Journal of Atmospheric and Solar-Terrestrial Physics 146 (2016) 110122

Contents lists available at ScienceDirect

Journal of Atmospheric and Solar-Terrestrial Physics

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

Estimation of global solar radiation using an articial neural network

based on an interpolation technique in southeast China
Ling Zou a, Lunche Wang b, Aiwen Lin a,n, Hongji Zhu a, Yuling Peng a, Zhenzhen Zhao a
a
b

School of Resource and Environmental Sciences, Wuhan University, Wuhan, Hubei Province 430079, China
Department of Geography, School of Earth Sciences, China University of Geosciences, Wuhan, Hubei Province 430074, China

art ic l e i nf o

a b s t r a c t

Article history:
Received 8 March 2016
Received in revised form
5 May 2016
Accepted 30 May 2016
Available online 31 May 2016

Solar radiation plays important roles in energy application, vegetation growth and climate change.
Empirical relations and machine-learning methods have been widely used to estimate global solar radiation (GSR) in recent years. An articial neural network (ANN) based on spatial interpolation is developed to estimate GSR in southeast China. The improved BristowCampbell (IBC) model and the improved ngstrmPrescott (IAP) model are compared with the ANN model to explore the best model in
solar radiation modeling. Daily meteorological parameters, such as sunshine duration hours, mean
temperature, maximum temperature, minimum temperature, relative humidity, precipitation, air pressure, water vapor pressure, and wind speed, along with station-measured GSR and a daily surface GSR
dataset over China obtained from the Data Assimilation and Modeling Center for Tibetan Multi-spheres
(DAM), are used to predict GSR and to validate the models in this work. The ANN model with the network
of 9-17-1 provides better accuracy than the two improved empirical models in GSR estimation. The rootmean-square error (RMSE), mean bias error (MBE), and determination coefcient (R2) are 2.65 MJ m 2,
0.94 MJ m 2, and 0.68 in the IA P model; 2.19 MJ m 2, 1.11 MJ m 2, and 0.83 in the IBC model;
1.34 MJ m 2, 0.11 MJ m 2, and 0.91 in the ANN model, respectively. The regional monthly mean GSR in
the measured dataset, DAM dataset, and ANN model is analyzed. The RMSE (RMSE %) is 1.07 MJ m 2
(8.91%) and the MBE (MBE %) is 0.62 MJ m 2 ( 5.21%) between the measured and ANN-estimated
GSR. The statistical errors of RMSE (RMSE %) are 0.91 MJ m 2 (7.28%) and those of MBE (MBE %) are
0.15 MJ m 2 ( 1.20%) between DAM and ANN-modeled GSR. The correlation coefcients and R2 are
larger than 0.95. The regional mean GSR is 12.58 MJ m 2. The lowest GSR is observed in the northwest
area, and it increases from northwest to southeast. The annual mean GSR decreases by
0.02 MJ m 2 decade 1 over the entire southeast China. The GSR in 52 stations experiences a decreasing
trend, and 21% of the stations are signicant at the 95% level.
& 2016 Elsevier Ltd. All rights reserved.

Keywords:
Articial neural network
Spatial interpolation
Empirical models
Global solar radiation

1. Introduction
Solar radiation is the primary energy source on earth and is
indispensable for numerous applications. For example, this radiant
energy provides energy for crop growth and affects the potential
yield and water use of crops (Hook and Mcclendon, 1992, Supit
and Van Kappel, 1998), it leads to different climate zones on earth
and determines the regional climate (Budyko, 1969, Tymvios et al.,
2005), and it signicantly affects the hydrological characteristic
and the balance of the ecological system of earth (Ramanathan
et al., 2001, Rivington et al., 2005, Zhang et al., 2015a, b). Continuous and accurate solar radiation measurement is imperative to
n
Correspondence to:School of Resource and Environmental Sciences, Wuhan
University, 129 Luoyu Road, Wuhan 430079 China.
E-mail addresses: aiwen_lin@163.com, awlin@whu.edu.cn (A. Lin).

http://dx.doi.org/10.1016/j.jastp.2016.05.013
1364-6826/& 2016 Elsevier Ltd. All rights reserved.

scientic research and energy application over the long term.

However, lack of direct measurement of solar radiation is a common issue in many remote and undeveloped areas (AntonanzasTorres et al., 2014, Davies et al., 1975). Therefore, predicting
methods for achieving accurate and continuous solar radiation
have been developed in recent decades (Bakirci, 2009, Mohandes
and Rehman, 2010).
Many researchers have exploited techniques for solar radiation
estimating, and satellite remote-sensing data have been used to
derive radiation since the 1960s (Pinker et al., 1995). Chen et al.
(2014) estimate the direct, diffuse, and global solar radiation (GSR)
during 20032011 based on moderate-resolution imaging spectroradiometer (MODIS) products, and a good agreement is observed between the estimated radiation and observations. Polo
et al. (2011) estimate the daily global and direct solar radiation in
India by using Meteosat-images from 2000 to 2007. Journee and
Bertrand (2010) improve the spatial and temporal resolution of

L. Zou et al. / Journal of Atmospheric and Solar-Terrestrial Physics 146 (2016) 110122

GSR with the site-based data and satellite images over Belgium.
Vignola et al. (2007) estimate hourly global, beam, and diffuse irradiance by using GOES8 satellite images. These researchers also
test and validate the model with the data from the University of
Oregon Solar Radiation Monitoring Network and from Kimberly
Idaho. According to Zhang et al. (2015a, b), satellite radiation
models show higher accuracy in generating daily irradiation than
measured data from ground stations. Qin et al. (2011) demonstrate
that remote-sensing data can estimate the radiation in remote
regions because they provide continuous signals in space and
time. However, the spatial resolution of satellite data is coarse, and
the time series of the data is short; thus, reconstructing sophisticated and historical solar radiation is difcult (Bindi and Miglietta,
1991).
Empirical relation models based on climatic variables are useful
in estimating solar radiation. Angstrom (1924) and Prescott (1940)
propose the ngstrmPrescott (AP) model to explore the linear
relationship between sunshine duration (SSD) and GSR. Almorox
and Hontoria (2004) employ the AP model and several improved
AP models to estimate the GSR in Spain using SSD as inputs.
Rehman (1999) propose a new model by improving the site-specic coefcients in the AP model and compare it with many existing SSD-based models to predict GSR at Saudi Arabia. Yang et al.
(2001, 2006) propose the Yang hybrid model on the basis of the A
P model, which is proven to be one of the best sunshine-based
broadband models for GSR estimating. Yaniktepe and Genc (2015)
compare and analyze eight different linear, second, and third order
polynomial models in Turkey and propose the best third-order
polynomial model to estimate the GSR in the study area. Temperature-based models that explore the relationship between solar radiation and maximum/minimum temperatures, such as the
BristowCampbell (BC) and RichardsonWright models, are important models for estimating GSR according to Bristow and
Campbell (1984) and Richardson and Wright (1984). Meza and
Varas (2000) estimate monthly mean GSR by using two temperature-based models in Chile. Rehman and Mohandes (2008)
estimate the GSR in Saudi Arabia by using the measured temperature and precipitation data during 19982002. Panday and
Katiyar (2010) use several temperature-based models to estimate
the GSR in India, and they prove that temperature-based models
can replace SSD-based models for estimating solar radiation. Other
meteorological data, such as precipitation, humidity, and cloud
cover, are used to model solar radiation worldwide (Ehnberg and
Bollen, 2005, Muneer and Gul, 2000, Thornton et al., 2000,
Thornton and Running, 1999, Wilks, 1999).
Spatial interpolation (Bindi and Miglietta, 1991, Hay and Suckling, 1979, Rivington et al., 2006) estimates solar radiation at one
site by employing interpolated values from nearby stations, but it
is rarely applied in areas where sparse stations are detected (Supit
and Van Kappel, 1998). Suckling (1985) used extrapolation to estimate daily GSR form nearby stations; Rehman and Ghori (2000)
employed the geo-statistical technique to estimate GSR, the kriging was used to predict GSR at stations without radiation values.
ANN is an efcient solar radiation estimation technique using
meteorological and geographical variables (Kashyap et al., 2015,
Mohandes et al., 1998, Teke et al., 2015). The ANN model has been
widely used in recent years to predict and forecast global, direct,
and diffuse solar radiations worldwide. For example, Mohandes
et al. (2000) use the radial basis functions in NN to estimate the
GSR in 41 sites in Saudi Arabia. Amrouche and Le Pivert (2014)
combine ANN and spatial modeling to predict the horizontal GSR
in France. Rehman and Mohandes (2009) use ANN with the inputs
of day, temperature, and relative humidity to estimate the diffuse
solar radiation in Saudi Arabia. Wang et al. (2016) use three types
of ANN and the improved BC model to estimate the solar radiation
in China.

