by
SVR-poly for Zahedan station were 4.01 and 4.47 MJ/(m2day), while
employing common parameters. This may be due to that each zone
the RMSE for Bojnurd station were 2.01 and 1.63 MJ/(m2day).
have several stations, and the data from different stations is averaged.
Table 9 has summarized other approaches applied to estimate daily
This treatment can reduce the amplitude of the changes of solar
global radiation, and has compared them with ANN models. Due to the
radiation and other parameters. Alvaro et al. [73] apply the satellite
scarcity of available studies, we cannot see clear difference between
data in the ANN model and the performance is very good. However, Lu
these approaches. There are five studies using support vector machine
et al. [74] do not get good performance, which may because the stations
to estimate daily global radiation, except the station of Zahedan in [61],
used for validation are not those used to train the ANN. Compared with
the accuracy of others is acceptable. However, these new approaches do
the SDF models used to estimate daily radiation, the ANN models do
not improve the accuracy of estimation to a new level.

325
J. Zhang et al. Renewable and Sustainable Energy Reviews 70 (2017) 314–329

Table 9 available hourly radiation data in two dimensions, in which the

The comparison of other techniques for daily global radiation. horizontal axis was the ordinal number of a day in a year while the
vertical axis was the ordinal number of an hour in a day. They
Methods Reference RMSE MAPE
concluded that there was strong information between the solar radia-
Radiation transfer model Tang et al. [75] 2.00 — tion intensities of the hour of interest and the adjacent hours. The
Artificial neural network Mohammadi et al. 1.84 9.31% adjacent hours were the hours before the hour of interest in the same
[22]
day, and the hours in the day before the day of interest. After that, they
Zeynab et al. [78] 3.70 —
Support vector machine Mohammadi et al. 1.42 7.00% applied linear correlations and neural networks to extract the informa-
[22] tion to finish the prediction. The best performance achieved by linear
Mohammadi et al. 2.00 10.45% correlation was, with known intensities of three hours as inputs, the
[36] RMSE equaled 39.37 W/m2. The best performance achieved by neural
Chen et al. [60] 1.79 to 2.38 —
network was, with known intensities of three hours as inputs, the
Zeynab et al. [77] 3.30, 3.40 —
Jamshid et al. [61] 4.01, 4.07, — RMSE equaled 34.57 W/m2. Philippe et al. [15] studied six approaches
2.01, 1.63 to forecast hourly clear sky index for three French islands. The hourly
Genetic programming Mohammadi et al. 1.85 9.32% global radiation was the product of hourly clear sky index and hourly
[22]
clear sky radiation. Two of the six approaches were persistence models,
Gorka et al. [63] 3.31 to 3.49 —
Autoregressive moving Mohammadi et al. 1.88 9.40% in which one considered that the intensities of two adjacent hours were
average [22] the same, while the other considered that the clear sky indices of two
Fuzzy linear regression Zeynab et al. [77] 3.91 — adjacent hours were the same. The other four approaches were the
Adaptive neural fuzzy Zeynab et al. [78] 3.8 — linear autoregressive model, neural network, support vector machine
inference system
and Gaussian process model. The general principle of these four
Gorka et al. [63] 3.14 to 3.33 —
models was to correlate the future clear sky index with the past values.
How many past values should be used was determined by a novel
5. Hourly global solar radiation technique that was based on Mutual Information. The forecast horizon
was the time interval between the hour of interest and the hours used
Hourly global solar radiation, with the unit of MJ/(m2hour), is the as inputs. As expected, the first persistence model had the worst
sum of direct radiation and diffuse radiation intercepted by horizontal performance. For Corsica Island, for the forecast horizon of an hour,
surface during an hour. It is much more accurate than daily radiation the relative RMSE of other five approaches ranged from 19.65% to
and monthly average daily radiation, because it has recorded the 21.00%, corresponding to 84.21 W/m2 to 89.99 W/m2. The errors
detailed change of solar radiation in a day. Certainly, the process of increased with the forecast horizon from 1 h to 6 h.
