Renewable and Sustainable Energy Reviews 61 (2016) 384–397

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews

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

Solar radiation prediction using different techniques: model evaluation

and comparison
Lunche Wang a,n, Ozgur Kisi b, Mohammad Zounemat-Kermani c, Germán Ariel Salazar d,
Zhongmin Zhu e,f, Wei Gong f,g
a
Laboratory of Critical Zone Evolution, School of Earth Sciences, China University of Geosciences, Wuhan 430074, China
b
Canik Basari University, Faculty of Architecture and Engineering, Civil Engineering Department, Samsun, Turkey
c
Department of Water Engineering, Shahid Bahonar University of Kerman, Kerman, Iran
d
Department of Physics, School of Exact Sciences, National University of Salta, Bolivia Avenue #5150, 4408 FVY Salta Capital, Argentina
e
Huazhong University of Science and Technology Wuchang Branch, Wuhan 430064, China
f
State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan, Hubei Province 430079, China
g
Collaborative Innovation Center for Geospatial Technology, Wuhan 430079,China

art ic l e i nf o a b s t r a c t

Article history: Daily observations of meteorological parameters, air temperature, air pressure, relative humidity, water
Received 27 December 2015 vapor pressure and sunshine duration hours observed at 12 stations in different climatic zones during
Received in revised form 1961–2014 are reported for testing, validating and comparing different solar radiation models. Three
12 March 2016
types of Artificial Neural Network (ANN)methods, Multilayer Perceptron (MLP), Generalized Regression
Accepted 7 April 2016
Neural Network (GRNN) and Radial Basis Neural Network (RBNN) are applied in this study for predicting
Available online 22 April 2016
the daily global solar radiation (Hg) using above meteorological variables as model inputs. The Bristow-
Keywords: Campbell model has also been improved by considering the factors influencing the incoming solar
Solar radiation radiation, such as relative humidity, cloud cover, etc. The results indicate that there are large differences
Generalized regression neural network
in model accuracies for each model at different stations, the ANN models can estimate daily Hg with
Multilayer perceptron
satisfactory accuracy at most stations in different climate zones, and MLP and RBNN models provide
Radial basis neural network
Improved Bristow-Campbell model better accuracy than the GRNN and IBC models, for example, the MAE and RMSE values range 1.53–2.29
Model evaluation and 1.94-3.27 MJ m
2 day
1, respectively for MLP model. The model performances also show some
differences at different stations for each model, for example, the RMSE values from MLP model are 1.94
and 3.27 MJ m
2 day
1at NN and HZ stations, respectively. Meanwhile, ANN models underestimate few
high radiation values at some stations, which may due to the differences in training and testing data
ranges and distributions of the stations. Finally, the differences in model performances from different
solar radiation models have been further analyzed.
& 2016 Elsevier Ltd. All rights reserved.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
2. Materials and methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
2.1. Sites and data processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
2.2. Solar radiation prediction models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388
2.2.1. Multilayer perceptron neural network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388
2.2.2. Radial basis neural network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
2.2.3. Generalized regression neural network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
2.2.4. Improved Bristow–Campbell model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
3. Model applications and results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
3.1. Comparisons of measures of fit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
3.2. Model performances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390

n
Corresponding author. Tel.: þ 86 13349889828.
E-mail addresses: wang@cug.edu.cn, lunchewang@whu.edu.cn (L. Wang).

http://dx.doi.org/10.1016/j.rser.2016.04.024
1364-0321/& 2016 Elsevier Ltd. All rights reserved.
 L. Wang et al. / Renewable and Sustainable Energy Reviews 61 (2016) 384–397 385

4. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
5. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396