111

In China, solar radiation measurement began since 1957. According to Tang et al. (2013) and Wang et al. (2015), approximately
700 routine meteorological stations and only 122 radiation stations have been established from 1961 over the entire country. The
radiation stations are sparsely distributed and mainly deployed in
relatively low-altitude and developed areas. Many researchers
have estimated and calibrated the solar radiation over China by
using satellite data, empirical model, and ANN model. For example, Chen and Li (2013) develop 20 empirical models based on
SSD and temperature to estimate the monthly mean daily GSR in
the Yangtze River Basin. Liu et al. (2015) explore the relationship
between SSD and GSR in large cities in China. Zhang et al. (2015a,
b) use MODIS atmospheric data and the digital elevation model to
estimate downward surface shortwave radiation. Jiang (2009)
proposes a feed-forward back-propagation (BP) network to estimate the daily GSR in eight typical cities in China and proves that
the ANN model is an excellent empirical model. However, these
studies have mainly focused on radiation estimation at a single site
or reconstructed long-time serial solar radiation at one site (Lam
et al., 2008, Tang et al., 2013).
The present study attempts to estimate solar radiation at
multiple sites in southeast China for a long-term period by developing an ANN model based on spatial interpolation. Different
meteorological variables, including sunshine duration (SSD),
average temperature (T), maximum/minimum temperature (TM/
Tm), relative humidity (Rh), precipitation (Pre), air pressure (Pa),
water vapor pressure (Pwv), and wind speed (Ws), are combined to
explore the optimal network. Two improved empirical models,
namely, IBC and IAP models, are compared with the ANN model
to evaluate their performances of solar radiation estimation in this
study. All model results are evaluated with the root-mean-square
errors (RMSEs), mean bias errors (MBEs), and determination
coefcient (R2) in this work. Interpolation is also used to analyze
the spatial characteristics of site-measured GSR, ANNpredicted,
and DAM-modeled GSR in southeast China. Finally, the long-term
trends of GSR in 60 stations are investigated in detail.

2. Material and method

2.1. Study area and data
Southeast China is located between 2032N and 108123E
(Fig. 1). Low hills and mountains are the main landforms in this
region, which mainly comprises a subtropical monsoon climate
zone that is hot and rainy in summer and warm and moist in
winter. The 10 radiation stations and 60 meteorological stations
from 1961 to 2010 are shown in Fig. 1, and the details of the
geographical location of the 10 radiation stations with measured
GSR are shown in Table 1. The annual mean GSR ranges from
10.89 MJ m 2 day 1 at station 57461 to 13.85 MJ m 2 day 1 at
station 59316. The annual mean temperature varies from 16.65 C
at station 58457 to 21.62 C at station 59316. The annual mean SSD
increases from 4.15 h at station 57957 to 5.58 h at station 59316.
The annual total Pre is between 1138.02 and 1627.08 mm in the 10
stations, and the annual mean Rh ranges from 74.91% at station
57993 to 80.23% at station 59316. The highest T, SSD, Rh, and GSR
are observed at station 59316, which may be due to its low latitude
(2323) and altitude (2.9 m). The geographical and meteorological
information of 50 meteorological stations without GSR data is
presented in Table 2. The annual mean SSD ranges from 3.42 h to
5.59 h, and the annual mean temperature ranges from 15.79 C to
22.20 C. The monthly variations in SSD, Pre, Rh, T and TM/Tm at
the 10 radiation stations are shown in Fig. 2. High daily T are
observed in July (28.82 C) and August (28.38 C). The cold months
are January (7.34 C) and February (9.26 C), and the annual mean

112

L. Zou et al. / Journal of Atmospheric and Solar-Terrestrial Physics 146 (2016) 110122

Fig. 1. Geographical location of southeast China and the distribution of meteorological and radiation stations.

Table 1
Geographic and meteorological parameters and annual mean daily GSR in 10 radiation stations.
Stations

Latitude (N)

Longitude (E)

Altitude (m)

T (C)

SSD (h)

Pa (h Pa)

Rh (%)

Pre (mm)

GSR (MJ m2)

Training

57461
57957
57993
58457
58606
59316

3042
2519
2552
3014
2836
2323

11117
11017
11500
12010
11555
11641

133.1
164.4
137.5
41.7
46.9
2.9

16.99
18.93
19.51
16.65
17.78
21.62

4.21
4.15
4.96
4.89
5.05
5.58

1002.24
994.97
999.90
1012.41
1009.91
1013.06

75.43
75.15
74.91
76.92
76.35
80.23

1138.02
1627.08
1382.46
1363.63
1533.07
1454.14

10.89
11.19
12.40
11.64
12.36
13.85

Validating

57679
58847

2812
2604

11304
11916

44.9
84.0

17.35
19.98

4.25
4.57

1008.74
1005.17

79.80
75.51

1397.64
1363.72

11.13
12.17

Testing

57494
59287

3037
2317

11408
11333

23.1
6.6

16.79
19.51

5.33
4.96

1013.43
999.90

76.68
74.91

1221.41
1382.46

12.17
11.67

daily largest Pre is observed in May (7.75 mm) and June (9.45 mm).
The driest months in the region are January (1.87 mm) and December (1.40 mm). The longest SSD time is 7.37 h in July, and the
largest Rh is found in June (80.88%).
GSR and meteorological data, including SSD, T, TM, Tm, Rh, Pre,
Pa, Pwv, and Ws, are obtained from the China Meteorological Administration (CMA) (http://data.cma.gov.cn/site/index.html). Two
types of radiation measurements are used in China. Before 1993,
the Yanishevsky thermoelectric actinometer and the Yanishevsky
thermoelectric pyranometer (imported from the Soviet Union)
were used to measure direct, diffuse, and global irradiance every
hour in China. China replaced its solar radiation measurement
instruments during 19901993. After 1993, GSR has been measured with a thermopile pyranometer, called DFY-4, manufactured
by China (Lin et al., 2016, Wang et al., 2015, Xia, 2010). According
to Shi et al. (2008), the errors in DFY-4 are less than 5% after 1993;
quality control of all measured data is performed by CMA. GSR is
further checked to ensure data quality with the detailed calibration process carried out by Shi et al. (2008) and Tang et al. (2010,
2011); the observed GSR does not exceed the extraterrestrial radiation, is larger than the lowest bound, and is 1.1 times less than

