Professional Documents
Culture Documents
PII: S1364-6826(16)30148-1
DOI: http://dx.doi.org/10.1016/j.jastp.2016.06.006
Reference: ATP4439
To appear in: Journal of Atmospheric and Solar-Terrestrial Physics
Received date: 6 April 2016
Revised date: 11 June 2016
Accepted date: 13 June 2016
Cite this article as: Saeid Mehdizadeh, Javad Behmanesh and Keivan Khalili,
Comparison of Artificial Intelligence Methods and Empirical Equations to
Estimate Daily Solar Radiation, Journal of Atmospheric and Solar-Terrestrial
Physics, http://dx.doi.org/10.1016/j.jastp.2016.06.006
This is a PDF file of an unedited manuscript that has been accepted for
publication. As a service to our customers we are providing this early version of
the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting galley proof before it is published in its final citable form.
Please note that during the production process errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal pertain.
Comparison of Artificial Intelligence Methods and Empirical Equations to Estimate Daily
Solar Radiation
Email: saied.mehdizadeh@gmail.com
Email: j.behmanesh@urmia.ac.ir
Email: khalili2006@gmail.com
*
Corresponding author Email address:
Abstract
In the present research, three artificial intelligence methods including Gene Expression Programming (GEP),
Artificial Neural Networks (ANN) and Adaptive Neuro-Fuzzy Inference System (ANFIS) as well as, 48
empirical equations (10, 12 and 26 equations were temperature-based, sunshine-based and meteorological
parameters-based, respectively) were used to estimate daily solar radiation in Kerman, Iran in the period of
1992 to 2009. To develop the GEP, ANN and ANFIS models, depending on the used empirical equations,
various combinations of minimum air temperature, maximum air temperature, mean air temperature,
extraterrestrial radiation, actual sunshine duration, maximum possible sunshine duration, sunshine duration
ratio, relative humidity and precipitation were considered as inputs in the mentioned intelligence methods. To
compare the accuracy of empirical equations and intelligence models, root mean square error (RMSE), mean
absolute error (MAE), mean absolute relative error (MARE) and determination coefficient (R2) indices were
used. The results showed that in general, sunshine-based and meteorological parameters-based scenarios in
ANN and ANFIS models presented high accuracy than mentioned empirical equations. Moreover, the most
accurate method in the studied region was ANN11 scenario with five inputs. The values of RMSE, MAE,
MARE and R2 indices for the mentioned model were 1.850 MJ m-2 day-1, 1.184 MJ m-2 day-1, 9.58 % and
0.935, respectively.
Keywords: Adaptive Neuro-Fuzzy Inference System, Artificial Neural Networks, Empirical Equations, Gene
Expression Programming, Solar Radiation
1
Tel: +98.9143435991.
2
Tel: +98.9141460272.
3
Tel: +98.9149393259.
1
Nomenclature
H: Daily solar radiation (MJ m-2 day-1) Igs: Solar constant (1367 W m-2)
Ho: Daily extraterrestrial solar radiation (MJ m-2 day-1) f: Eccentricity correction factor
S: Sunshine duration (h) : Latitude of the station ()
So: Maximum possible sunshine duration (h) w s: Sunrise hour angle ()
Tmin: Daily minimum air temperature (C) : Solar declination ()
Tmax: Daily maximum air temperature (C) J: Julian day (starting from the first January)
T: Daily mean air temperature (C) Ps: Atmospheric pressure at the station (KPa)
RH: Daily mean relative humidity (%) P: Atmospheric pressure at sea level (KPa)
ST: Daily mean soil temperature (C) RMSE: Root mean square error (MJ m-2 day-1)
P: Precipitation (cm in empirical Eq. 25 and mm in empirical Eq. 35) MAE: Mean absolute error (MJ m -2 day-1)
Tk: Daily mean absolute air temperature (K) R2: Determination coefficient
w: Atmospheric precipitable water vapor per unit volume of air (cm) MARE: Mean absolute relative error (%)
Hmod: Daily estimated solar radiation corrected for systematic bias (MJ m-2 day-1)
Z: Elevation of the station (m in empirical Eq. 5 and Km in empirical Eqs. 13 and 16)
a to g: Calibrated coefficients of empirical equations : Membership function
1. Introduction
Solar radiation is one of the most important parameters which is used to design solar systems (Mousavi et al.,
2015), atmospheric energy balance studies, agricultural studies and meteorological forecasting (Ozgoren et
al., 2012), solar energy applications (Janjai et al., 2011), determining irrigation water requirements and
potential yield of crops (Almorox et al., 2013) and agronomy (Liu et al., 2015). But, despite the importance
of the measurement of this parameter, solar radiation is not a routinely measured meteorological parameter as
temperature or rainfall (Liu et al., 2015; Iziomon and Mayer, 2002; Almorox et al., 2005; Mubiru et al.,