accurate estimation becomes more difficult, because the parameters of
sunshine duration and temperature are no longer usable to estimate 5.2. Estimation models
hourly radiation. Accurate estimation of hourly radiation needs to
consider the scattering, absorption and reflection of the atmospheric Mefti et al. [86] developed a satellite-based model to estimate the
components in detail. hourly clear sky index. They processed the grey level of the pixels of
An hour is commonly the smallest time interval in the measure- high-resolution visible Meteosat images to obtain the cloud cover
ments of official meteorological stations. To some extent, it is reason- index. Then, they applied regression equations to link the clear sky
able to transform the hourly radiation to mean irradiance for many index with cloud cover index. In the data pre-processing period, the
engineering applications, with the unit of W/m2. However, for some available data about the clear sky index and cloud cover index was
precise researches, such as the performance testing of solar collector divided into six groups with respect to the change of solar altitude.
and the control strategy of solar field, the real-time measurement is Meanwhile the sky condition was classified into clear sky, partly cloudy
preferred. Real-time measurement records the solar radiation data on sky and overcast sky. The regression coefficients were then obtained
much smaller time intervals, such as one minute and several minutes, respectively in each subset. They validated the model with nearly
but the maintenance and calibration work is costly. In addition, the 25,000 pairs of data from 25 French meteorological stations, and the
value of real-time measurement is greatly affected by the time and relative RMSE for estimation ranged from 11% to 32%. Shamim et al.
location. As a matter of fact, usually only professional research [87] developed an improved technique to estimate hourly global solar
institutes and universities will perform this type of measurement for radiation, in which they applied a meso-scale meteorological model to
some short periods. determine the air pressure and relative humidity for different atmo-
Through investigating the literature, we have found that a large spheric layers. Then, they used the air pressure and relative humidity to
proportion of studies are to predict the global solar radiation for compute the cloud cover index, which was then empirically related to
impending hours, such as [14,15,79–84]. The basic principle of them is an atmospheric transmission factor by using available measured data.
to correlate the solar radiation of future hours with the values of past In the end, they utilized the obtained transmission factor and clear sky
hours in applying different approaches. However, [85] tested the radiation to calculate the actual hourly global solar radiation by an
prediction abilities of autoregressive moving average model and time empirical correlation. The clear sky radiation was estimated by using a
delay neural network using the data series of monthly average hourly meteorological radiation transfer model, in which the scattering,
radiation, since the trend of the averaged data was much more stable. absorption and transmittance of atmospheric components were con-
Two studies concerning the "prediction" of solar radiation instead of sidered explicitly. One year of data was used for training while the data
"estimation" are briefly introduced in the following section, to distin- from another year was used for testing. The RMSE in testing period
guish the difference between these two terms. After that, several studies was 110.83 W/m2 and the MBE was 24.66 W/m2. Yao et al. [88]
aiming at estimating hourly radiation are carefully explained. evaluated eleven existing mathematical expressions of the ratio of
hourly solar radiation to daily solar radiation, and then developed three
new expressions. In this kind of model, the daily global radiation is
5.1. Forecasting models known and the key is to determine the ratio of hourly radiation to daily
radiation. Six of the models were similar to the widely cited Collares-
Fatih et al. [14] proposed a two-dimensional representation model Pereira and Rabl model, while four of them depended on Gaussian
to predict hourly solar radiation. The basic theory was, they tabulated functions. They used 10979 pairs of data from three years to validate

326
J. Zhang et al.
the models. Except one model (0.382 MJ/(m2hour)), the RMSE of
other ten models ranged from 0.324 to 0.356 MJ/(m2hour), corre-
sponding to 90.00 to 98.89 W/m2. The RMSE of the three new
developed models ranged from 0.318 to 0.321 MJ/(m2hour), corre-
sponding to 88.33 to 89.17 W/m2. After that, they tested all the models
with six months of data from another site in the same city. The RMSE
of the ten models ranged from 0.473 to 0.512 MJ/(m2hour), corre-
sponding to 131.39 to 142.22 W/m2. The RMSE of the three new
developed models ranged from 0.474 to 0.481 MJ/(m2hour), corre-
sponding to 131.67 to 133.61 W/m2.
5.2.1. Monthly average hourly global radiation
Wan Nik et al. [89] tested six mathematical expressions of the ratio
of hourly solar radiation to daily radiation, in which most of them were
fitted Gaussian functions. The estimation was for long-term monthly
average hourly radiation. Three years of data from three sites of
Malaysian was used to test the models. They found that the Collares-
Pereira and Rabl model gave the best performance and the relative
RMSE for the three sites ranged from 8.22% to 26.49%. Janjai et al.