the clear-sky GSR. The likely noisy data and data with evident
systematic and operational errors are further removed. In this
study, the unit of daily solar radiation is MJ m 2, and the daily
meteorological data and solar radiation data are used to predict
and validate the GSR in empirical and ANN models after quality
control.
A 50-year period (19612010) daily surface solar radiation dataset over China is obtained from DAM (http://dam.itpcas.ac.cn/
data/daily_solar_radiation_dataset_over_China.zip). The dataset
includes the radiation data of 716 CMA stations, and it combines
observed data and the data from two models. One model is the
hybrid model proposed by Yang et al. (2001, 2006) that is used to
estimate the daily solar radiation dataset in 716 stations between
1961 and 2010. The other one is an ANN-based model, which is
trained by quality-controlled observation data from 2003 to 2010
and validated by data from 1994 to 2002. This trained ANN-based
model is then applied to estimate the daily solar radiation from
1961 to 2010 in 96 stations. According to the research of Tang et al.
(2011, 2013), the RMSEs between the original and corrected datasets in the Yellow River Basin for the same time period are
compared. The mean RMSEs are 2.01 and 1.78 MJ m 2 in the

L. Zou et al. / Journal of Atmospheric and Solar-Terrestrial Physics 146 (2016) 110122

113

Table 2
Geographical and meteorological information of solar-radiation-predicting stations.
Stations

Latitude (N)

Longitude (E)

Altitude (m)

SSD (h)

T (C)

Stations

Latitude (N)

Longitude (E)

Altitude (m)

SSD (h)

T (C)

57378
57385
57447
57476
57483
57554
57584
57598
57655
57662
57682
57745
57799
57853
57866
57872
57972
58319
58334
58402
58424
58436
58467
58506
58527

3110
3137
3017
3021
3040
2923
2923
292
2827
293
2842
2727
273
2644
2614
2653
2548
313
3119
3044
3032
3037
3012
2934
2918

11233
11348
10928
1129
11310
11010
1134
11434
11024
11141
11333
10941
11455
11038
11136
11235
1131
11657
11823
11540
1172
11859
12115
11559
11712

65.8
93.3
457.1
32.2
34.1
322.2
53.0
146.8
151.6
35.0
106.3
272.2
71.2
341.0
172.6
104.9
184.9
85.4
14.8
123.8
19.8
89.4
4.5
1164.5
61.5

5.18
5.48
3.42
4.76
4.84
3.27
4.72
4.49
3.86
4.47
4.50
3.98
4.56
3.98
4.18
4.24
3.92
4.82
5.59
5.27
5.09
5.06
5.35
4.82
5.04

16.23
15.79
16.29
16.55
16.66
16.33
17.32
16.67
16.75
17.07
16.93
16.63
18.55
16.80
17.90
18.19
18.17
16.26
16.12
16.48
16.90
15.68
16.53
11.74
17.53

58531
58549
58608
58626
58646
58665
58730
58820
58921
58926
58933
59058
59072
59082
59102
59117
59134
59265
59271
59293
59298
59453
59462
59658
59663

2942
297
283
2818
2827
2837
2745
2653
2557
2518
2655
2412
2447
2441
2457
2415
2429
2329
2338
2348
234
2238
2245
218
2149

11816
11939
11532
11713
11955
12125
1181
11710
11720
11725
11859
11030
11223
11335
11539
1165
1183
11117
11226
11444
11425
11010
11133
11017
11158

142.7
62.6
30.4
51.2
59.7
4.6
222.1
342.9
206.0
205.3
869.5
145.7
98.3
61.0
303.9
87.8
139.4
114.8
57.3
70.8
22.4
81.8
53.3
53.3
89.9

4.85
5.27
4.78
4.86
4.65
4.98
4.39
4.38
4.63
4.42
4.54
4.13
4.05
4.63
4.72
5.27
5.52
4.85
4.52
5.15
5.21
4.55
4.78
5.25
5.01

16.51
17.60
17.92
18.48
18.29
17.29
18.55
17.28
19.45
20.76
15.24
19.85
19.70
20.44
19.02
21.43
20.73
21.12
21.00
21.52
22.05
22.00
22.20
23.26
22.49

Fig. 2. Monthly variation in meteorological parameters in 10 stations.

original and corrected datasets, respectively. The new dataset is

proven more accurate than the observed dataset for its strict
quality control and model correction. Therefore, the DAM-modeled radiation dataset is suitable to validate the performance of the
ANN model proposed in this study.
2.2. Empirical models
In this work, the IAP and IBC models are developed and
compared with ANN model to estimate the GSR in southeast

China. Angstrom (1924) and Prescott (1940) propose a linear

model for investigating the linear relationship between relative
SSD and ground level and extraterrestrial solar radiation. The
model can be expressed as follows:

H
S
= a +b ,
H0
S0

(1)

where H and H0 are the horizontal surfaces at ground level and at

extraterrestrial solar radiation, respectively, S/S0 is the relative
SSD, a and b are constants. Benson et al. (1984) and Rietveld (1978)

114

L. Zou et al. / Journal of Atmospheric and Solar-Terrestrial Physics 146 (2016) 110122

Fig. 3. Schematic architecture of MLP neural network.

improve the AP model by correcting its spectral terms at different


regions. Akinoglu
and Ecevit (1990) and Aksoy (1997) use a
quadratic equation to estimate solar radiation by adding nonlinear


terms. Bahel et al. (1987) and Togrul
and Togrul
(2002) propose a
third-order polynomial nonlinear model based on the AP model
for GSR estimation. In this study, we improve the AP model by
adding nonlinear analysis and considering other meteorological
variables, such as T, Rh, and Pre. The improved model is depicted
as

S 2
S 3
H
S
= a0
+ a1 + a2 + a3 T + a4RH + a5Pre + b,
H0
S0
S0
S0

(2)

where a0 , a1, a5, and b are constants determined by changing the

values until the squared errors of the measured and estimated GSR
reach the minimum values. The constant values of the IAP model
in southeast China are 0.37, 0.34, 0.1, 0.001, 0.001, 0.001, and
0.25 for a0 to a5 and b, respectively.
Bristow and Campbell (1984) suggest a relationship between
GSR and temperature differences as follows:

H
= A 1 exp (BT C ),
H0

(3)

where T is the difference between the maximum and minimum

temperatures, A, B, and C are constants, A refers to the expected
maximum radiation in clear days, and B and C balance the inuence of temperature on solar radiation when A approaches the
threshold values. Thornton and Running (1999) reformulate the BC
model by considering clear-sky transmittance and by expanding
the effects of Rh and Pre. Winslow et al. (2001) propose the vapor
pressure radiation model on the basis of the BC model, which uses
TM/Tm, Pre, T, site latitude, and site elevation. Meza and Yebra
(2015) modify the BC model by considering T, TM/Tm, Pre, and Rh
to satisfactorily represent the atmospheric transmittance. Considering that SSD is closely correlated with land surface solar radiation, we expand the BC model by adding SSD, Pre, and Rh to
relate the effects of cloud transmittances.