2007). Moreover, direct measurements of solar radiation have been faced many problems. Some of these
problems are: calibration problems, problems with accumulation of dirt and water on the sensors
(Rahimikhoob, 2010). Even, at stations where solar radiation is measured, there could be many days which
radiation data are missing or lie outside the expected range due to the equipment failure and other problems
(Hunt et al., 1998). These problems have been encouraged researchers to use the empirical models for
estimating solar radiation (Menges et al., 2006; Ertekin and Evrendilek, 2007; Bakirci, 2009a; Sonmete et al.,
2010; Besharat et al., 2013; Yao et al., 2014). Besharat et al. (2013) reviewed empirical models for
estimating solar radiation. They classified existent models into four categories; sunshine-based, cloud-based,
temperature-based and other meteorological parameters-based models. Then, they have carried out a case
study in Yazd, Iran. Their results showed that sunshine-based models had the highest accuracy. Yao et al.
(2014) evaluated 108 empirical models for estimation of solar radiation in Shanghai, China. They reported
that polynomial models were the most accurate models between other used models. Das et al. (2015) applied
17 sunshine duration-based models (6 linear and 11 nonlinear models) in South Korea. They concluded that
2
one of nonlinear model had the best performance which means that nonlinear relationship exist between
sunshine duration and clearness index.
Beside the empirical equations, artificial intelligence methods are as a powerful tool to estimate
meteorological parameters in non-linear systems such as solar radiation. In the recent years, artificial
intelligence methods e.g. Gene Expression Programming (GEP), Artificial Neural Networks (ANN) and
Adaptive Neuro-Fuzzy Inference System (ANFIS) have been used successfully in many engineering sciences
(e.g. estimating the solar radiation). However, conducted studies in the field of solar radiation estimation
using GEP are rare. Reddy and Ranjan (2003) estimated solar radiation in 13 locations in north and south of
India using existent meteorological data and geographical coordinates. They used Angstrom-Prescott,
Hargreaves, Supit and Van Kappel models and ANN technique. It is concluded that ANN was the most
accurate model. Tymvios et al. (2005) predicted solar radiation using ANN and Angstrom-Prescott linear
model at Athallasa in Cyprus radiometric station. ANN model with inputs of sunshine duration, maximum
possible sunshine duration and maximum temperature had the highest accuracy. Ozgoren et al. (2012) by
considering 10 different scenarios from monthly meteorological parameters and also geographical
coordinates estimated monthly solar radiation in 31 stations in Turkey using ANN and MNLR (Multiple
Non-Linear Regression). The results showed that ANN with full inputs was the best model. Landeras et al.
(2012) compared the ability of three artificial intelligence techniques including GEP, ANN and ANFIS with
two empirical equations of Hargreaves-Samani and Mahmood and Hubbard to estimate solar radiation in four
stations in Basque region in northern Spain. They developed used artificial intelligence techniques based on
five different defined scenarios using different combinations of extraterrestrial radiation, minimum and
maximum air temperatures, day of the year and corrected clear sky solar radiation. The results indicated that
ANN structure with four inputs of extraterrestrial radiation, minimum and maximum air temperatures, day of
the year and 10 neurons in the hidden layer was the most accurate model among studied models. Citakoglu
(2015) compared the accuracy of ANN, ANFIS, MLR (Multiple Linear Regression) to estimate monthly
solar radiation with four empirical equations including Angstrom, Abdalla, Bahel and Hargreaves-Samani in
163 stations in Turkey. In the mentioned research, 11 different combinations of month number,
extraterrestrial radiation, relative humidity, mean air temperature as well as geographical coordinates were
considered as inputs for ANN, ANFIS, and MLR models. The results showed that the estimation accuracy of
ANN was greater than ANFIS, MLR and used empirical equations in the studied region. Piri and Kisi (2015)
compared the ANFIS, NN-ARX (neural network auto regressive model with exogenous inputs) and empirical
models in two Iranian stations, Zahedan and Bojnurd. ANFIS models are found to perform better than other
methods. Mohammadi et al. (2015a) applied ANFIS and six empirical equations to estimate daily solar
radiation in Tabass, Iran by using day of the year parameter as a sole input. The results revealed the
superiority of ANFIS model.