[90] developed a satellite-based model to estimate monthly average
hourly global solar radiation, in which the visible channel data from the
GMS5 satellite was utilized. In this model, the earth's atmosphere was
assumed to consist of two layers, in which one was from the top
atmosphere to the top layer of cloud and the other one was from the top
layer of cloud to the ground. A solar radiation budget was built between
the two layers using five various coefficients of transmittance, absorp-
tion and reflectance. Two of them, the surface albedo and broadband
earth-atmospheric albedo were obtained by using satellite data. The
estimations were performed for eight hours from 8:00 in the morning
to 16:00 in the afternoon. They validated the model with one year of
data from 25 stations in Thailand and the relative RMSE for all the
hours ranged from 7.5% to 14.7%. The biggest errors happened in the
early morning and late afternoon. The relative RMSE for hours during
9:00 and 15:00 ranged from 7.5% to 10.7%.
6. Conclusions
Solar radiation data is critical in the design and assessment of solar
energy technologies. It is very important to understand the variation
characteristics, measurement modes and estimation methods about
solar radiation. The models used to estimate the monthly average daily,
daily and hourly global radiation on horizontal surface are reviewed,
classified and compared in this paper. The terms of long-term monthly
average daily radiation and long-term daily radiation are also dis-
cussed. The results of corresponding studies are collected to illustrate
their performances. The detailed discussions and noteworthy questions
have already been given in each category, here we just make a brief
summary about them. On the whole, the estimation error increases
when the time span becomes short, and the ANN models and some new
developed techniques have not shown desired improvement compared
with empirical models. The accuracies of different types of models are
shown below.
1) The RMSEs of the sunshine duration fraction models are in the
range of 0.36–2.05 MJ/(m2day) and 1.11–4.50 MJ/(m2day) for
monthly average daily and daily radiation respectively.
2) The RMSEs of the non-sunshine duration models are in the range of
0.71–2.43 MJ/(m2day) and 2.05–4.70 MJ/(m2day) for monthly
average daily and daily radiation respectively.
3) The RMSEs of the artificial neural networks are in the range of
0.32–2.05 MJ/(m2day) and 1.24–4.2 MJ/(m2day) for monthly
average daily and daily radiation respectively.
4) The RMSEs of the models used to estimate hourly radiation are in
the range of 88.33–142.22 W/m2.
Although the collection of related studies is not exhaustive, it can
327
Renewable and Sustainable Energy Reviews 70 (2017) 314–329
reasonably reflect the performances of different types of models. Also,
this paper has provided an effective way to compare the results from
different studies. The parameter of sunshine duration has the greatest
influence on the accuracy of estimation models, and the ambient
temperature is considered most when the sunshine duration data is
not available.
Some research gaps in this field are discussed here. Most of the
existing empirical models are related to sunshine duration and ambient
temperature, almost there is no room for further improvement if the
basic principle does not change. The daily radiation and hourly
radiation data is much more useful for engineering uses, and should
be paid attention to. There is a lack of in-depth theoretical analysis
about the methods of estimating solar radiation. The artificial intelli-
gence techniques are popular in recent years but the practical effec-
tiveness and superiority to estimate solar radiation need careful
examinations.
Acknowledgement
This study was financial supported by National High Technology
Research and Development Program (863 Program) of China under
Grant number 2015AA050403, and National Natural Science
Foundation of China under Grant number 51476110.
References
[1] Zang H, Xu Q, Bian H. Generation of typical solar radiation data for different
climates of China. Energy 2012;38:236–48.
[2] Angstrom A. Solar and terrestrial radiation. Q J R Meteor Soc 1924;50:121–5.
[3] Page J. The estimation of monthly mean values of daily total shortwave radiation on
vertical and inclined surfaces from sunshine records for latitude 40。N-40。S.
Proceedings UN conference on new sources of energy; 1961. p. 378-390.
[4] Bakirci K. Models of solar radiation with hours of bright sunshine: a review. Renew
Sustain Energy Rev 2009;13:2580–8.
[5] 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.
[6] Khatib T, Mohamed A, Sopian K. A review of solar energy modeling techniques.