Hg /H0 = b0 + b1 sin( M ) + b2 cos( M ) + b3RH + b4 Pre + b5S

1 exp( b6(TM Tm)(b7)),

)
(4)

where M=2j/365, j is the Julian day, and n is the SSD. The values of
constants b0 to b7 in the model are calculated with the same approach the constants in Eq. (2) are measured. These constants are
0.55, 0.02, 0.52, 0.004, 0.001, 0.07, 0.07, and 0.94 for b0 to b7,
respectively, in southeast China. The detailed model performances
of IAP and IBC are analyzed in Section 3.

2.3. Articial neural network

ANN is a massively parallel distributed processor, which is
composed of a large number of highly inter-connected neurons
inspired by biological neural systems. Multilayer perceptron (MLP)
NN is one of the most widely used NNs with nonlinear approximation capability to identify the nonlinear separable data (Rezrazi
et al., 2015, Rumelhart et al., 1986). This NN obtains an optimal
result based on the evaluation of the nonlinear relationship among
multiple factors (Riahi-Madvar et al., 2011). The MLP network includes three layers, namely, input, hidden, and output layers.
Neurons (nodes) are the fundamental elements in each layer, and
every neuron in one layer is associated and interacts with other
layers. The outputs of every neuron in the hidden and output
layers are determined by the previous output ( wijxj , xj is the

input signals), activation function ( f wijxj ), and weighting

coefcients (wij). The MLP network is processed as follows: the
information ow is input into the input layer and passes through
the hidden and output layers to achieve the output information.
The Tan-sigmoid activation function is used in the neurons of
hidden and output layers in this work (Fig. 3). In detail, neurons
sum the weight-controlled input production to apply the nonlinear activation function as follows:
n

ai =

wijxj ,
j=1

yi = f ( ai ) =

2
1,
1+e2ai

(5)

(6)

where wij is the weight coefcient, xj is the input signals, ai is the

summation of the weighted inputs, and yi is the output at neuron i
of the output layer.
One of the most commonly used learning algorithms, the BP
algorithm is used for network training in this study (Landeras
et al., 2012, Moghaddamnia et al., 2009). This algorithm can improve a network by back propagating the neuron bias, thereby
remedying the defect of the MLP network. The BP algorithm also
evaluates the errors between the network outputs and target to
adjust the network weight.

e=

1
S

(Ok Tk)2,
k=1

(7)

where e denotes the errors, n is the number of neurons, Ok is the

network output in the kth processor, and Tk is the target. In our
study, three datasets are considered in ANN applications, namely,
training, validating, and testing datasets. The training dataset is
employed to train a set of input data with random weights; the
output of this stage is compared with the target data and adjusts

L. Zou et al. / Journal of Atmospheric and Solar-Terrestrial Physics 146 (2016) 110122

115

the network weights accordingly until the most accurate results

are produced. The validating dataset processes new input data
with trained weights. The testing dataset is applied to check the
overall performance of a trained and validated network.
2.4. Spatial interpolation
According to Tobler (1970), all attribute values on a geographic
surface are related to one another, in which the closer values are
more strongly related than the distant ones. Thus, spatial interpolation estimates the unknown solar radiation data based on the
nearby observed data in a given area because nearby sites share
more similarities than those that are far apart (Hammer and Beyer,
2012). 10 solar radiation stations are homogeneously distributed
over the study area, which can represent the entire distribution of
GSR. The meteorological variables and GSR in these stations are
used as inputs and outputs in the ANN model to train and validate
the nonlinear relationship between meteorological parameters
and GSR in southeast China. Therefore, the trained and validated
ANN model can nearly analyze the nonlinear relationship between
meteorological parameters and GSR at any site in the study area.
Kriging interpolation is used to analyze the spatial distribution of
GSR in this study. This study aims to predict GSR by considering
the variation and distance between the data at known and unknown sites. The Kriging predictions are a linear combination of
s
known values and location weights, i.e., z0 = i = 1 zxwx , where z0 is
the predicted values, zx is the known data, and wx is the weights.

Fig. 4. Correlation between meteorological parameter and GSR.

Table 3
Input parameters and statistical errors of four different ANN networks.
Model Input variable
combination

1
2
3
4

2.5. Model performance evaluation

TTMTmPa
PwvWs
SSDRhPrePa
PwvWs
SSDRhPreT
TMTm
SSDRhPreT
TMTmPaPwv
Ws

Network R2

RMSE
(MJ
m2)

RMSE
(%)

MBE
(MJ
m2)

MBE
(%)

6101

0.88 1.77

14.80

0.39

3.28

6101

0.89 1.45

12.11

0.11

0.92

6101

0.89 1.45

12.10

0.07

0.61

9171

0.91 1.34

11.25

0.10

0.90

RMSE, MBE and R2 are used to evaluate the model accuracy in

this study, they can be expressed as:

RMSE= 1/M

i=1

MBE=1/M

M
i=1

(x m, i xe, i )2 ,

(8)

(x m,i x e,i),

(9)

R2 =

( I = 1 (x m, i xm)(xe, i xe ))2
M
i = 1 (x m, i

xm)2 i = 1 (xe, i xe )2

,
(10)

where M is the number of measurement, x m, i is the ith measured

value, and xe, i is the ith value estimated by the model.

3. Results and discussion

3.1. Application of models
Ten radiation stations are used to train, validate, and test the
ANN model in this study. Stations 57461, 57957, 57993, 58457,
58606, and 59316 are used as the training set, stations 57679 and
58847 are the validation set, and stations 57494 and 59287 are the
testing set. SSD, T, TM, Tm, Rh, Pre, Pa, Pwv, and Ws are used as
input parameters to identify the optimal model. Different combinations of meteorological variables are tried as the ANN model
input according to their correlation with GSR. Fig. 4 presents the
correlation coefcients between GSR and meteorological variables
at the 10 stations. SSD, T (T, TM, Tm), and Pwv are highly correlated
with GSR at all stations. The average coefcients are 0.87, 0.85,
0.82, 079, and 0.78 for SSD, TM, T, Tm, and Pwv, respectively. Pa is
proven to be negatively correlated with GSR at all stations. The
average coefcient of Pa is 0.74. Rh, Pre, and Ws show a relatively
low correlation with GSR as illustrated in Fig. 4.