1. Introduction elaborated in Schmetz [29], Noia et al. [30], Pinker et al. [31] and
Perez et al. [32]. Rigollier et al. [33] demonstrated a clear-sky
Solar radiation reaching the Earth's surface plays an important model, which was developed in the framework of the new digital
role in the energy balances of numerous physical, chemical, and European Solar Radiation Atlas and compared with the Heliosat
biological processes [1–3]. The changes in the amount of solar method. The above model was validated as one of the most
radiation greatly influences the fluxes of sensible and latent heat, accurate with respect to robustness and accuracy because it con-
the hydrological cycle, terrestrial ecological ecosystems and the sidered the Linke turbidity factor and the elevation of the sites.
climate [4,5]. Meanwhile, the solar energy has a much lower Kambezidis et al. [34] reported the recent improvements of the
environmental pollution than the conventional sources like fossil meteorological radiation model in predicting solar radiation under
fuels [6], and it is the most abundant of all renewable and sus- all-sky conditions at Athens, Greece, which indicated that the
tainable energy resources at places around the world, which can inclusion of the aerosol properties in the radiation model can
be harnessed for commercial uses through large solar array farms significantly improve the estimations. Shamim et al. [35] pre-
to meet the global energy challenges [7,8]. Thus, accurate deter- sented an improved technique (Mesoscale meteorological model)
mination and clear understanding of the spatial-temporal varia- that utilizes information from a numerical weather prediction
bility of solar radiation is of great importance to meteorological model for determining the cloud cover index and solar radiation at
and hydrological processes, photosynthesis, ecological functions, Brue catchment situated in the southwest of England. The results
agricultural and industrial production, energy development and clearly showed an improvement in the estimated Hg in compar-
utilization [9,10]. ison to the prevailing approach.
Though Meteonorm version 6.0 is a global climatological Meanwhile, artificial intelligence is a particularly promising
database designed for planners of active solar systems like PV approach for modeling solar radiation variation in recent years
plants or solar thermal systems, which contains monthly mean [36,37], a number of ANN methods have been optimized for esti-
values of Hg of several databases [11,12], the radiation data has not mating solar radiation in different regions of the world [38,39].
been routinely observed at most meteorological stations around Olatomiwa et al. [40] developed an adaptive neuro-fuzzy approach
the world due to the high instrument cost and technical require- for predicting solar radiation in Nigeria using TM, Tm and h. Park
ments [13], for example, the ratio between stations observing solar et al. [8] tried to estimate the spatial distribution of solar radiation
radiation and those observing Ta is lower than 1:100 in America using topographic factor and h in South Korea. Aguiaret al. [41]
[14]. Therefore, developing and applying proper methods to esti- proposed the Markov transitions matrix approach for estimating
mate solar radiation has been the focus of numerous studies in daily radiation values using only the clearness index as input.
locations without direct radiation measurements in recent years Aguiar and Collares-Pereira [42] also developed a time-dependent,
[15,16]. One of the most widely used methods is to establish the autoregressive, Gaussian model for generating synthetic hourly
relationships between solar radiation and other measured radiation, which has been widely used and modified in predicting
meteorological parameters such as Ta, h and water vapor contents solar radiation [43]. Amrouche and Pivert [44] predicted daily G
[17,18], for example, Yacef et al. [19] estimated the daily Hg from Ta with satisfactory accuracy at two sites in France using combined
in Algeria; Li et al. [20] calculated the Hg in Tibet, China from h. spatial modeling and ANN techniques. Olatomiwa et al. [45]
Among the temperature-based models, the BC model can relate developed an efficient support vector machines firefly algorithm,
diurnal air temperature range (TM
Tm) to incoming solar radia- ANN and Genetic Programming models for estimating solar
tion, which has been widely used for modeling solar energy [21], radiation at the Iranian city. Linares-Rodríguez et al. [46] applied
for example, Almorox et al. [22] estimated the daily Hg from ANN for predicting solar radiation in Spain based on latitude,
measured Ta at Cañada de Luque, Córdoba, Argentina. Due to the longitude, day of the year and general climatic parameters, and the
effects of geographical, meteorological and terrestrial factors results showed that RMSE values were in the range of 13.52–14.2%.
(albedo, aerosol, cloudiness, etc.), the amount of solar radiation Emad et al. [47] predicted the monthly average Hg using ANN
reaching the surface are greatly affected, above empirical models model in Qena, Upper Egypt, the RMSE and R2 values were
should be recalibrated [23], for example, the Ångström-Prescott 115 Wh/m2 and 0.977, respectively. Shamshirband et al. [48] pro-
model has been modified to the quadratic, cubic, exponential and posed a hybrid support vector machine-firefly optimization
logarithmic forms in many places of the world [24]. method for estimating monthly mean Hg in Iran, the results
The physical radiation models take into considerations of revealed that this method was greatly capable to give favorable
radiative transferring process (aerosol absorption and scattering), predictions with much higher precision than other examined
which is proved as an effective method for predicting solar methods. Rizwan et al. [49] used fuzzy logic technique to estimate
radiation around the world, for example, Pyrina et al. [25] inves- monthly mean Hg in four Indian stations using different input
tigated the cloud effects on the shortwave, longwave and all-wave data. They reported that the developed model was accurate since
radiation budget of the Mediterranean basin. Gueymard et al. [26] the amounts of obtained errors are limited. Bhardwaj et al. [50]
developed an atmospheric transmittance model for calculating the introduced a hybrid approach which includes hidden Markov
clear-sky beam, diffuse and global photosynthetically active models and generalized fuzzy models to prediction solar irradia-
radiation. Yang et al. [27] estimated the hourly, daily and monthly tion in India. The results indicated that the predicted values
solar radiation by importing global data sets using a hybrid model, obtained using the proposed model are in favorable agreements
which was also validated as one of the best broadband radiation with the measured data. Aguiar et al. [51] employed a library of
models [28]. Hybrid models that coupled both the physical and Markov transition matrices, each corresponding to a specific
empirical aspects have also been developed over the years as interval in clearness indices, and explained how they were used
386 L. Wang et al. / Renewable and Sustainable Energy Reviews 61 (2016) 384–397
Nomenclature
ANN Artificial Neural Network
MLP Multilayer Perceptron
GRNN Generalized Regression Neural Network
RBNN Radial Basis Neural Network
BC Bristow-Campbell model
IBC Improved Bristow-Campbell model
ANFIS Adaptive network-fuzzy inference system
Ta Air temperature
PA Air pressure
TM Maximum air temperature
G0 extraterrestrial radiation
for generating radiation sequences. Mohanty et al. [52] reviewed
and compared different models and techniques used for prediction
of solar radiation using general climatic parameters. Mohammadi
et al. [53] investigated the potential of ANFIS for predicting daily
Hg by day of the year. Yadav and Chandel [54] reviewed different
ANN techniques to identify suitable methods for forecasting solar
radiation in literature. The results indicated that the prediction
accuracy was dependent on input parameter combinations,
training algorithm and architecture configurations.
Moreover, satellite imagines are also widely used for studying
the spatial-temporal variations of solar radiation around the world
[55], for example, Hay [56] introduced the modeling approaches
for satellite based estimates of solar irradiance at the Earth's sur-
face. Cano et al. [57] proposed a method for determining Hg from
meteorological satellite data. Cros et al. [58] simulated the
Meteosat-7 broadband radiances using two visible channels of
Meteosat-8. Polo et al. [59] estimated the solar radiation over India
using Meteosat satellite images. Antonanzas-Torres et al. [60]
conducted the comparative assessment of Hg from a satellite
estimate model and on-ground measurements in Spain. Zhang
et al. [61] also developed an integrated algorithm to estimate
shortwave solar radiation on clear-sky days in rugged terrain using
aerosol optical depth and precipitable water vapor from MODIS
atmospheric products. Quesada-Ruiz et al. [62] developed an
optimized ANN ensemble model to derive hourly Hg estimates
from Meteosat Second Generation imagery, the results obtained
with the proposed model reduced the RMSE value of the Heliosat-
2 model a 22% for all-sky conditions and a 42% for overcast
conditions.
Some researchers compared and analyzed the model perfor-
mances for modeling solar radiation using different techniques, for
example, Kumar et al. [63] compared the regression and ANN
models for estimation of Hg; Rahimikhoob et al. [64] compared the
statistical and ANN's methodologies for deriving Hg from NOAA
satellite images; Polo et al. [65] analyzed the sensitivity of
satellite-based methods for deriving solar radiation to different
choice of aerosol input and models. Ahmad and Tiwari [66]
reviewed different solar radiation models and it was found that
the Collares-Pereira and Rabl model as modified by Gueymard
provided the best accuracy for estimating mean hourly Hg under
clear sky conditions for Indian regions, and the Ertekin and Yaldiz
model yielded the best accuracy against measured data of Konya,
Turkey. Citakoglu [67] also compared the accuracies of the ANFIS,
ANN, and MLR models, and of Ångström, Abdalla, Bahel and
Hargreaves-Samani empirical equations, the final results indicated
that the ANN model generally performed superior to the ANFIS,
MLR and the empirical equations in estimating monthly Hg at 163
stations in Turkey. Engerer and Mills [68] compared nine of the
most prominent beam and global clear sky radiation models at 14
B biases
Hg Global solar radiation
MAE Mean absolute error
RMSE Root mean square error
R2 Determination coefficient
h Sunshine durations
OPp occurrence of precipitation
MLR Multiple linear regression
RH Relative humidity
PWV Water vapor pressure
Tm minimum air temperature
w weighting coefficient
BP back-propagation
sites in Australia, the results showed that the Solis, Esra and REST2
approaches performed the best, while the Iqbal, Esra and REST2
methods are the most proficient clear skybeam models. Mghouchi
et al. [69] compared and validated three solar radiation models
under all sky conditions at Tetuan city in northern of Morocco, the
results indicated that some models can be preferred to estimate
the solar radiation intensities for the studied site and for other
locations that have similar climate conditions. Wu and Wang [70]
proposed a novel hybrid model for short term solar radiation
prediction, which outperforms than the BP model and the Auto-
regressive Integrated Moving Average model. Despotovic et al. [71]
used statistical analysis to evaluate performance of analyzed dif-
fuse solar radiation models using long term measurements at 267
different sites around the world, which was visually presented by
means of Taylor diagrams. Despotovic et al. [72] further conducted
a detailed statistical analysis and comparison of Hg models, a total
of 101 different radiation models were tested on long term
meteorological data corresponding to 924 sites throughout the
world, and the results showed wide range of calculated statistical
indicators, from very poor to satisfactory.
It can be seen from above analysis that the estimated results of
solar radiation using remote sensing methods may not be as good
as above empirical, physical and artificial intelligence methods,
although it can provide the spatial distributions of solar radiation
in regional or global scales [66,73]. The model performance may
also differ greatly from different observation stations for each
model, and there are relative less studies focusing on comparing
the model accuracy for stations at different climates, so it is not
able to scientifically and comprehensively evaluate the model
performances with different input variables [74,75]. For instance,
the coefficients for the origin BC model are site specific (local
geographical and atmospheric conditions), which should be
modified to expand its application by incorporating more
meteorological data to derive the atmospheric transmissivity
[21,76].
The aim of this study is to review recent studies related to
application of artificial intelligence methods to Hg and to apply
four different methods, including MLP, GRNN, RBNN and IBC, for
modeling daily Hg in different climate zones (Plateau Climate,
Cold-Temperate, Mid-Temperate, Warm-Temperate, Subtropical
and Tropical climate zones, respectively). Daily observations of
general meteorological parameters, Ta, PA, RH, PWV and h at 12
stations in different ecosystems of China during 1961–2014 are
used for training and testing each solar radiation model. The
information about model principles and operation steps will be
described in detail and the model results will be evaluated
through the measures of fit, RMSE, MAE and R2. The model
accuracies will be further compared at different stations (and cli-
mate zones) for comprehensively evaluate different radiation
 L. Wang et al. / Renewable and Sustainable Energy Reviews 61 (2016) 384–397
models, which might be helpful in selection of the most appro-
priate and accurate model for mapping solar energy in regional or
global scale based on the available meteorological data.
2. Materials and methods
2.1. Sites and data processing
In order to test the model performances (MLP, GRNN, RBNN
and IBC model) at different climate zones, daily measurements of
radiation and meteorological variables during 1961–2014 at 12
stations across China were selected, the geographical distributions
of above stations can be seen in Fig. 1 and Table 1. LSA and GEM
stations are located at the Qinghai-Tibet Plateau, the climate in
this area belongs to the Plateau Climate Zone, which is char-
acterized by extremely cold winters with large ranges between the
annual TM and Tm. For example, the annual mean Ta, TM and Tm at
LSA station is about 8.18, 15.95 and 1.78 °C, respectively; the RH
and PWV at both LSA (RH and PWV are 43% and 5.15 hPa, respec-
tively) and GEM stations (RH and PWV are 32% and 3.21 hPa,
respectively) are the lowest among all the stations. The HEB sta-
tion is located at the Cold-Temperate Zone, which is characterized
by long and cold winter, short and cool summer; the annual Ta, TM
and Tm at HEB station is 4.79, 10.66 and
0.89 °C, respectively. The
ELHT, BJ, ALT and LZ stations are in Northern China (Mid-Tempe-
rate zone), which is hot in summer, cold in winter months. The
detailed information about the meteorological conditions can be
clearly seen in Table 1, for example, the annual Ta, TM, Tm, RH and
PWV at BJ station are 12.64 °C, 18.50 °C, 7.22 °C, 54% and 10.34 hPa,
respectively. HT station is located at the Warm-Temperate Zone,
the annual mean daily h is about 7.88 h; the annual Ta, TM and Tm
are 13.10, 19.69 and 7.33 °C, respectively, which is obviously higher
than those at ALT station (Table 1) in Northwestern China. The
WH, HZ and NN stations belong to the Subtropical Climate Zone,
which is characterized by hot summer (with abundant rainfall)
and warm winter; for example, the annual Ta, TM, Tm, RH and PWV at
HZ station are 18.18 °C, 23.36 °C,14.13 °C, 72% and 17.20 hPa,
respectively. The HK station is located in the Hainan Peninsula, and
the climate here belongs to typical tropical zone, the annual Ta, RH
Fig. 1. Spatial distribution of the solar radiation observation stations.
387
and PWV (25.22 °C, 82% and 26.68 hPa, respectively) in this area are
the highest among all the stations.
It is reported that two different types of the radiometers have
been used for observing solar radiation in China before 1994 and
afterwards, respectively, so the daily observations of Hg at above
stations should be checked to ensure the data quality [77,78]. The
radiation data are controlled by following principles: the mea-
sured radiation should not exceed the G0 and not less than the
lower bound; the surface observed Hg should not exceed the clear-
sky daily solar radiation too much; the data with evident sys-
tematic and operational errors are further removed to exclude the
likely noisy data [79]. After the quality control, the remaining data
are used for model development and validation in this study.
The monthly variations of h, Ta, RH, Hg, TM-Tm and PWV at each
station are shown in Fig. 2, it is clear that most of the meteor-
ological variables (h, Ta, Hg and PWV) are higher in summer and
lower in winter months, for example, the monthly mean Hg at ALT
station is 25.25 and 5.69 MJ m
2 day
1 in June and December,
respectively; the monthly mean values of Ta at GEM station is
18.22 °C in July, while this value decreases to
9.19 °C in January.
There are also obvious differences at different stations for each
meteorological variable, for example, the monthly Hg is generally
higher at LSA and GEM stations (larger than 15 MJ m
2 day
1 for
most months) and the lower values of Hg are observed at HEB and
HZ stations (lower than 18 MJ m
2 day
1 for most months); the
highest monthly mean PWV is observed at HK and NN stations,
which are generally higher than 12 hPa throughout the year, while
the largest monthly PWV at GEM, LSA and ELHT stations are lower
than 12 hPa; the monthly mean Ta at HK station is generally higher
than other stations throughout the year and the lowest Ta is
observed at ALT station; the monthly h is highest at ALT and ELHT
stations and the lowest values are observed at NN and HZ stations.
However, the monthly values of (TM
Tm) at some stations are
lower in summer months, for example, the highest monthly
(TM
Tm) at BJ station is 12.97 °C in May, while the lowest values
are observed in August (9.48 °C). Similar phenomenon has also
been observed at other stations except HK station: the monthly
(TM
Tm) is higher in spring and lower in winter months at HK
station, and the values of (TM
Tm) are generally lower than 8 °C
during the year. This relative small temperature differences
between day and night are associated with the larger specific heat
388 L. Wang et al. / Renewable and Sustainable Energy Reviews 61 (2016) 384–397