Renew Sustain Energy Rev 2012;16:2864–9.
[7] Yadav AK, Chandel SS. Solar radiation prediction using Artificial Neural Network
techniques: a review. Renew Sustain Energy Rev 2014;33:772–81.
[8] Despotovic M, Nedic V, Despotovic D, Cvetanovic S. Review and statistical analysis
of different global solar radiation sunshine models. Renew Sustain Energy Rev
2015;52:1869–80.
[9] Kumar R, Aggarwal RK, Sharma JD. Comparison of regression and artificial neural
network models for estimation of global solar radiations. Renew Sustain Energy
Rev 2015;52:1294–9.
[10] Qazi A, Fayaz H, Wadi A, Raj RG, Rahim NA, Khan WA. The artificial neural
network for solar radiation prediction and designing solar systems: a systematic
literature review. J Clean Prod 2015;104:1–12.
[11] Teke A, Yıldırım HB, Çelik Ö. Evaluation and performance comparison of different
models for the estimation of solar radiation. Renew Sustain Energy Rev
2015;50:1097–107.
[12] Piri J, Kisi O. Modelling solar radiation reached to the Earth using ANFIS, NN-
ARX, and empirical models (Case studies: Zahedan and Bojnurd stations). J Atmos
Sol Terr Phys 2015;123:39–47.
[13] Inman RH, Pedro HTC, Coimbra CFM. Solar forecasting methods for renewable
energy integration. Prog Energy Combust Sci 2013;39:535–76.
[14] Hocaoğlu FO, Gerek ÖN, Kurban M. Hourly solar radiation forecasting using
optimal coefficient 2-D linear filters and feed-forward neural networks. Sol Energy
2008;82:714–26.
[15] Lauret P, Voyant C, Soubdhan T, David M, Poggi P. A benchmarking of machine
learning techniques for solar radiation forecasting in an insular context. Sol Energy
2015;112:446–57.
[16] Duffie JA, Beckman WA. Solar engineering of thermal processes. Hoboken, New
Jersy: John Wiley & Sons, Inc; 2013.
[17] El-Sebaii AA, Al-Ghamdi AA, Al-Hazmi FS, Faidah AS. Estimation of global solar
radiation on horizontal surfaces in Jeddah, Saudi Arabia. Energy Policy
2009;37:3645–9.
[18] El-Sebaii AA, Al-Hazmi FS, Al-Ghamdi AA, Yaghmour SJ. Global, direct and diffuse
solar radiation on horizontal and tilted surfaces in Jeddah, Saudi Arabia. Appl
Energy 2010;87:568–76.
[19] Jin Z, Yezheng W, Gang Y. General formula for estimation of monthly average daily
global solar radiation in China. Energy Convers Manag 2005;46:257–68.
[20] Shamshirband S, Mohammadi K, Yee PL, Petković D, Mostafaeipour A. A
comparative evaluation for identifying the suitability of extreme learning machine
to predict horizontal global solar radiation. Renew Sustain Energy Rev
2015;52:1031–42.
[21] Khorasanizadeh H, Mohammadi K. Introducing the best model for predicting the
J. Zhang et al.
monthly mean global solar radiation over six major cities of Iran. Energy
2013;51:257–66.
[22] Mohammadi K, Shamshirband S, Tong CW, Arif M, Petković D, Ch S. A new hybrid
support vector machine–wavelet transform approach for estimation of horizontal
global solar radiation. Energy Convers Manag 2015;92:162–71.
[23] Olatomiwa L, Mekhilef S, Shamshirband S, Mohammadi K, Petković D, Sudheer C.
A support vector machine–firefly algorithm-based model for global solar radiation
prediction. Sol Energy 2015;115:632–44.
[24] Liu J, Liu J, Linderholm HW, Chen D, Yu Q, Wu D, et al. Observation and
calculation of the solar radiation on the Tibetan Plateau. Energy Convers Manag
2012;57:23–32.
[25] Bakirci K. Correlations for estimation of daily global solar radiation with hours of
bright sunshine in Turkey. Energy 2009;34:485–501.
[26] Katiyar AK, Pandey CK. Simple correlation for estimating the global solar radiation
on horizontal surfaces in India. Energy 2010;35:5043–8.