From literature review, it is concluded that conducted studies in the field of the use of GEP for estimation of
solar radiation are rare. Therefore, the objectives of the present study are: 1) evaluate the performance of 48
empirical equations including 10 temperature-based, 12 sunshine-based and 26 different meteorological
parameters-based models, 2) investigating the capability of artificial intelligence methods including Gene
Expression Programming (GEP), Artificial Neural Networks (ANN) and Adaptive Neuro-Fuzzy Inference
3
System (ANFIS) and 3) comparative analysis of empirical equations with intelligence methods for estimating
daily solar radiation in Kerman, Iran.
Fig. 1
Table 1
Table 2a
Table 2b
Table 2c
Nomenclature
It can be seen that in all empirical equations, extraterrestrial solar radiation (Ho) is required for estimating
solar radiation. Ho can be calculated by Eq. 1 (Duffie and Beckman, 1991),
[ ] (1)
4
Where, Ho is the extraterrestrial radiation (MJ m-2 day-1), Igs is the solar constant (1367 W m-2), f is the
eccentricity correction factor, is the latitude of the station (), is the solar declination () and is the
sunrise hour angle (). f, and are calculated using Eqs. 2-4 (Duffie and Beckman, 1991),
( ) (2)
( ) (3)
(4)
Where, J is the Julian day starting from first January. The maximum day length or maximum sunshine
duration (So) is calculated by Eq. 5 (Duffie and Beckman, 1991),
(5)
Table 3
5
is usually employed (input layer, hidden layer, output layer). Input layer is a transferring layer and as a tool
for generating data. Output layer includes the estimated values by network and hidden layer is formed from
processor nodes so that these nodes are the place for data processing. In each layer, there are one or more
processing elements (neuron) which have the relationship with all neurons in the next layer. In fact, the
number of neurons in input layer is the number of input parameters. In the present study, input parameters
were included different combinations of Tmin, Tmax, Ho, S, So, S/So, RH, T and P. Output layer was daily solar
radiation. Number of neurons in the hidden layer was determined by trial and error procedure. For example,
Fig. 2 shows the structure of ANN1 model. It should be noted that in Fig. 2, Wij and Bj are weights and biases
between input and hidden layers, respectively. Also, Wjk is weight between hidden layers and output layer.
Neurons are the main components of neural networks. The input pattern to a node can be shown using a
vector with N components (X=(X1, X2, , Xn)). Summation of inputs multiplication in correspond weights
can be shown by a scalar quantity (S),
(6)
Where, W=(W1, W2, , Wn) is the weight vector of neurons. To obtain the output, the quantity of S enters
into the nonlinear function (f),
In the hidden layer, the logistic and tangent sigmoid functions are generally used as the transfer function.
In ANN process, the output of y can be the entrance of next layer in multilayer networks.
To develop the ANN models, ANN toolbox in MATLAB R2014a was used. In this study, feed forward
neural networks with learning algorithm of Levenberg-Marquardt (LM) were employed. The LM algorithm is
claimed to be superior to the standard back-propagation in terms of fast convergence and thus needs lesser
learning cycles (Zanetti et al., 2007).