SSD and temperatures are closely associated with GSR. Four

different combinations are proposed in Table 3. The rst model has
inputs of T, TM, Tm, Rh, Pwv, and Ws as inputs; the second model
has SSD, Rh, Pre, Pwv, and Ws as inputs; the third model has SSD, Rh,
Pre, T, TM and Tm, as inputs; the fourth model has all meteorological variables as inputs. The scatter plot of the measured and
modeled GSR of the entire dataset in the four models is shown in
Fig. 5. The RMSEs (RMSE %) are 1.77 (14.80%) MJ m 2, 1.45
(12.11%) MJ m 2, and 1.45 (12.10%) MJ m 2; the MBEs are 0.39
0.11 ( 0.92%) MJ m 2, and
0.07
( 3.28%) MJ m 2,
2
2
( 0.61%) MJ m ; the R values are 0.88, 0.89, and 0.89 for models
1 to 3, respectively. Model 4 shows the best performance on GSR
estimation with the lowest RMSE (RMSE %) [1.34 MJ m 2 (11.25%)]
and the highest R2 values (0.91). Therefore, model 4 is proven to be
the optimal ANN network for GSR estimation in southeast China.
The ANN model and improved empirical models (IAP and IBC
models) are compared to analyze their performance in our work.
The ANN model is composed of 9 inputs (i.e., SSD, Pre, Rh, T, TM,
Tm, Pa, Pwv, and Ws,), 17 hidden neurons, and 1 output. Fig. 6 shows
the scatter plot of the measured and estimated GSR in the three
models at 10 stations. The RMSEs (RMSE %) are 2.65
(22.68%) MJ m 2, 2.19 (18.39%) MJ m 2, and 1.34 (11.25%) MJ m 2;
the MBEs are 0.94 ( 7.90%) MJ m 2, 1.11 (9.34%) MJ m 2, and
0.11 ( 0.90%) MJ m 2; the R2 values are 0.68, 0.83, and 0.91 for
the IAP, IBC, and ANN models, respectively. The statistical errors
and R2 values in the ANN, IBC, and IAP models at each station are
compared in Table 4. The RMSE ranges from 1.09 MJ m 2 to
2.65 MJ m 2, from 1.49 MJ m 2 to 3.79 MJ m 2, and from
2.44 MJ m 2 to 3.03 MJ m 2; MBE ranges from 0.48 MJ m 2 to
2.1 MJ m 2, from 0.13 MJ m 2 to 3.46 MJ m 2, and from 1.38 to
0.67 MJ m 2; R2 ranges from 0.80 to 0.94, from 0.76 to 0.91, and

116

L. Zou et al. / Journal of Atmospheric and Solar-Terrestrial Physics 146 (2016) 110122

Fig. 5. Scatter plot of the measured and estimated GSR using four ANN models.

Fig. 6. Scatter plot of the measured and estimated GSR for ANN, IBC, and IAP models.

L. Zou et al. / Journal of Atmospheric and Solar-Terrestrial Physics 146 (2016) 110122

117

Table 4
Statistical errors and R2 values in the ANN, IBC, and IAP models at each station.
Stations

ANN

IBC

RMSE (MJ m2)

MBE (MJ m2)

IA-P

R2

RMSE (MJ m2)

MBE (MJ m2)

R2

RMSE (M Jm2)

MBE (MJ m2)

R2

Training

57993
58606
57461
58457
57957
59316

1.29
2.65
1.58
1.16
1.34
1.34

0.05
2.10
0.41
0.22
0.12
0.48

0.91
0.80
0.87
0.94
0.91
0.93

1.63
3.79
2.20
2.50
2.07
1.49

0.71
3.46
1.11
2.01
0.81
0.53

0.89
0.81
0.85
0.90
0.86
0.92

2.84
2.50
2.54
2.77
2.55
2.51

0.98
0.74
0.67
0.84
1.04
0.78

0.61
0.57
0.67
0.69
0.72
0.78

Validating

57679
58847

1.21
1.64

0.44
0.09

0.88
0.88

2.12
2.01

1.03
0.72

0.76
0.84

2.48
3.03

1.38
1.14

0.64
0.68

Testing

57494
59287

1.09
1.27

0.01
0.01

0.94
0.91

1.72
1.54

0.74
0.13

0.91
0.90

2.44
2.76

1.09
0.75

0.78
0.64

Fig. 7. Scatter plot of the monthly mean daily measured and ANN-estimated GSR at 10 stations.

from 0.57 to 0.78 in the ANN, IBC, and IAP models, respectively.
The average RMSEs are 1.4, 2.11, and 2.64 MJ m 2; the average
MBEs are 0.07, 0.86, and 0.94 MJ m 2; the average R2 values are
0.90, 0.86, and 0.86 in the ANN, IBC, and IAP models, respectively.
These results indicate that the ANN model is better than the IBC
and IAP models. The scatter plots of the measured and estimated

monthly mean GSR by the ANN model at 10 stations are shown in

Fig. 7. The correlation coefcients range from 0.92 to 0.97 at the 10
stations in the ANN model. We conclude that ANN is well trained
with six stations and shows good modeling results at validating
and testing stations. Therefore, the ANN model can be used to
predict the GSR at other sites in southeast China.

118

L. Zou et al. / Journal of Atmospheric and Solar-Terrestrial Physics 146 (2016) 110122

Fig. 8. Scatter plot of monthly mean daily (a) measured and ANN-predicted GSR and (b) DAM-modeled and ANN-predicted GSR during 19612010.

3.2. GSR in southeast china

We predict the daily GSR at 50 stations using model 4. A total of
60 stations are then used to analyze the regional GSR in southeast
China. To validate the accuracy of the ANN-modeled GSR, we
compare it with the measured and DAM-modeled values in Fig. 8.
The scatter plot of the monthly mean GSR at ANN and measuring
sites is shown in Fig. 8a. The correlation coefcient and R2 are
larger than 0.95, the RMSE (%) is 1.07 MJ m 2 (8.91%), and the MBE
(%) is 0.62 MJ m 2 ( 5.21%). Fig. 8b presents the scatter plot of
the monthly mean GSR in the ANN and DAM models. The correlation coefcient and R2 are larger than 0.95, the RMSE (%) is
0.91 MJ m 2 (7.28%), and the MBE (%) is 0.15 MJ m 2 ( 1.20%).
The results of the ANN model have high accordance with the
measured and DAM-modeled values because high correlation and
low errors are observed between them. Therefore, the proposed
ANN model has a great performance in solar radiation estimation
in southeast China.
In addition to scatter plot analysis, we compare the spatial
characteristics of different GSR in this study. The interpolation of
the annual mean daily GSR is shown in Fig. 9. Fig. 9a presents the
interpolation distribution of the measured GSR at 10 stations. The
annual mean daily GSR ranges from 10.90 MJ m 2 to 13.51 MJ m 2
in this area; the lowest value is located in the northwest area. The
spatial distribution of the annual mean daily GSR reveals an increasing trend from northwest to southeast. Most areas range from
11 MJ m 2 to 13 MJ m 2, and the regional annual mean daily GSR
is 11.95 MJ m 2. Fig. 9b indicates the interpolation distribution of
DAM-modeled GSR at 56 stations in southeast China. The annual
mean daily GSR ranges from 10.16 MJ m 2 to 14.23 MJ m 2. Similar to Fig. 9a, the lowest value of Fig. 9b is in the northwest area,
and the annual mean daily GSR increases from northwest to
southeast. By contrast, the south area in the Guangdong Province
receives more solar radiation; it is larger than 13 MJ m 2. The