Table 1
The geographical locations of each radiation station and its annual mean meteorological parameters.

Station Lon Lat Altitude (m) Hs (h) Tmean (°C) Tmax (°C) Tmin ( °C) RH (%) PwvG (hPa)

ALT 88°050 E 47°440 N 735.30 8.89 5.35 11.78
0.69 56 6.05 16.22
BJ 116°280 E 39°480 N 31.3 8.09 12.64 18.50 7.22 54 10.34 15.58
ELHT 111°580 E 43°390 N 964.7 9.01 4.39 11.99
2.31 47 4.86 17.52
GEM 94°540 E 36°250 N 945.4 8.58 5.26 13.01
1.39 32 3.21 19.31
HEB 126°460 E 45°450 N 142.30 7.46 4.79 10.66
0.89 64 8.10 13.85
HK 110°210 E 20°020 N 13.9 6.90 25.22 29.56 22.32 82 26.68 16.39
HT 79°560 E 37°080 N 1374.5 7.88 13.10 19.69 7.33 41 6.69 16.85
HZ 120°100 E 30°140 N 41.7 6.79 18.18 23.36 14.13 72 17.20 15.29
LSA 91°080 E 29°400 N 3648.7 8.26 8.18 15.95 1.78 43 5.15 20.47
LZ 103°530 E 36°030 N 1035.3 7.7 9.30 16.61 3.51 56 7.52 15.90
NN 108°130 E 22°380 N 121.6 6 23.62 28.97 19.94 78 23.41 15.65
WH 114°080 E 30°370 N 23.1 7.35 18.39 23.72 14.09 74 17.67 15.58

The unit of G is MJ m
2 day
1.

Fig. 2. Monthly variations of metrological parameters for each station.

capacity of water at HK station, in which is surrounded by the sea.

Meanwhile, the monthly RH at subtropical and tropical stations are
relatively stable during the year, however, the RH values are gen-
erally lower at stations in Mid-Temperate or Qinghai-Tibet Plateau
(HT, GEM and ELHT), which may be associated with the Mongolia
and Siberian high pressure system.

2.2. Solar radiation prediction models

2.2.1. Multilayer perceptron neural network
ANNs are information processing systems inspired from the
biological neural network, which can be considered as a mathe-
matical model developed by the concepts of neural biology. MLP
neural networks are known as a class of nonlinear models with the
ability to discover patterns, simulation and time series forecasting
adaptively from the data [80,81]. The effort in an MLP is based on
the intrinsic relationship established between data, nonlinear
mapping between independent and dependent variables. Each
Fig. 3. Schematic architecture of a MLP neural network.
MLP model is consisted of input layer, hidden layer and output
layer. In each hidden and output layer, there are several processing
elements (also called as neurons or nodes). Each processor is
associated and interconnected with all the processors in the next
layer. The relationship between network layers is possible to
investigate according to the weighting coefficients (w) and biases
(B) for each processor and activation function (Fig. 3). After
establishing the MLP architecture, the information given to MLP is
 L. Wang et al. / Renewable and Sustainable Energy Reviews 61 (2016) 384–397
propagated from input layer to the output layer through hidden
layer(s). The output can be regarded as assimilation of the network
after the training process. In the training process it is necessary to
adopt appropriate learning algorithm to reduce errors. It should be
noted that the learning algorithms for a MLP model are based on
BP technique which is normally a steepest gradient descent
method. The main objective of the BP method is to reduce the
amount of network error that can be calculated using [82]:
X
N division and summation units. The summation and output layers
e ¼ 0:5  ðok
t k Þ2
k¼1
where N is the number of processors, ok is the network output in
the kth processor and tk denotes the target value.
2.2.2. Radial basis neural network
A RBNN model can be regarded as a type of MLP with a single
hidden layer. The main differences between the MLP and RBF
neural networks can be given as follows: (1) the connections
between the input layer and the hidden layer for RBNN models are
not weighted, and (2) the transfer (activation) functions on the
nodes of the hidden layer are designed to be radically symmetrical.
The most common transfer function which is designated to RBNN
networks are Gaussian function, whilst and other options such as
multiquadric, inverse quadric and polyharmonic spline functions
might also be applied [83].
Fig. 4a indicates schematic diagram of a RBNN model with
several inputs and one output neuron. RBNN parameters include
the centers (Uj) and spread (σj) of the transfer functions in the
hidden layer nodes and consist of synaptic weights w in the output
layer node. Having the RBNN centers in each separate point in the
input space is desirable, but practically a few entry points of the
total available points might be selected through a process called
clustering. For Xi as an input vector, jth hidden node shows the
response of hj based on [84]:
" #