Fig. 2
6
Where, H is the Hessian matrix of coefficients (matrix of second derivatives) and is a parameter for
stabilizing. Small values of parameter result in a Gauss-Newton update and large values of result in a
steepest descent update. The parameter is initialized to be large so that first updates are small steps in the
steepest descent direction. If an iteration happens to result in a worse approximation, is increased. As the
solution improves, is decreased, the LM method approaches the Gauss-Newton method and the solution
typically accelerates to the local minimum (Marquardt, 1963; Madsen et al., 2004; Lourakis, 2005).
7
To develop ANFIS models, MATLAB R2014a was used. In ANFIS model, there are two common methods
including grid partition (GP) and subtractive clustering (SC). In the present study, GP method with back
propagation learning algorithm was used in Rs estimating process.
Fig. 3
| | (21)
| | (22)
(23)
[ ][ ]
Where, Hi,m is the ith observed daily solar radiation, Hm,avg is the mean of observed values, Hi,c is the ith
estimated daily solar radiation using empirical equations and intelligence methods, Hc,avg is the mean of
estimated values and N is the number of daily solar radiation values.
8
showed good performance so that the differences of statistical indices were negligible. Moreover, the worst
model was model 13. In fact, models 11 and 13 are Angstrom-Prescott equation. However, equation 11 is
calibrated using H/Ho and S/So and equation 13 is calibrated using altitude, latitude and solar declination.
Therefore, in Kerman, calibrating Angstrom-Prescott model using regression of H/Ho and S/So instead of
geographical coordinates is recommended.
In the various meteorological parameters-based models (models 23 to 48), the different results with high
differences in the statistical indicators were observed. The values of RMSE, MAE, MARE and R2 indices
change between 1.894 MJ m-2 day-1 and 4.513 MJ m-2 day-1, 1.266 MJ m-2 day-1 and 3.545 MJ m-2 day-1, 10.07
% and 27.55 %, 0.633 and 0.934, respectively. Among the meteorological parameters-based models, model
34 (with input parameters of S/So, Tmax and RH) was the best model due to the lowest values of RMSE, MAE
and MARE (1.894 MJ m-2 day-1, 1.266 MJ m-2 day-1 and 10.07 %) and the higher value of R2 (0.933). After
the mentioned model, models 33, 24 and 32 had high accuracy in comparison of other models.
It is seen from Table 5 that models 17 (sunshine-based) and 34 (meteorological parameters-based) show
similar results and were most accurate models among empirical equations. Therefore, in the study area, the
mentioned models were recommended to estimate solar radiation. Similar to obtained results, Besharat et al.
(2013) showed that the selected sunshine-based model had the best performance in Yazd, Iran.
Table 4
Table 5
9
based models, respectively) have been presented in Table 7. Also, expression tree of the best GEP model
(GEP8) is shown in Fig. 4.
In ANN model, similar to GEP, various input combinations were utilized (see input combinations column in
Table 8). Output layer was formed from one node, daily solar radiation values. Also, number of neurons in
the hidden layer to achieve higher accuracy, was determined by using trial and error procedure. In the present
study, it is found that optimum number of neurons varied between 5 and 9. Moreover, transfer functions of
tangent-sigmoid (Tansig) and linear (Purelin) for all used scenarios, had good performance in hidden and
output layers, respectively. Values of statistical indices for ANN model in testing stage are presented in Table
8. In temperature-based scenarios, ANN2 (with Tmin, Tmax and Ho inputs) presented the good results than
ANN1 (RMSE = 3.087 MJ m-2 day-1, MAE =2.230 MJ m-2 day-1, MARE = 17.07 % and R2 = 0.819).
Therefore, it can be concluded that extraterrestrial radiation is important parameter as input in ANN model.
In sunshine-based scenarios, similar results were obtained. However, ANN3 with S and H o inputs was
slightly better than ANN4 and ANN5. In various meteorological parameters-based scenarios, performance
difference of the models was negligible. However, ANN11 with full inputs was superior model (RMSE =
1.850 MJ m-2 day-1, MAE =1.184 MJ m-2 day-1, MARE = 9.58 % and R2 = 0.935). Weights and biases
between input and hidden layers as well as weights between hidden and output layers for the best ANN
scenario (ANN11) are presented in Table 9.