regional annual mean daily GSR is 12.33 MJ m 2, and most areas

range from 12 MJ m 2 to 13 MJ m 2.
Fig. 9c shows the annual mean daily GSR interpolated by the
ANN-estimated values at 60 stations. The distribution of the annual mean daily GSR is generally similar to that of the measured
and DAM-modeled ones. The lowest value is obtained in the
northwest area, and the spatial distribution of the annual mean
daily GSR shows an increasing trend from northwest to southeast.
Contrarily, a huge amount of solar radiation is received in the west
and south areas of the Fujian Province, in which GSR exceeds
13 MJ m 2. GSR exceeds 14 MJ m 2 in the south coastal area,
whereas relatively low values are observed in the east Zhejiang
Province. The regional annual mean daily GSR is 12.58 MJ m 2,
and most of the areas present GSR that ranges from 12 MJ m 2 to
14 MJ m 2. According to the MBE in Table 4, the ANN model
overestimates GSR in some areas, but it generally shows high estimation precision. Fig. 9b and c show detailed distribution characteristics when more stations are detected. The spatial distribution of the GSR modeled by ANN is highly consistent with the
results obtained by Tang et al. (2010, 2013). The spatial distribution
of GSR is uneven in this area; it gradually decreases from east to
west and from south to north. Latitude and distance from the sea
are the main factors that inuence the distribution of solar radiation. Terrain and economy are also important factors affecting
the local differences in solar radiation.
Fig. 10 shows the seasonal and annual mean daily GSR in
southeast China. The annual mean GSR is 12.58 MJ m 2 during
19612010. According to the smoothed and t lines in Fig. 10,
southeast China experienced a dimming phenomenon during 1961
1990 when the GSR decreased by 0.05 MJ m 2 decade 1, whereas
the GSR showed a stable trend from 1991 to 2010. The variation is
consistent with that indicated in the studies of Xia (2010) and Tang
et al. (2011), that is, dimming occurred over southern China during
19601989 and a transition happened thereafter. A similar trend

L. Zou et al. / Journal of Atmospheric and Solar-Terrestrial Physics 146 (2016) 110122

119

Fig. 9. Spatial distribution of the annual mean daily measured, DAM-developed, and ANN-estimated GSR during 19612010 in southeast China.

was observed in seasonal GSR variation; it decreased from 1960 to

1990 and increased later. The seasonal mean daily GSR values are
11.73, 16.75, 13.15, and 8.68 MJ m 2 day 1 in spring, summer, autumn, and winter, respectively, in southeast China.
The long-term trend slopes of the ANN-estimated GSR from
1961 to 2010 at 60 stations are shown in Fig. 11. Solid blue cycles
are negative values, which represent a decrease trend of GSR;
double cycles are positive values that imply an increase trend of
GSR in stations; red star denotes that the variation is signicant at
the 95% level. Of 60 stations, 52 are negative in southeast China
from 1961 to 2010, and 11 stations are signicant at the 95% level.
The trend slopes are positive at eight stations, and four of them are
signicant at the 95% level. Overall, southeast China experienced a
decrease in solar radiation from 1961 to 2010. The rapid economic
development and increasing aerosol emission may explain this
decrease in southeast China according to Wang et al. (2009) and
Tang et al. (2011).

4. Conclusion
In this study, an MLP NN using the BP learning algorithm based
on spatial interpolation is proposed to estimate the GSR in
southeast China. Two improved empirical models are compared
with the ANN model to explore the best model for solar radiation
modeling. The measured meteorological data at 60 stations and
the solar radiation data at 10 stations obtained from the CMA are
used as inputs and outputs in all models. A high-accuracy daily
surface solar radiation dataset during 19612010 developed by
DAM is used to validate the ANN-estimated values in this study.
Different input combinations of meteorological parameters
from 10 solar radiation stations, such as SSD, T, TM, Tm, Rh, Pre, Pa,
Pwv, and Ws, are used to investigate the optimal ANN network.
Among four different network models, model 4 with the network
of 9-17-1 shows the best performance in estimating GSR. The
lowest RMSE (RMSE %) (1.34 MJ m 2 (11.25%)) and the highest R2
values (0.91) are observed in model 4. Compared with two

120

L. Zou et al. / Journal of Atmospheric and Solar-Terrestrial Physics 146 (2016) 110122

Fig. 10. Variations in annual and seasonal mean daily GSR in southeast China during 19612010; the blue line is the smoothed line of variation, and the red lines are the t
lines of variation during 19611989 and 19902010.

Fig. 11. Spatial distribution of long-term trend slopes of the ANN-model-estimated

GSR at 60 stations.

improved empirical models, model 4 has higher accuracy in predicting GSR at 10 stations. The RMSE, MBE, and R2 in the IAP
model are 2.65 MJ m 2, 0.94 MJ m 2, and 0.68, respectively, the
RMSE, MBE, and R2 in the IBC model are 2.19 MJ m 2, 1.11 MJ m 2,
and 0.83, respectively, and the RMSE, MBE, and R2 in the ANN
model are 1.34 MJ m 2,
0.11 MJ m 2, and 0.91, respectively.
Hence, the proposed ANN model is well trained and shows high
accuracy in GSR estimation.
The ANN model is used to estimate the GSR at 50 meteorological stations. The regional monthly mean GSR in measured
dataset, DAM dataset, and ANN model is analyzed. The statistical

errors between the measured and ANN-estimated monthly mean

daily GSR are 1.07 MJ m 2 (8.91%) for RMSE (RMSE %) and
0.62 MJ m 2 ( 5.21%) for MBE (MBE %). The statistical errors between DAM and ANN are 0.91 MJ m 2 (7.28%) for RMSE (RMSE %)
and 0.15 MJ m 2 ( 1.20%) for MBE (MBE %). The spatial distribution of GSR is also investigated. The lowest value is observed in the
northwest area, and it increases from northwest to southeast. The
regional annual mean GSR is 12.58 MJ m 2, and most of the areas
present GSR ranging from 12 MJ m 2 to 14 MJ m 2. The annual
mean GSR decreases by 0.02 MJ m 2 decade 1 over the entire
southeast China. Regarding the long-term trend of GSR at every
single site, 52 out of 60 stations show a decreasing trend in southeast
China from 1961 to 2010, and 11 stations are signicant at the 95%
level. The rapid economic development and increasing aerosol
emission may contribute to the decline in the GSR in southeast China.
The proposed ANN model shows a good performance in GSR
estimation, which improves our understanding of the long-term
variations in the GSR in southeast China. The ANN based on spatial
interpolation is of vital importance to accurately estimate the
spatialtemporal characteristics of GSR. This model can be used to
predict GSR at any region without solar radiation measurements in
the world. However, the application of this model needs measured
GSR data to train the network. Therefore, further analysis will focus on the inuence radius of solar radiation site and the accuracy
of this model in different spatial scales.

Acknowledgments
This work was supported by the National Natural Science
Foundation of China (Program: No. 41301586). The geographical,

L. Zou et al. / Journal of Atmospheric and Solar-Terrestrial Physics 146 (2016) 110122

meteorological, and solar radiation parameters used in this study

were obtained from the China Meteorological Administration/National Meteorological Center. We appreciate the daily surface solar
radiation dataset over China provided by DAM. We also express
our sincere gratitude to the editor and reviewers.