‖Xi
Uj ‖
hj ¼ exp ð2Þ
2σ 2j
here, ‖Xi
Uj ‖ represents the distance between Xi and the center
of the jth hidden node to which is measured by a norm operator
(e.g. Euclidean norm). Eventually, the output of RBNN model is
provided in the kth output node via
X
L
yik ¼ hj wkj
j¼1
2.2.3. Generalized regression neural network
A GRNN model is a kind of RBNN which is based on kernel
regression networks. Originally, GRNN model is proposed as an
alternative to the BP training algorithm for MLP model, but unlike
MLP, they do not require an iterative training procedures such as
back-propagation algorithm. GRNNs have the ability to approx-
imate any nonlinear function between input and output vectors in
a direct function drawing from the training data [85]. Moreover, in
GRNNs, the larger the size of the training set, the smaller the
Fig. 4. General structure of (a) RBNN and (b) GRNN models.
389
estimation error would be. As shown in Fig. 4b, GRNN model
consists of four layers: the input layer, pattern layer, summation
layer and output layer. The number of entries in the input layer
will depend on the total number of observed parameters. The first
layer, in which each neuron presents a training pattern and pro-
vides its output, is connected to the pattern layer. Pattern layer is
attached to the summation layer. The third layer (summation
layer) consists two different types of summation including single
ð1Þ work together to normalize the output vector. In training process
of network, linear and radial basis transfer functions are used in
the output and hidden layers. Each neuron in the pattern layer is
connected to both neurons in the summation layer (summation
neurons of S and D). S-summation neuron calculates the sum of
the weighted responses in pattern layer, whereas, D-summation
neuron computes the unweighted output neuron in the pattern
layer. The output layer just divides each S-summation neuron on
0
D-summation neuron and gives predicted Y i for an unspecified
input vector and the value of x as the following expression [85,86]:
P
n
yi :exp½
Dðx; xi Þ m 
0 X xi
xik 2
Y i ¼ i ¼ n1 ; Dðx; xi Þ ¼ ð4Þ
P σ
exp½
Dðx; xi Þ k¼1
i¼1
yi denotes the weight relationship between neuron i-th in the
pattern layer and S neuron in the summation layer. D is the
Gaussian function, n and m stand for the number of training pat-
terns and the number of elements of an input vector, respectively.
xik is kth element of and xi, also, σ is the spread (smoothness)
parameter whose optimal value can be experimentally evaluated.
2.2.4. Improved Bristow–Campbell model
Bristow and Campbell [76] proposed a method for estimating
the daily Hg from diurnal temperature range (TM
Tm), the atmo-
spheric transmittance, the ratio between Hg and G0, is calculated
as a function of (TM
Tm), which can be expressed as
Hg
¼ A½1
expð
BðT M
T m ÞC Þ ð5Þ
G0
where A is the maximum radiation expected on a clear day, being
distinctive for each location and depending on air quality and
ð3Þ
altitude; coefficients B and C control the rates at which A is
approached as the temperature difference increases. The above BC
model does not take into considerations of other factors influen-
cing the incoming surface solar radiation, such as relative
humidity, cloud cover, etc. So more meteorological variables [for
example, TM, Tm, RH and OPp (binary value)] should be employed to
represent the actual atmospheric transmittance in this study, the
IBC model is proposed as follow
 
dm
G0 ¼ 37:54   ½ðH S Þ sin ϕ sin δ þ cos ϕ cos δ sin H S  ð6Þ
d
390 L. Wang et al. / Renewable and Sustainable Energy Reviews 61 (2016) 384–397

!2
n 
P  
Xmi
Xm Xoi
Xo
i¼1
R2 ¼ n 
ð11Þ
P 2 P
n  2
Xmi
Xm Xoi
Xo
i¼1 i¼1

where N and bar respectively indicate the number of data and

mean of the variable, Xm and Xo are the modeled and observed
daily G.

3.2. Model performances

In present work, the accuracies of above four different meth-

ods, MLP, GRNN, RBNN and IBC models, were compared for
modeling daily Hg in different climate zones (Plateau Climate,
Cold-Temperate, Mid-Temperate, Warm-Temperate, Subtropical
and Tropical climate zones, respectively). Daily meteorological
parameters, Ta, PA, RH, PWV and h at 12 stations in China during
1961-2014 were used as model inputs to the applied models for
estimating solar radiation. 70% of the entire dataset was used for
training and remaining 30% was used for testing phases.
Before applying MLP, GRNN and RBNN methods to the data, the
Fig. 5. Variation of spread constant vs RMSE for the GRNN model in ALT station.
input and output values were standardized using the accom-
panying mathematical statement
W ¼ 2πj=365 and j is the Julian day. The solar radiation at the
top of the atmosphere can be calculated as xi
xmin
c1 þ c2 ð12Þ
xmax
xmin
Hg
¼ ðb0 þ b1 sin ðWÞ þ b2 cos ðWÞ þ b3 RH þ b4 OPpÞ
G0 where xmin and xmax indicate the extreme values of the dataset; xi
b6 is the observed value of the variable at time i. Scaling elements c1
½1
expð
b5 ðT M
T m Þ Þ ð7Þ
and c2 can take distinctive values. In this study, 0.8 and 0.2 were
dm allotted for the c1 and c2, respectively; in this way, training data
d
is known as the correction factor for the Sun-Earth distance,
which can be obtained from were scaled in the reach [0.2, 0.8].

dm pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
¼ 1:00011 þ 0:034221 cos ξ þ 0:034221 sin ξ þ0:000719 cos 2ξ þ 0:000077 sin 2ξ ð8Þ
d

The daily angle (ξ, in rad) is calculated as a function of the For the GRNN models, distinctive spread constants were attemp-
Julian day (ξ ¼ 2πðj
1Þ=365); Hs and δ are the solar angle at sunset ted and the ideal values that gave the lowest RMSE values were
and solar declination, respectively. obtained for each model. As a sample, the variety of spread constant
Then, the coefficients of the IBC model (b0, b1, b2, b3, b4, b5 and versus RMSE for the GRNN model at ALT station is illustrated in Fig. 5.
b6) can be obtained by minimizing the sums of the squares of Diverse number of hidden nodes was tried for the MLP models and
deviations between observed and expected values. These deriva- the ideal ones were achieved for each station. As an illustration, the
tives are numerically computed using finite differences, the variety of hidden node number versus RMSE for the MLP model at
Newton-Raphson algorithm is used for multivariate nonlinear ALT station is shown at Fig. 6. For the RBNN models, the ideal spread
optimization, which is conducted in the SPSS 22.0 software. The constants and hidden node numbers were obtained by trial and error
detailed procedures for calculating the model parameters for IBC method for each station. As an illustration, the variety of hidden node
model can be seen in Meza et al. [21] and Pan et al. [23]. number and spread constant versus RMSE for the RBNN model at ALT
station is given in Fig. 7. Comparisons of ideal ANN and IBC models
are made in Table 2 for above 12 stations. For the GRNN, IBC, MLP and
RBNN models, the RMSE values range 2–3.29, 3.13–4.58, 1.94-3.27,
3. Model applications and results 1.96-3.25 MJ m
2 day
1, respectively. For the GRNN, MLP and RBNN
model, the most extreme RMSE values (3.29, 3.27 and
3.1. Comparisons of measures of fit 3.25 MJ m
2 day
1) were found for the station HZ. For the IBC model,
however, the most extreme RMSE value (4.58 MJ m
2 day
1) was
The measures of fit used in the present study includes RMSE, obtained for the ALT station. The best accuracy for GRNN
MAE and R2, which can be expressed as (RMSE¼ 2 MJ m
2 day
1), MLP (RMSE¼1.94MJ m
2 day
1) and
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi RBNN (RMSE¼1.96 MJ m
2 day
1) model was found at the NN sta-
u N
u1 X tion while the IBC model (RMSE¼3.13 MJ m
2 day
1) gave the best
RMSE ¼ t ðXmi
Xoi Þ2 ð9Þ
Ni¼1 accuracy at HT station. It is clear from Table 2 that each ANN model
generally performs superior to the IBC models in forecasting global
solar radiation in different climate zones. The MLP model produces
1Xn
MAE ¼ jXmi
Xoi j ð10Þ the highest accuracy in seven of twelve stations (ALT, BJ, ELHT, HT, HZ,
Ni¼1
LSA and LZ station). The RBNN model also brings higher model
 L. Wang et al. / Renewable and Sustainable Energy Reviews 61 (2016) 384–397 391