[27] Li H, Ma W, Lian Y, Wang X, Zhao L. Global solar radiation estimation with
sunshine duration in Tibet, China. Renew Energy 2011;36:3141–5.
[28] Behrang MA, Assareh E, Noghrehabadi AR, Ghanbarzadeh A. New sunshine-based
models for predicting global solar radiation using PSO (particle swarm optimiza-
tion) technique. Energy 2011;36:3036–49.
[29] Duzen H, Aydin H. Sunshine-based estimation of global solar radiation on
horizontal surface at Lake Van region (Turkey). Energy Convers Manag
2012;58:35–46.
[30] Teke A, Yıldırım HB. Estimating the monthly global solar radiation for Eastern
Mediterranean Region. Energy Convers Manag 2014;87:628–35.
[31] Park J-K, Das A, Park J-H. A new approach to estimate the spatial distribution of
solar radiation using topographic factor and sunshine duration in South Korea.
Energy Convers Manag 2015;101:30–9.
[32] Chelbi M, Gagnon Y, Waewsak J. Solar radiation mapping using sunshine duration-
based models and interpolation techniques: application to Tunisia. Energy Convers
Manag 2015;101:203–15.
[33] Yao W, Li Z, Wang Y, Jiang F, Hu L. Evaluation of global solar radiation models for
Shanghai, China. Energy Convers Manag 2014;84:597–612.
[34] Almorox J, Benito M, Hontoria C. Estimation of monthly Angström–Prescott
equation coefficients from measured daily data in Toledo, Spain. Renew Energy
2005;30:931–6.
[35] İnci Türk Toğrula HT, Evinb Dugyu. Estimation of monthly global solar radiation
from sunshine duration measurement in Elaziğ. Renew Energy 2000;19:587–95.
[36] Mohammadi K, Shamshirband S, Anisi MH, Alam KA, Petković D. Support vector
regression based prediction of global solar radiation on a horizontal surface. Energy
Convers Manag 2015;91:433–41.
[37] Manzano A, Martín ML, Valero F, Armenta C. A single method to estimate the daily
global solar radiation from monthly data. Atmos Res 2015;166:70–82.
[38] Dumas A, Andrisani A, Bonnici M, Graditi G, Leanza G, Madonia M, et al. A new
correlation between global solar energy radiation and daily temperature variations.
Sol Energy 2015;116:117–24.
[39] Adaramola MS. Estimating global solar radiation using common meteorological
data in Akure, Nigeria. Renew Energy 2012;47:38–44.
[40] Korachagaon I, Bapat VN. General formula for the estimation of global solar
radiation on earth's surface around the globe. Renew Energy 2012;41:394–400.
[41] Qin J, Chen Z, Yang K, Liang S, Tang W. Estimation of monthly-mean daily global
solar radiation based on MODIS and TRMM products. Appl Energy
2011;88:2480–9.
[42] Şenkal O, Kuleli T. Estimation of solar radiation over Turkey using artificial neural
network and satellite data. Appl Energy 2009;86:1222–8.
[43] Şenkal O. Modeling of solar radiation using remote sensing and artificial neural
network in Turkey. Energy 2010;35:4795–801.
[44] Ozgoren M, Bilgili M, Sahin B. Estimation of global solar radiation using ANN over
Turkey. Expert Syst Appl 2012;39:5043–51.
[45] Mohandes MA. Modeling global solar radiation using particle swarm optimization
(PSO). Sol Energy 2012;86:3137–45.
[46] Yadav AK, Malik H, Chandel SS. Selection of most relevant input parameters using
WEKA for artificial neural network based solar radiation prediction models. Renew
Sustain Energy Rev 2014;31:509–19.
[47] Yadav AK, Malik H, Chandel SS. Application of rapid miner in ANN based
prediction of solar radiation for assessment of solar energy resource potential of 76
sites in Northwestern India. Renew Sustain Energy Rev 2015;52:1093–106.
[48] Jiang H, Dong Y, Wang J, Li Y. Intelligent optimization models based on hard-ridge
penalty and RBF for forecasting global solar radiation. Energy Convers Manag
2015;95:42–58.
[49] Olatomiwa L, Mekhilef S, Shamshirband S, Petković D. Adaptive neuro-fuzzy
approach for solar radiation prediction in Nigeria. Renew Sustain Energy Rev
2015;51:1784–91.