References

Akinoglu,
B.G., Ecevit, A., 1990. Construction of a quadratic model using modied
ngstrom coefcients to estimate global solar radiation. Sol. Energy 45, 8592.
Aksoy, B., 1997. Estimated monthly average global radiation for Turkey and its
comparison with observations. Renew. Energy 10, 625633.
Almorox, J., Hontoria, C., 2004. Global solar radiation estimation using sunshine
duration in Spain. Energy Convers. Manag. 45, 15291535.
Amrouche, B., Le Pivert, X., 2014. Articial neural network based daily local forecasting for global solar radiation. Appl. Energy 130, 333341.
Angstrom, A., 1924. Solar and terrestrial radiation. Report to the international
commission for solar research on actinometric investigations of solar and atmospheric radiation. Q. J. R. Meteor. Soc. 50, 121126.
Antonanzas-Torres, F., Martnez-De-Pisn, F.J., Antonanzas, J., Perpinan, O., 2014.
Downscaling of global solar irradiation in complex areas in R. J. Renew. Sustain.
Energy 6, 063105.
Bahel, V., Bakhsh, H., Srinivasan, R., 1987. A correlation for estimation of global
solar-radiation. Energy 12, 131135.
Bakirci, K., 2009. Models of solar radiation with hours of bright sunshine: A review.
Renew. Sustain. Energy Rev. 13, 25802588.
Benson, R.B., Paris, M.V., Sherry, J.E., Justus, C.G., 1984. Estimation of daily and
monthly direct, diffuse and global solar-radiation from sunshine duration
measurements. Sol. Energy 32, 523535.
Bindi, M., Miglietta, F., 1991. Estimating daily global radiation from air temperature
and rainfall measurements. Clim. Res. 1, 117124.
Bristow, K.L., Campbell, G.S., 1984. On the relationship between incoming solar
radiation and daily maximum and minimum temperature. Agric. For. Meteorol.
31, 159166.
Budyko, M.I., 1969. The effect of solar radiation variations on the climate of the
Earth. Tellus 21, 611619.
Chen, J.L., Li, G.S., 2013. Estimation of monthly average daily solar radiation from
measured meteorological data in Yangtze River Basin in China. Int. J Climatol.
33, 487498.
Chen, M., Zhuang, Q.L., He, Y.J., 2014. An efcient method of estimating downward
solar radiation based on the MODIS observations for the use of land surface
modeling. Remote. Sens.-Basel 6, 71367157.
Davies, J.A., Schertzer, W., Nunez, M., 1975. Estimating global solar radiation.
Bound.-Layer. Meteorol. 9, 3352.
Ehnberg, J.S.G., Bollen, M.H.J., 2005. Simulation of global solar radiation based on
cloud observations. Sol. Energy 78, 157162.
Hammer, A., Beyer, H.G., 2012. Solar Radiation, Spatial and Temporal Variability. In:
Meyers, R.A. (Ed.), Encyclopedia of Sustainability Science and Technology.
Springer, New York, New York, NY, pp. 97449758.
Hay, J.E., Suckling, P.W., 1979. An assessment of the networks for measuring and
modelling solar radiation in British Columbia and adjacent areas of western
Canada. Can. Geogr. 23, 222238.
Hook, J.E., Mcclendon, R.W., 1992. Estimation of solar-radiation data missing from
long-term meteorological records. Agron. J 84, 739742.
Jiang, Y., 2009. Computation of monthly mean daily global solar radiation in China
using articial neural networks and comparison with other empirical models.
Energy 34, 12761283.
Journee, M., Bertrand, C., 2010. Improving the spatio-temporal distribution of surface solar radiation data by merging ground and satellite measurements. Remote Sens. Environ. 114, 26922704.
Kashyap, Y., Bansal, A., Sao, A.K., 2015. Solar radiation forecasting with multiple
parameters neural networks. Renew. Sust. Energy Rev. 49, 825835.
Lam, J.C., Wan, K.K.W., Yang, L., 2008. Solar radiation modelling using ANNs for
different climates in China. Energy Convers. Manag. 49, 10801090.
Landeras, G., Lopez, J.J., Kisi, O., Shiri, J., 2012. 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. 62, 113.
Lin, A., Zou, L., Wang, L., Gong, W., Zhu, H., Salazar, G.A., 2016. Estimation of atmospheric turbidity coefcient over Zhengzhou, China during 1961-2013
using an improved hybrid model. Renew. Energy 86, 11341144.
Liu, J.D., Linderholm, H., Chen, D.L., Zhou, X.J., Flerchinger, G.N., Yu, Q., Du, J., Wu, D.
R., Shen, Y.B., Yang, Z.B., 2015. Changes in the relationship between solar radiation and sunshine duration in large cities of China. Energy 82, 589600.
Meza, F., Varas, E., 2000. Estimation of mean monthly solar global radiation as a
function of temperature. Agric. For. Meteorol. 100, 231241.
Meza, F.J., Yebra, M.L., 2015. Estimation of daily global solar radiation as a function
of routine meteorological data in Mediterranean areas. Theor. Appl. Climatol. .
http://dx.doi.org/10.1007/s00704-015-1519-6
Moghaddamnia, A., Remesan, R., Kashani, M.H., Mohammadi, M., Han, D., Piri, J.,
2009. Comparison of LLR, MLP, Elman, NNARX and ANFIS Models-with a case
study in solar radiation estimation. J. Atmos. Sol.-Terr. Phys. 71, 975982.