Table 2
The test statistics of different models in predicting the solar radiation

Training Testing

MAE RMSE R2 Mean MAE RMSE R2 Mean

ALT GRNN 1.87 2.43 91 16.26 2.15 2.72 89 16.15

ALT IBC 2.18 3.16 85 16.6 3.41 4.58 70 15.35
ALT MLP 1.97 2.59 90 16.26 2.08 2.64 90 16.15
ALT RBNN 2.06 2.68 89 16.26 2.16 2.74 89 16.15
BJ GRNN 1.82 2.35 88 15.81 1.83 2.27 88 15.05
BJ IBC 2.37 3.08 79 15.81 2.93 3.86 66 15.04
BJ MLP 1.87 2.43 87 15.81 1.80 2.02 89 15.05
BJ RBNN 1.96 2.51 86 15.81 1.87 2.28 88 15.05
ELHT GRNN 1.85 2.49 89 17.52 2.09 2.76 87 17.53
ELHT IBC 2.3 3.24 82 17.6 3.02 4.14 70 17.35
ELHT MLP 1.97 2.67 87 17.52 1.97 2.62 88 17.53
ELHT RBNN 2.04 2.73 87 17.52 2.00 2.68 87 17.53
GEM GRNN 1.75 2.28 90 19.56 1.97 2.43 90 18.72
GEM IBC 2.35 3.14 82 19.62 2.51 3.45 77 18.56
GEM MLP 1.83 2.39 88 19.56 1.75 2.19 92 18.72
GEM RBNN 1.88 2.44 88 19.56 1.77 2.19 92 18.72
HEB GRNN 2.19 2.91 83 13.81 2.26 2.92 83 13.93
HEB IBC 2.34 3.27 80 14.1 3.19 4.16 63 13.26
HEB MLP 2.26 2.99 82 13.81 2.22 2.86 83 13.93
HEB RBNN 2.23 2.98 83 13.81 2.21 2.85 84 13.93
HK GRNN 2.3 2.96 75 16.04 2.03 2.59 82 17.21
Fig. 6. Variation of hidden node number vs RMSE for the MLP model in ALT station. HK IBC 2.88 3.63 63 17.26 3.35 4.15 53 14.36
HK MLP 2.26 2.92 75 16.04 2.02 2.59 82 17.21
HK RBNN 2.28 2.94 75 16.04 2.03 2.59 82 17.21
HT GRNN 1.56 2.02 90 16.97 2.06 2.69 85 16.58
HT IBC 1.99 2.67 82 16.99 2.32 3.13 77 16.54
HT MLP 1.61 2.08 89 16.97 1.92 2.53 87 16.58
HT RBNN 1.68 2.17 88 16.97 1.97 2.58 87 16.58
HZ GRNN 2.2 2.82 80 15.02 2.34 3.29 72 15.92
HZ IBC 2.64 3.41 69 16.17 3.11 4.02 64 13.25
HZ MLP 2.1 2.71 82 15.02 2.29 3.27 73 15.92
HZ RBNN 2.28 2.91 79 15.02 2.32 3.25 72 15.92
LSA GRNN 2.6 3.34 67 20.55 2.13 2.63 77 20.29
LSA IBC 2.67 3.45 66 20.37 2.86 3.76 52 20.71
LSA MLP 2.66 3.44 65 20.55 1.83 2.30 81 20.29
LSA RBNN 2.74 3.52 63 20.55 1.89 2.36 80 20.29
LZ GRNN 1.82 2.53 89 15.9 2.02 2.56 86 15.89
LZ IBC 2.34 3.20 82 15.77 3.19 4.26 64 16.21
LZ MLP 1.95 2.68 87 15.9 1.89 2.41 88 15.89
LZ RBNN 1.99 2.72 87 15.9 1.93 2.45 87 15.89
NN GRNN 1.97 2.57 79 15.52 1.58 2.00 86 15.97
NN IBC 2.4 3.05 66 16.59 2.85 3.62 55 13.47
NN MLP 2 2.61 78 15.52 1.53 1.94 86 15.97
NN RBNN 2.03 2.63 78 15.52 1.54 1.96 86 15.97
WH GRNN 2.49 3.21 76 15.61 2.10 2.76 78 15.51
WH IBC 2.92 3.76 65 16.27 3.03 3.96 61 13.98
Fig. 7. Variation of hidden node number and spread constant vs RMSE for the WH MLP 2.44 3.16 77 15.61 1.98 2.61 81 15.51
RBNN model in ALT station. WH RBNN 2.45 3.18 77 15.61 1.94 2.59 81 15.51

performances at four stations (GEM, HEB, HK and NN) while the
GRNN model only produces high accuracy at only one station
in China.
The MAE values given in Table 2 range 1.58–2.34, 2.32–3.41, 1.53–
2.29, 1.54–2.32 MJ m
2 day
1 for the GRNN, IBC, MLP and RBNN
models, respectively. Like the RMSE criterion, the lower MAE values
(1.58, 1.53 and 1.54 MJ m
2 day
1, respectively) of the GRNN, MLP
and RBNN models were found at NN station while the minimum
MAE value (2.32 MJ m
2 day
1) for the IBC model was obtained for
HT station. Furthermore, the larger MAE values for GRNN (MAE ¼
2.34 MJ m
2 day
1), MLP (MAE¼2.29 MJ m
2 day
1) and RBNN
(MAE ¼ 2.32 MJ m
2 day
1) model was obtained at HZ station
while the IBC model (MAE¼3.41 MJ m
2 day
1) is demonstrated as
the worst method at ALT station. Table 2 clearly shows that the ANN
models generally outperform the IBC model with respect to MAE
statistics. It is observed that the MLP model produces the best
accuracy in seven of twelve stations (ALT, BJ, ELHT, HT, HZ, LSA and
LZ) with respect to MAE and R2 values. The RBNN and MLP models
gave similar statistical indices at four stations (GEM, HEB, HK and NN
stations) while the RBNN model was observed to be superior to the
The unit of MAE, RMSE and Mean are MJ m
2 day
1.
alternate models at only one station. For the GRNN, IBC, MLP and
RBNN models, the most compelling determination coefficients R2
are 90, 77, 92 and 92% at GEM station, respectively. The minimum
determination coefficient values were found to be 72, 73 and 72% for
the GRNN, MLP and RBNN models at HZ station while the IBC model
has the lowest R2 (52%) at LSA station. It is clearly seen from Table 2
that all the ANN models produce more accurate results than the IBC
models, for example, the GRNN, MLP and RBNN model estimate the
annual mean global solar radiation (16.22 MJ m
2 day
1) as
16.15 MJ m
2 day
1 with underestimations of 0.43% while the IBC
model results in 15.35 MJ m
2 day
1 with an underestimation of
5.4% at ALT station; similar model performances were also obtained
for other stations.
The observed and predicted global solar radiation values by
GRNN, MLP and RBNN models in training and testing stages are
plotted in Figs. 8–10. It is clear from the figures that the higher
solar radiation values ( 430 MJ m
2 day
1) were significantly
underestimated for the stations ALT, ELHT, HK and HZ. The main
392 L. Wang et al. / Renewable and Sustainable Energy Reviews 61 (2016) 384–397