[50] Sun H, Yan D, Zhao N, Zhou J. Empirical investigation on modeling solar radiation
series with ARMA–GARCH models. Energy Convers Manag 2015;92:385–95.
[51] Ayodele TR, Ogunjuyigbe ASO. Prediction of monthly average global solar radiation
based on statistical distribution of clearness index. Energy 2015;90:1733–42.
[52] Janjai S, Pankaew P, Laksanaboonsong J, Kitichantaropas P. Estimation of solar
radiation over Cambodia from long-term satellite data. Renew Energy
2011;36:1214–20.
[53] Li H, Ma W, Lian Y, Wang X. Estimating daily global solar radiation by day of year
in China. Appl Energy 2010;87:3011–7.
[54] Khorasanizadeh H, Mohammadi K. Prediction of daily global solar radiation by day
of the year in four cities located in the sunny regions of Iran. Energy Convers
Manag 2013;76:385–92.
[55] Khorasanizadeh H, Mohammadi K, Jalilvand M. A statistical comparative study to
328
Renewable and Sustainable Energy Reviews 70 (2017) 314–329
demonstrate the merit of day of the year-based models for estimation of horizontal
global solar radiation. Energy Convers Manag 2014;87:37–47.
[56] Mohammadi K, Shamshirband S, Tong CW, Alam KA, Petković D. Potential of
adaptive neuro-fuzzy system for prediction of daily global solar radiation by day of
the year. Energy Convers Manag 2015;93:406–13.
[57] El Mghouchi Y, Ajzoul T, El Bouardi A. Prediction of daily solar radiation intensity
by day of the year in twenty-four cities of Morocco. Renew Sustain Energy Rev
2016;53:823–31.
[58] Wan KKW, Tang HL, Yang L, Lam JC. An analysis of thermal and solar zone
radiation models using an Angstrom–Prescott equation and artificial neural net-
works. Energy 2008;33:1115–27.
[59] Zhao N, Zeng X, Han S. Solar radiation estimation using sunshine hour and air
pollution index in China. Energy Convers Manag 2013;76:846–51.
[60] Chen J-L, Li G-S, Wu S-J. Assessing the potential of support vector machine for
estimating daily solar radiation using sunshine duration. Energy Convers Manag
2013;75:311–8.
[61] Piri J, Shamshirband S, Petković D, Tong CW. Rehman MHu. prediction of the
solar radiation on the earth using support vector regression technique. Infrared
Phys Technol 2015;68:179–85.
[62] Fortin JG, Anctil F, Parent L-É, Bolinder MA. Comparison of empirical daily
surface incoming solar radiation models. Agric Meteor 2008;148:1332–40.
[63] Landeras G, López JJ, Kisi O, Shiri J. Comparison of Gene Expression
Programming with neuro-fuzzy and neural network computing techniques in
estimating daily incoming solar radiation in the Basque Country (Northern Spain).
Energy Convers Manag 2012;62:1–13.
[64] Liu X, Mei X, Li Y, Wang Q, Jensen JR, Zhang Y, et al. Evaluation of temperature-
based global solar radiation models in China. Agric Meteor 2009;149:1433–46.
[65] Li M-F, Liu H-B, Guo P-T, Wu W. Estimation of daily solar radiation from routinely
observed meteorological data in Chongqing, China. Energy Convers Manag
2010;51:2575–9.
[66] Li M-F, Tang X-P, Wu W, Liu H-B. General models for estimating daily global solar
radiation for different solar radiation zones in mainland China. Energy Convers
Manag 2013;70:139–48.
[67] Almorox J, Hontoria C, Benito M. Models for obtaining daily global solar radiation
with measured air temperature data in Madrid (Spain). Appl Energy
2011;88:1703–9.
[68] Almorox J, Bocco M, Willington E. Estimation of daily global solar radiation from
measured temperatures at Cañada de Luque, Córdoba, Argentina. Renew Energy
2013;60:382–7.
[69] Pan T, Wu S, Dai E, Liu Y. Estimating the daily global solar radiation spatial
distribution from diurnal temperature ranges over the Tibetan Plateau in China.
Appl Energy 2013;107:384–93.