In ANFIS model, used input and output parameters were similar to GEP and ANN methods. In this model to
obtain the higher accuracy, suitable membership functions and optimum number of membership functions
should be used. Suitable membership functions and their optimum numbers determined by trial and error. In
the present research, Trimmf (Triangular membership function) and Gaussmf (Gaussian membership
function) for input parameters showed good performances. Moreover, constant and linear membership
functions for output variable were tested. Statistical indices values for ANFIS model in testing stage have
been presented in Table 10. The results of ANFIS model were very similar to ANN. In temperature-based
scenarios, similar to ANN and slightly GEP, ANFIS2 showed higher accuracy than ANFIS1. The
performance difference between sunshine-based and various meteorological parameters-based scenarios was
not considerable. However, ANFIS5 with S, So and Ho inputs (RMSE = 1.889 MJ m-2 day-1, MAE =1.259 MJ
m-2 day-1, MARE = 10.08 % and R2 = 0.933) and ANFIS10 with inputs of S/So, RH, P, Ho (RMSE = 1.842
MJ m-2 day-1, MAE =1.239 MJ m-2 day-1, MARE = 9.73 % and R2 = 0.936) were the best between two studied
scenarios.
In general, the performance of sunshine-based and meteorological parameters-based scenarios was better
than temperature-based scenarios in GEP, ANN and ANFIS models. Similar results were reported by Chen
and Li (2014) at 15 stations in China using support vector machine intelligence technique, so that sunshine-
based scenarios were superior to temperature-based models. Also, the most accurate scenarios in the studied
region were GEP8 (MARE = 10.15 %), ANN11 (MARE = 9.58 %) and ANFIS10 (MARE = 9.73 %).
Table 6
Table 7
Fig. 4
Table 8
10
Table 9
Table 10
Table 11
4. Conclusion
In the present study, the performance of 48 empirical equations (10 temperature-based, 12 sunshine-based
and 26 meteorological parameters-based) to estimate daily solar radiation was investigated in Kerman, Iran.
Then, the ability of three artificial intelligence methods including GEP, ANN and ANFIS was evaluated and
compared with empirical equations. Among the empirical equations, model 17 (sunshine-based) and model
34 (meteorological parameters-based) had the best performance in comparison of other equations. In general,
sunshine-based equations were the best methods and temperature-based equations had the worst
performance. Modeling results of GEP, ANN and ANFIS showed that the mentioned methods were able to
estimate daily solar radiation. However, in all three types of scenarios (temperature-based, sunshine-based
and various meteorological parameters-based), ANN and ANFIS models had minimum error than GEP.
Between artificial intelligence methods, ANN11 with full inputs was the most accurate scenario (RMSE =
1.850 MJ m-2 day-1, MAE = 1.184 MJ m-2 day-1, MARE = 9.58 % and R2 = 0.935). Also generally, estimation
accuracy of sunshine-based and meteorological parameters-based scenarios in ANN and ANFIS methods
were better than mentioned empirical equations. However, calculations in ANN and ANFIS models were
heavier and time-consuming. The results of this study can be used in similar climates in Iran. In the following
the present research, the use of other empirical equations as well as, other artificial intelligence methods such
as support vector machine and wavelet neural networks are recommended to estimate solar radiation.
References
11
Abdalla, Y.A.G., 1994. New correlation of global solar radiation with meteorological parameters for Bahrain.
Int. J. Sol. Energy 16, 111120.
Allen, R.G., 1997. Self-calibrating method for estimating solar radiation from air temperature. J. Hydrol.
Eng. 2, 5667.
Almorox, J., Hontoria, C., 2004. Global solar radiation estimation using sunshine duration in Spain. Energy
Convers. Manage. 45, 15291535.
Almorox, J., Benito, M., Hontoria, C., 2005. Estimation of monthly AngstromPrescott equation coefficients
from measured daily data in Toledo, Spain. Renew. Energy 30, 931936.
Almorox, J., Bocco, M., Willington, E., 2013. Estimation of daily global solar radiation from measured
temperatures at Caada de Luque, Crdoba, Argentina. Renew. Energy 60, 382387.
Ampratwum, D.B., Dorvlo, A.S.S., 1999. Estimation of solar radiation from the number of sunshine hours.
Appl. Energy 63, 161167.