Fig. 8. Comparison between daily measured solar radiation and estimates using GRNN model during the training (a) and validation (b) period for each station.

reason may be the high differences in training and testing data
ranges and the distributions for these stations. All the models gave
the least scattered estimates at NN station, which is also confirmed
by the various statistics provided in Table 2. Comparisons between
above three ANN methods indicate that the MLP and RBNN models
generally give better accuracy than the GRNN model. This can also
be clearly seen from the fit line equations and R2 statistics. Fig. 11
illustrates the estimates of IBC model both in training and testing
stages, model comparisons between IBC and ANN methods indi-
cate that the IBC model generally provide higher scattered esti-
mates and the fit lines of the ANN models are closer to the ideal
line (45° line) than those of the IBC model.
 L. Wang et al. / Renewable and Sustainable Energy Reviews 61 (2016) 384–397 393

Fig. 9. Comparison between daily measured solar radiation and estimates using MLP model during the training (a) and validation (b) period for each station.

4. Discussion promising research area due to the simplicity of utilization and

The model results at above 12 stations in diverse ecosystems
indicate that all the ANN models performed superior to the IBC
model. Scientifically, an ANN model can be acknowledged as a
universal approximator [87], which has already become a
straightforward definition [88]. Among the ANN methods, the MLP
and RBNN models were found to have preferred accuracies over
the GRNN model in predicting global solar radiation. This outcome
is consistent with the relevant literatures, for example, Kisi [89]
investigated the accuracies of MLP, RBNN and GRNN models in
394 L. Wang et al. / Renewable and Sustainable Energy Reviews 61 (2016) 384–397

Fig. 10. Comparison between daily measured solar radiation and estimates using RBNN model during the training (a) and validation (b) period for each station.

estimating daily suspended sediment concentration and the
results showed that the MLP and RBNN methods gave preferable
results over the GRNN model. The capability of above three dis-
tinctive ANN techniques are also investigated in modeling refer-
ence evapotranspiration by Kisi [90], the GRNN model gave the
second rate results when compared with MLP and RBNN models.
Seckin et al. [91] predicted the flood peaks of different return
periods at ungauged sites and the model performances further
indicated that the MLP and RBNN performed superior to the GRNN
method. Pinar et al. [92] analyzed the performances of MLP, RBNN
and GRNN models in evaluating backwater through arched bridge
constrictions with normal and skewed crossings, and the MLP and
 L. Wang et al. / Renewable and Sustainable Energy Reviews 61 (2016) 384–397 395

Fig. 11. Comparison between daily measured solar radiation and estimates using IBC model during the training (a) and validation (b) period for each station.