121

Mohandes, M., Balghonaim, A., Kassas, M., Rehman, S., Halawani, T.O., 2000. Use of
radial basis functions for estimating monthly mean daily solar radiation. Sol.
Energy 68, 161168.
Mohandes, M., Rehman, S., 2010. Global solar radiation maps of Saudi Arabia. J.
Energy Power Eng. 4, 5763.
Mohandes, M., Rehman, S., Halawani, T.O., 1998. Estimation of global solar radiation
using articial neural networks. Renew. Energy 14, 179184.
Muneer, T., Gul, M.S., 2000. Evaluation of sunshine and cloud cover based models
for generating solar radiation data. Energy Convers. Manag. 41, 461482.
Panday, C.K., Katiyar, A.K., 2010. Temperature base correlation for the estimation of
global solar radiation on horizontal surface. Int. J. Energy Environ. 1, 737744.
Pinker, R.T., Frouin, R., Li, Z., 1995. A review of satellite methods to derive surface
shortwave irradiance. Remote Sens. Environ. 51, 108124.
Polo, J., Zarzalejo, L.F., Cony, M., Navarro, A.A., Marchante, R., Martn, L., Romero, M.,
2011. Solar radiation estimations over India using Meteosat satellite images.
Sol. Energy 85, 23952406.
Prescott, J., 1940. Evaporation from a water surface in relation to solar radiation.
Trans. R. Soc. South Aust. 64, 114118.
Qin, J., Chen, Z., Yang, K., Liang, S., Tang, W., 2011. Estimation of monthly-mean daily
global solar radiation based on MODIS and TRMM products. Appl. Energy 88,
24802489.
Ramanathan, V., Crutzen, P.J., Kiehl, J.T., Rosenfeld, D., 2001. Atmosphere: Aerosols,
climate, and the hydrological cycle. Science 294, 21192124.
Rehman, S., 1999. Empirical model development and comparison with existing
correlations. Appl. Energy 64, 369378.
Rehman, S., Ghori, S.G., 2000. Spatial estimation of global solar radiation using
geostatistics. Renew. Energy 21, 583605.
Rehman, S., Mohandes, M., 2008. Articial neural network estimation of global
solar radiation using air temperature and relative humidity. Energy Policy 36,
571576.
Rehman, S., Mohandes, M., 2009. Estimation of diffuse fraction of global solar radiation using articial neural networks. Energy Sources Part A: Recovery Util.
Environ. Eff. 31, 974984.
Rezrazi, A., Hanini, S., Laidi, M., 2015. An optimisation methodology of articial
neural network models for predicting solar radiation: a case study. Theor. Appl.
Climatol. 123, 769783.
Riahi-Madvar, H., Ayyoubzadeh, S.A., Atani, M.G., 2011. Developing an expert system for predicting alluvial channel geometry using ANN. Expert. Syst. Appl. 38,
215222.
Richardson, C.W., Wright, D.A., 1984. WGEN: A model for generating daily weather
variables, ARS-8. US Department of Agriculture, Agricultural Research Service,
Washington, DC, USA.
Rietveld, M.R., 1978. A new method for estimating the regression coefcients in the
formula relating solar radiation to sunshine. Agric. Meteorol. 19, 243252.
Rivington, M., Bellocchi, G., Matthews, K.B., Buchan, K., 2005. Evaluation of three
model estimations of solar radiation at 24 UK stations. Agr. For. Meteorol. 132,
228243.
Rivington, M., Matthews, K.B., Bellocchi, G., Buchan, K., 2006. Evaluating uncertainty introduced to process-based simulation model estimates by alternative sources of meteorological data. Agric. Syst. 88, 451471.
Rumelhart, D.E., Hinton, G.E., Williams, R.J., 1986. Learning Representations by
Back-Propagating Errors. Nature 323, 533536.
Shi, G.Y., Hayasaka, T., Ohmura, A., Chen, Z.H., Wang, B., Zhao, J.Q., Che, H.Z., Xu, L.,
2008. Data quality assessment and the long-term trend of ground solar radiation in China. J. Appl. Meteorol. Clim. 47, 10061016.
Suckling, P.W., 1985. Estimating daily solar radiation values in selected mid-latitude
regions by extrapolating measurements from nearby stations. Sol. Energy 35,
491495.
Supit, I., Van Kappel, R.R., 1998. A simple method to estimate global radiation. Sol.
Energy 63, 147160.
Tang, W., Qin, J., Yang, K., Niu, X., Zhang, X., Yu, Y., Zhu, X., 2013. Reconstruction of
daily photosynthetically active radiation and its trends over China. J. Geophys.
Res. Atmos. 118, 1329213302.
Tang, W., Yang, K., He, J., Qin, J., 2010. Quality control and estimation of global solar
radiation in China. Sol. Energy 84, 466475.
Tang, W.J., Yang, K., Qin, J., Cheng, C.C.K., He, J., 2011. Solar radiation trend across
China in recent decades: A revisit with quality-controlled data. Atmos. Chem.
Phys. 11, 393406.
Tang, W.J., Yang, K., Qin, J., Min, M., 2013. Development of a 50-year daily surface
solar radiation dataset over China. Sci. China Earth Sci. 56, 15551565.
Teke, A., Yildirim, H.B., elik, ., 2015. Evaluation and performance comparison of
different models for the estimation of solar radiation. Renew. Sustain. Energy
Rev. 50, 10971107.
Thornton, P.E., Hasenauer, H., White, M.A., 2000. Simultaneous estimation of daily
solar radiation and humidity from observed temperature and precipitation: An
application over complex terrain in Austria. Agric. For. Meteorol. 104, 255271.
Thornton, P.E., Running, S.W., 1999. An improved algorithm for estimating incident
daily solar radiation from measurements of temperature, humidity, and precipitation. Agric. For. Meteorol. 93, 211228.
Tobler, W.R., 1970. A computer movie simulating urban growth in the Detroit region. Econ. Geogr. 46, 234240.


Togrul,
I.T., Togrul,
H., 2002. Global solar radiation over Turkey: Comparison of
predicted and measured data. Renew. Energy 25, 5567.
Tymvios, F.S., Jacovides, C.P., Michaelides, S.C., Scouteli, C., 2005. Comparative study
of ngstrm's and articial neural networks' methodologies in estimating
global solar radiation. Sol. Energy 78, 752762.

122

L. Zou et al. / Journal of Atmospheric and Solar-Terrestrial Physics 146 (2016) 110122

Vignola, F., Harlan, P., Perez, R., Kmiecik, M., 2007. Analysis of satellite derived beam
and global solar radiation data. Sol. Energy 81, 768772.
Wang, K., Dickinson, R.E., Liang, S., 2009. Clear sky visibility has decreased over land
globally from 1973 to 2007. Science 323, 14681470.
Wang, K., Ma, Q., Li, Z., Wang, J., 2015. Decadal variability of surface incident solar
radiation over China: Observations, satellite retrievals, and reanalyses. J. Geophys. Res. Atmos. 120, 65006514.
Wang, L., Kisi, O., Zounemat-Kermani, M., Salazar, G.A., Zhu, Z., Gong, W., 2016.
Solar radiation prediction using different techniques: model evaluation and
comparison. Renew. Sustain. Energy Rev. 61, 384397.
Wilks, D.S., 1999. Simultaneous stochastic simulation of daily precipitation, temperature and solar radiation at multiple sites in complex terrain. Agric. For.
Meteorol. 96, 85101.
Winslow, J.C., Hunt Jr, E.R., Piper, S.C., 2001. A globally applicable model of daily
solar irradiance estimated from air temperature and precipitation data. Ecol.
Model. 143, 227243.

Xia, X., 2010. A closer looking at dimming and brightening in China during 1961
2005. Ann. Geophys.-Ger. 28, 11211132.
Yang, K., Huang, G.W., Tamai, N., 2001. A hybrid model for estimating global solar
radiation. Sol. Energy 70, 1322.
Yang, K., Koike, T., Ye, B., 2006. Improving estimation of hourly, daily, and monthly
solar radiation by importing global data sets. Agric. For. Meteorol. 137,
4355.
Yaniktepe, B., Genc, Y.A., 2015. Establishing new model for predicting the global
solar radiation on horizontal surface. Int. J Hydrog. Energy 40, 1527815283.
Zhang, H., Xin, X., Li, L., Liu, Q., 2015a. Estimating global solar radiation using a
hybrid parametric model from MODIS data over the Tibetan Plateau. Sol. Energy 112, 373382.
Zhang, Y., Li, X., Bai, Y., 2015b. An integrated approach to estimate shortwave solar
radiation on clear-sky days in rugged terrain using MODIS atmospheric products. Sol. Energy 113, 347357.