[70] Behrang MA, Assareh E, Ghanbarzadeh A, Noghrehabadi AR. The potential of
different artificial neural network (ANN) techniques in daily global solar radiation
modeling based on meteorological data. Sol Energy 2010;84:1468–80.
[71] Rehman S, Mohandes M. Artificial neural network estimation of global solar
radiation using air temperature and relative humidity. Energy Policy
2008;36:571–6.
[72] Lam JC, Wan KKW, Yang L. Solar radiation modelling using ANNs for different
climates in China. Energy Convers Manag 2008;49:1080–90.
[73] Linares-Rodriguez A, Ruiz-Arias JA, Pozo-Vazquez D, Tovar-Pescador J. An
artificial neural network ensemble model for estimating global solar radiation from
Meteosat satellite images. Energy 2013;61:636–45.
[74] Lu N, Qin J, Yang K, Sun J. A simple and efficient algorithm to estimate daily global
solar radiation from geostationary satellite data. Energy 2011;36:3179–88.
[75] Tang W, Yang K, He J, Qin J. Quality control and estimation of global solar
radiation in China. Sol Energy 2010;84:466–75.
[76] Yang K, Koike T, Ye B. Improving estimation of hourly, daily, and monthly solar
radiation by importing global data sets. Agric Meteor 2006;137:43–55.
[77] Ramedani Z, Omid M, Keyhani A, Khoshnevisan B, Saboohi H. A comparative study
between fuzzy linear regression and support vector regression for global solar
radiation prediction in Iran. Sol Energy 2014;109:135–43.
[78] Ramedani Z, Omid M, Keyhani A, Shamshirband S, Khoshnevisan B. Potential of
radial basis function based support vector regression for global solar radiation
prediction. Renew Sustain Energy Rev 2014;39:1005–11.
[79] Yang D, Jirutitijaroen P, Walsh WM. Hourly solar irradiance time series forecasting
using cloud cover index. Sol Energy 2012;86:3531–43.
[80] Huang J, Korolkiewicz M, Agrawal M, Boland J. Forecasting solar radiation on an
hourly time scale using a Coupled AutoRegressive and Dynamical System (CARDS)
model. Sol Energy 2013;87:136–49.
[81] Akarslan E, Hocaoğlu FO, Edizkan R. A novel M-D (multi-dimensional) linear
prediction filter approach for hourly solar radiation forecasting. Energy
2014;73:978–86.
[82] Akarslan E, Hocaoglu FO. A novel adaptive approach for hourly solar radiation
forecasting. Renew Energy 2016;87:628–33.
[83] Mazorra Aguiar L, Pereira B, David M, Díaz F, Lauret P. Use of satellite data to
improve solar radiation forecasting with Bayesian Artificial Neural Networks. Sol
Energy 2015;122:1309–24.
[84] Kaplanis S, Kaplani E. Stochastic prediction of hourly global solar radiation for
Patra, Greece. Appl Energy 2010;87:3748–58.
[85] Ji W, Chee KC. Prediction of hourly solar radiation using a novel hybrid model of
ARMA and TDNN. Sol Energy 2011;85:808–17.
[86] Mefti A, Adane A, Bouroubi MY. Satellite approach based on cloud cover
classification: estimation of hourly global solar radiation from meteosat images.
Energy Convers Manag 2008;49:652–9.
[87] Shamim MA, Remesan R, Bray M, Han D. An improved technique for global solar
J. Zhang et al.
radiation estimation using numerical weather prediction. J Atmos Sol Terr Phys
2015;129:13–22.
[88] Yao W, Li Z, Xiu T, Lu Y, Li X. New decomposition models to estimate hourly global
solar radiation from the daily value. Sol Energy 2015;120:87–99.
[89] Wan Nik WB, Ibrahim MZ, Samo KB, Muzathik AM. Monthly mean hourly global
solar radiation estimation. Sol Energy 2012;86:379–87.
329
Renewable and Sustainable Energy Reviews 70 (2017) 314–329
[90] Janjai S, Pankaew P, Laksanaboonsong J. A model for calculating hourly global
solar radiation from satellite data in the tropics. Appl Energy 2009;86:1450–7.
[91] Rahimikhoob A. Estimating global solar radiation using artificial neural network
and air temperature data in a semi-arid environment. Renew Energy
2010;35:2131–5.