RBNN models produced more accurate forecasts than the other
methods.
The primary focal points of MLP contrasted to RBNN model are
(1) the MLP model makes global approximations, while the RBNN
model estimates locally nonlinear input-output relationships;
(2) MLP model may require less number of parameters than the
RBNN for obtaining the same precision [93]. Despite its inferior
results, the GRNN model might likewise be favored rather than the
MLP method in view of the accompanying points of interest:
(1) The MLP precision is highly sensitive to randomly assigned
initial weight values while this issue is not confronted in GRNN
model [94]; (2) The GRNN model does not require an iterative
396 L. Wang et al. / Renewable and Sustainable Energy Reviews 61 (2016) 384–397
training procedure as that in the MLP model [95]; (3) The local
minima issue is not confronted in GRNN model.
5. Conclusion
The applicability of three ANN models, MLP, GRNN and RBNN
models in modeling daily surface solar radiation at different cli-
matic regions was investigated. The major contribution is the use
of routine meteorological parameters as inputs to overcome the
lack of rarely measured radiation database. Long-term con-
tinuously observations of meteorological data, i.e.: h, Ta, RH, Hg,
TM
Tm and PWV at 12 stations were used in the present study for
analyzing their variation characteristics and solar radiation model
development, validation and comparisons.
An improved Bristow–Campbell (IBC) model has been applied by
considering the influencing factors of surface solar radiation, for
example, TM, Tm, RH and OPp have been used to represent the actual
atmospheric transmittance. The model performances have been
evaluated and compared through the statistical indices RMSE, MAE
and R2 for each model. The results indicate that the MLP and RBNN
models generally provide better accuracies than the GRNN and IBC
models in predicting daily solar radiation, for example, the RMSE
values range 2–3.29, 3.13–4.58, 1.94–3.27, 1.96–3.25 MJ m
2 day
1
for GRNN, IBC, MLP and RBNN models, respectively, and the most
extreme RMSE values (3.29, 3.27 and 3.25 MJ m
2 day
1 for the
GRNN, MLP and RBNN models) were found for the HZ station; the
MAE values given in Table 2 range 1.58–2.34, 2.32–3.41, 1.53–2.29,
1.54–2.32 MJ m
2 day
1 for the GRNN, IBC, MLP and RBNN models,
respectively. The statistical results also show differences at different
stations for each model, for example, the higher accuracy for GRNN
(RMSE¼2 MJ m
2 day
1), MLP (RMSE¼ 1.94 MJ m
2 day
1) and
RBNN (RMSE¼1.96 MJ m
2 day
1) models were found at the NN
station while the IBC model (RMSE¼3.13 MJ m
2 day
1) gave the
best accuracy at HT station. Meanwhile, it is observed that the ANN
models underestimated high radiation values (430 MJ m
2
day
1) at some stations and the IBC model generally provided higher
scattered estimates, which may due to the differences in training and
test data ranges and distribution for the stations.
From above, it can be concluded that the MLP and RBNN
models are more accurate in estimating solar radiation at different
climatic zones in China, which is of vital importance for surface
energy budget, climate change and energy applications. Certainly,
the models should be further applied and tested at other places in
the world for extending the international applicability. More
attentions will be paid on the atmospheric radiative transferring
mechanism and mapping the regional and global radiation dis-
tributions using remote sensing techniques.
Acknowledgments
This work was financially supported by the Special Fund for
Basic Scientific Research of Central Colleges, China University of
Geosciences, Wuhan (No. CUG150631), and the Fundamental
Research Funds for the Central Universities (No. 2042016kf0165).
We would like to thank the China Meterological Administration
(CMA) for providing the meteorological and radiation data.
References
[1] Beer C, Reichstein M, Tomelleri E, Ciais P. Terrestrial gross carbon dioxide
uptake: global distribution and covariation with climate. Science
2010;329:834–8.
[2] Khatib T, Mohamed A, Sopian K. A review of solar energy modeling techni-
ques. Renew Sustain Energy Rev 2012;16(5):2864–9.
[3] Islam MD, Kubo I, Ohadi M, Alili MM. Measurement of solar energy radiation
in Abu Dhabi, UAE. Appl Energy 2009;86:511–5.
[4] Meza F, Varas E. Estimation of mean monthly solar global radiation as a
function of temperature. Agric Forest Meteorol 2000;100:231–41.
[5] Kisi O. Modeling solar radiation of Mediterranean region in Turkey by using
fuzzy genetic approach. Energy 2014;64:429–36.
[6] Ayodele TR, Ogunjuyigbe ASO. Prediction of monthly average global solar
radiation based on statistical distribution of clearness index. Energy
2015;90:733–1742.
[7] Park JK, Das A, Park JH. 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.
[8] Purohit I, Purohit P. Inter-comparability of solar radiation databases in Indian
context. Renew Sustain Energy Rev 2015;50:735–47.
[9] Wild M. Global dimming and brightening: a review. J Geophys Res 2009. http:
//dx.doi.org/10.1029/2008JD011470.
[10] Asaf D, Rotenberg E, Tatarinov F, Dicken U, Montzka SA, Yakir D. Ecosystem
photosynthesis inferred from measurements of carbonyl sulphide flux. Nat
Geosci 2014;6:186–90.
[11] Remund J, Kunz S, Schilter C. METEONORM Version 6.0. Meteotest. Bern,
Switzerland: Fabrikstrasse; 2007 www.meteonorm.com.
[12] Badescu V. Modeling Solar Radiation at the Earth's Surface. Springer; 2014.
[13] Besharat F, Dehghan AA, Faghih AR. Empirical models for estimating global
solar radiation: A review and case study. Renew Sustain Energy Rev
2013;21:798–821.
[14] Luis FZ, Jesús P, Luis M, Lourdes R, Bella E. A new statistical approach for deriving
global solar radiation from satellite images. Sol Energy 2009;83:480–4.
[15] Teke A, Başak Yıldırım H, Çelik Ö. Evaluation and performance comparison of
different models for the estimation of solar radiation. Renew Sustain Energy
Rev 2015;50:1097–107.
[16] Janjai S, Pankaewa P, Laksanaboonsong J. A model for calculating hourly global
solar radiation from satellite data in the tropics. Appl Energy 2009;86:1450–7.
[17] 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.
[18] Benghanem M, Mellit A. A simplified calibrated model for estimating daily
global solar radiation in Madinah, Saudi Arabia. Theor Appl Climatol
2014;115:197–205.
[19] Yacef R, Benghanem M, Mellit A. Prediction of daily global solar irradiation
data using Bayesian neural network: a comparative study. Renew Energy
2012;48:146–54.
[20] Li R, Zhao L, Wu TH, Ding YJ, Xin YF, Zou DF, et al. Temporal and spatial
variations of global solar radiation over the Qinghai-Tibetan Plateau during
the past 40 years. Theor Appl Climatol 2013;113:573–83.
[21] Meza FJ, Yebra ML. Estimation of daily global solar radiation as a function of
routine meteorological data in Mediterranean areas. Theor Appl Climatol 2015.
http://dx.doi.org/10.1007/s00704-015-1519-6.
[22] 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.
[23] Pan T, Wu S, Dai E, et al. Estimating the daily global solar radiation spatial
distribution from diurnal temperature ranges over the Tibetan Plateau in
China. Appl Energy 2013;107:384–93.
[24] Yaniktepe B, Genc YA. Establishing new model for predicting the global solar
radiation on horizontal surface. Int J Hydrog Energy 2015;40:15278–83.
[25] Pyrina M, Hatzianastassiou N, Matsoukas C, Fotiadi A, Papadimas C, Pavlakis K,
et al. Cloud effects on the solar and thermal radiation budgets of the Medi-
terranean basin. Atmos Res 2015;152:14–28.
[26] Gueymard C. An atmospheric transmittance model for the calculation of the
clear sky beam, diffuse and global photosynthetically active radiation. Agric
Forest Meteorol 1989;45:215–29.
[27] Yang K, Koike T, Ye B. Improving estimation of hourly, daily, and monthly solar
radiation by importing global data sets. Agric Forest Meteorol 2006;137:43–55.
[28] Gueymard CA, Myers DR. Evaluation of conventional and high-performance
routine solar radiation measurements for improved solar resource, climato-
logical trends, and radiative modeling. Sol Energy 2009;83:171–85.
[29] Schmetz J. Towards a surface radiation climatology: retrieval of downward
irradiances from satellites. Atmos Res 1989;23:287–321.
[30] Noia M, Ratto C, Festa R. Solar irradiance estimation from geostationary
satellite data: I. Statistical models, II. Physical models. Sol Energy
1993;51:449–65.
[31] Pinker R, Frouin R, Li Z. A review of satellite methods to derive surface
shortwave irradiance. Remote Sens Environ 1995;51:108–24.
[32] Perez R, Aguiar R, Collares-Pereira M, Dumortier D, Estrada Cajigal V, Guey-
mard C, Ineichen P, Littlefair P, Lund H, Michalsky J, Olseth J, Renné D, Rymes
M, Skartveit A, Vignola F, Zelenka A. Solar resource assessment – a review. In:
Zelenka A. (Ed.). Sol Energy – The Statae of the Art. James and James, London.
[33] Rigollier C, Bauer O, Wald L. On the clear sky model of the ESRA-European
Solar Radiation Atlas — with respect to the heliosat method. Sol Energy
2000;68:33–48.
[34] Kambezidis H, Psiloglou B, Karagiannis D, Dumka U, Kaskaoutis D. Recent
improvements of the Meteorological Radiation Model for solar irradiance
estimates under all-sky conditions. Renew Energy 2016;93:142–58.
 L. Wang et al. / Renewable and Sustainable Energy Reviews 61 (2016) 384–397
[35] Shamim M, Remesan R, Bray M, Han D. An improved technique for global solar
radiation estimation using numerical weather prediction. J Atmos Sol – Terre
Phys 2015;129:13–22.
[36] Kashyap Y, Bansal A, Sao AK. Solar radiation forecasting with multiple para-
meters neural networks. Renew Sustain Energy Rev 2015;49:825–35.
[37] Jiang Y. Computation of monthly mean daily global solar radiation in China
using artificial neural networks and comparison with other empirical models.
Energy 2009;34:1276–83.
[38] 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.
[39] 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.
[40] 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.
[41] Aguiar J, Collares-Perrira M, Conde S. Simple procedure for generating of daily
radiation values using a library of Markov Transition Matrices. Sol Energy
1988;49:29–279.
[42] Aguiar R, Collares-Pereira M. TAG: a time-dependent, autoregressive, Gaussian
model for generating synthetic hourly radiation. Sol Energy 1992;49:167–74.
[43] Ghayekhloo M, Ghofrani M, Menhaj M, Azimi R. A novel clustering approach
for short-term solar radiation forecasting. Sol Energy 2015;122:1371–83.
[44] Amrouche B, Pivert X. Artificial neural network based daily local forecasting
for global solar radiation. Appl Energy 2014;130:333–41.
[45] Olatomiwa L, Mekhilef S, Shamshirband S, Mohammadi K, Petkovic D, Sudheer
Ch. A support vector machine-firefly algorithm-based model for global solar
radiation prediction. Sol Energy 2015;115:632–44.
[46] Linares-Rodríguez A, Ruiz-Arias J, Pozo-Vázquez D, Tovar-Pescador J. Genera-
tion of synthetic daily global solar radiation data based on ERA-Interim rea-
nalysis and artificial neural networks. Energy 2011;36:5356–65.
[47] Emad A, El-Nouby Adam M. Estimate of global solar radiation by using arti-
ficial neural network in Qena,Upper Egypt. J Clean Energy Technol
2013;1:148–50.
[48] Shamshirband S, Mohammadi K, Tong C, Zamani M, Motamedi S, Ch S. A
hybrid SVM-FFA method for prediction of monthly mean global solar radia-
tion. Theor Appl Climatol 2015. http://dx.doi.org/10.1007/s00704-015-1482-2.
[49] Rizwan M, Jamil M, Kirmani S, Kothari DP. Fuzzy logic based modeling and
estimation of global solar energy using meteorological parameters. Energy
2014;70:685–91.
[50] Bhardwaj S, Sharma V, Srivastava S, Sastry OS, Bandyopadhyay B, Chandel SS.
Estimation of solar radiation using a combination of hidden markov model
and generalized fuzzy model. Sol Energy 2013;93:43–54.
[51] Aguiar R, Collares-Pereira M, Conde J. Simple procedure for generating
sequences of daily radiation values using a library of Markov transition
matrices. Sol Energy 1988;40:269–79.
[52] Mohanty S, Patra P, Sahoo S. Prediction and application of solar radiation with
soft computing over traditional and conventional approach – a comprehensive
review. Renew Sustain Energy Rev 2016;56:778–96.
[53] Mohammadi K, Shamshirband S, Tong C, Alam K, Petkovic 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.
[54] Yadav A, Chandel S. Solar radiation prediction using Artificial Neural Network
techniques: a review. Renew Sustain Energy Rev 2014;33:772–81.
[55] Hay J, Hanson K. A satellite-based methodology for determining solar irra-
diance at the ocean surface during GATE. Bull Am Meteol Soc 1978;59:1549.
[56] Hay J. Satellite based estimates of solar irradiance at the Earth's surface-I.
Modelling approaches. Renew Energy 1993;3:381–93.
[57] Cano D, Monget JM, Albuisson M, Guillard H, Regas N, Wald L. A method for
the determination of the global solar radiation from meteorological satellite
data. Sol Energy 1986;37:31–9.
[58] Cros S, Albuisson M, Wald L. Simulating Meteosat-7 broadband radiances
using two visible channels of Meteosat-8. Sol Energy 2006;80:361–7.
[59] Polo J, Zarzalejo LF, Cony M, et al. Solar radiation estimations over India using
Meteosat satellite images. Sol Energy 2011;85(9):2395–406.
[60] Antonanzas-Torres F, Cañizares F, Perpiñán O. Comparative assessment of
global irradiation from a satellite estimate model (CM SAF) and on-ground
measurements (SIAR): a Spanish case study. Renew Sustain Energy Rev
2013;21:248–61.
[61] Zhang Y, Li X, Bai Y. An integrated approach to estimate shortwave solar
radiation on clear-sky days in rugged terrain using MODIS atmospheric pro-
ducts. Sol Energy 2015;113:347–57.
[62] Quesada-Ruiz S, Linares-Rodríguez A, Ruiz-Arias J, Pozo-Vázquez D, Tovar-
Pescador J. An advanced ANN-based method to estimate hourly solar radiation
from multi-spectral MSG imagery. Sol Energy 2015;115:494–504.
[63] Kumar R, Aggarwal RK, Sharm JD. Comparison of regression and artificial
neural network models for estimation of global solar radiations. Renew Sus-
tain Energy Rev 2015;52:1294–9.
[64] Rahimikhoob A, Behbahani SMR, Banihabib ME. Comparative study of statis-
tical and artificial neural network's methodologies for deriving global solar
radiation from NOAA satellite images. Int J Climatol 2013;33(2):480–6.
397
[65] Polo J, Antonanzas-Torres F, Vindel JM, et al. Sensitivity of satellite-based
methods for deriving solar radiation to different choice of aerosol input and
models. Renew Energy 2014;68:785–92.
[66] Ahmad M, Tiwari G. Solar radiation models – a review. Int J Energy Res
2011;35:271–90.
[67] Citakoglu H. Comparison of artificial intelligence techniques via empirical
equations for prediction of solar radiation. Comput Electron Agr 2015;118:28–37.
[68] Engerer N, Mills F. Validating nine clear sky radiation models in Australia. Sol
Energy 2015;120:9–24.
[69] Mghouchi Y, Bouardi A, Sadouk A, Fellak I, Ajzoul T. Comparison of three solar
radiation models and their validation under all sky conditions – case study:
Tetuan city in northern of Morocco. Renew Sustain Energy Rev 2016;58:1432–44.
[70] Wu Y, Wang J. A novel hybrid model based on artificial neural networks for
solar radiation prediction. Renew Energy 2016;89:268–84.
[71] Despotovic M, Nedic V, Despotovic D, Cvetanovic S. Evaluation of empirical
models for predicting monthly mean horizontal diffuse solar radiation. Renew
Sustain Energy Rev 2016;56:246–60.
[72] 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.
[73] Linares-Rodriguez A, Ruiz-Arias JA, Pozo-Vazquez D, et al. An artificial neural
network ensemble model for estimating global solar radiation from Meteosat
satellite images. Energy 2013;61:636–45.
[74] Phakamas N, Jintrawet A, Patanothai A, et al. Estimation of solar radiation
based on air temperature and application with the DSSAT v4. 5 peanut and
rice simulation models in Thailand. Agric Forest Meteorol 2013;180:182–93.
[75] Halthore RN, Crisp D, Schwartz SE, Anderson GP, Berk A, Bonnel B, et al.
Intercomparison of shortwave radiative transfer codes and measurements.
J Geophys Res 2005;110:D11206. http://dx.doi.org/10.1029/2004JD005293.
[76] Bristow KL, Campbell GS. On the relationship between incoming solar radia-
tion and daily maximum and minimum temperature. Agric Forest Meteorol
1984;31(2):159–66.
[77] Shi GY, Hayasaka T, Ohmura A, Chen ZH, Wang B, Zhao JQ, et al. Data quality
assessment and the long-term trend of ground solar radiation in China. J Appl
Meteorol Climatol 2008;47:1006–16.
[78] Tang WJ, Yang K, Qin J, Cheng C, He J. Solar radiation trend across China in
recent decades: a revisit with quality-controlled data. Atmos Chem Phys
2011;11:393–406.
[79] Tang W, Yang K, He J, Qin J. Quality control and estimation of global solar
radiation in China. Sol Energy 2010;84:466–75.
[80] Mishra AK, Desai VR. Drought forecasting using feed-forward recursive neural
network. Ecol Model 2006;198:127–38.
[81] Zounemat-Kermani M. Hourly predictive Levenberg–Marquardt ANN and
multi linear regression models for predicting of dew point temperature.
Meteorol Atmos Phys 2012;117:181–92.
[82] Riahi-Madvar H, Ayyoubzadeh SA, Gholizadeh Atani M. Developing an expert
system for predicting alluvial channel geometry using ANN. Expert Syst Appl
2011;38:215–22.
[83] Mukherjee S, Ashish K, Baran H, Chattopadhyay S. Modeling depression data:
feed forward neural network vs. radial basis function neural network. Am J
Biomed Sci 2014;6:166–74.
[84] Zounemat-Kermani M, Kisi O, Rajaee T. Performance of radial basis and LM-
feed forward artificial neural networks for predicting daily watershed runoff.
Appl Soft Comput 2013;13:4633–44.
[85] Firat M, Gungor M. Generalized Regression Neural Networks and Feed For-
ward Neural Networks for prediction of scour depth around bridge piers. Adv
Eng Softw 2009;40:731–7.
[86] Zounemat-Kermani M. Principal component analysis (PCA) for estimating
chlorophyll concentration using forward and generalized regression neural
networks. Appl Artif Intell 2014;28:16–29.
[87] Ouammi A, Zejli D, Dagdougui H, Benchrifa R. Artificial neural network ana-
lysis of Moroccan solar potential. Renew Sustain Energy Rev 2012;16:4876–89.
[88] ASCE Task Committee. Artificial neural networks in hydrology II: hydrological
applications. J Hydrol Eng ASCE 2000;5(2):124–37.
[89] Kisi O. Daily pan evaporation modelling using a neuro-fuzzy computing
technique. J Hydrol 2013;329:636–46.
[90] Kisi O. Multi-layer perceptrons with Levenberg-Marquardt training algorithm
for suspended sediment concentration prediction and estimation. Hydrol Sci J
2004;49(6):1025–40.
[91] Seckin N, Cobaner M, Yurtal R, Haktanir T. Comparison of artificial neural
network methods with L-moments for estimating flood flow at ungauged
sites: the Case of East Mediterranean River Basin, Turkey. Water Resour Manag
2013;27:2103–24.
[92] Pinar E, Paydas K, Seckin G, Akilli H, Sahin B, Cobaner M, et al. Artificial neural
network approaches for prediction of backwater through arched bridge con-
strictions. Adv Engine Soft 2010;41(4):627–35.
[93] Haykin S. Neural networks – a comprehensive foundation. 2nd ed.Prentice-
Hall; 1998.
[94] Cigizoğlu HK. Application of generalized regression neural networks to inter-
mittent flow forecasting and estimation. J Hydrol Eng ASCE 2005;10(4):336–41.
[95] Kisi O. Generalized regression neural networks for evapotranspiration mod-
elling. Hydrol Sci J 2006;51(6):1092–105.