Angstrom, A., 1924. Solar and terrestrial radiation. Q. J. R. Meteorol. Soc. 50, 121125.
Annandale, J.G., Jovanic, N.Z., Benade, N., Allen, R.G., 2002. Software for missing data error analysis of
PenmanMonteith reference evapotranspiration. Irrig. Sci. 21, 5767.
Bahel, V., Bakhsh, H., Srinivasan, R., 1987. A correlation for estimation of global solar radiation. Energy 12,
131135.
Bakirci, K., 2009a. Models of solar radiation with hours of bright sunshine: A review. Renew. Sustain.
Energy Reviews 13, 25802588.
Bakirci, K., 2009b. Correlations for estimation of daily global solar radiation with hours of bright sunshine in
Turkey. Energy 34, 485501.
Besharat, F., Dehghan, A.A., Faghih, A.R., 2013. Empirical models for estimating global solar radiation: A
review and case study. Renew. Sustain. Energy Reviews 21, 798821.
Chen, R., Ersi, K., Yang, J., Lu, S., Zhao, W., 2004. Validation of five global radiation models with
measured daily data in China. Energy Convers. Manag. 45, 17591769.
Chen, J.L., Li, G.S., 2014. Evaluation of support vector machine for estimation of solar radiation from
measured meteorological variables. Theor. Appl. Climatol. 115, 627-638.
Citakoglu, H., 2015. Comparison of artificial intelligence techniques via empirical equations for prediction of
solar radiation. Comput. Electron. Agric. 118, 2837.
Das, A., Park, J.K., Park, J.H., 2015. Estimation of available global solar radiation using sunshine duration
over South Korea. J. Atmos. Sol. Terr. Phys. 134, 2229.
Duffie, J.A., Beckman, W.A., 1991. Solar engineering of thermal processes. New York; Wiley.
Ertekin, C., Evrendilek, F., 2007. Spatio-temporal modeling of global solar radiation dynamics as a function
of sunshine duration for Turkey. Agric. For. Meteorol. 145, 3647.
Ferreira, C., 2001. Gene expression programming: a new adaptive algorithm for solving problems. Complex
Syst. 13, 87129.
Garg, H.P., Garg, S.T., 1982. Prediction of global solar radiation from bright sunshine hours and other
meteorological parameters, Solar-India, proceedings on national solar energy convention. New Delhi:
Allied Publishers, p. 1.0047.
12
Gariepy, J., 1980. Estimation of global solar radiation. International report, Service of meteorology,
Government of Quebec, Canada.
Gopinathan, K.K., 1988. A simple method for predicting global solar radiation on a horizontal surface. Sol.
Wind Technol. 5, 581583.
Hargreaves, G.H., Samani, Z.A., 1982. Estimating potential evapotranspiration. J. Irrig. Drain. Eng. 108,
223230.
Hunt, L.A., Kuchar, L., Swanton, C.J., 1998. Estimation of solar radiation for use in crop modeling. Agric.
For. Meteorol. 91, 293300.
Iziomon, M.G., Mayer, H., 2002. Assessment of some global solar radiation parameterizations. J. Atmos. Sol.
Terr. Phys. 64, 16311643.
Jang, J.S.R., 1993. ANFIS: adaptive-network-based fuzzy inference system. IEEE Syst. Man. Cyb. Soc. 23,
665685.
Janjai, S., Sricharoen, K., Pattarapanitchai, S., 2011. Semi-empirical models for the estimation of clear sky
solar global and direct normal irradiances in the tropics. Appl. Energy 88, 47494755.
Jiang, Y., 2009. Estimation of monthly mean daily diffuse radiation in China. Appl. Energy 86, 14581464.
Kilic, A., Ozturk, A., 1983. Solar energy. Istanbul: Kipas Yayincilik [in Turkish].
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. Manage. 62, 113.
Lewis, G., 1983. Estimates of irradiance over Zimbabwe. Sol. Energy 31, 609612.
Liu, J., Linderholm, H., Chen, D., Zhou, X., Flerchinger, G.N., Yu, Q., Du, J., Wu, D., Shen, Y., Yang, Z.,
2015. Changes in the relationship between solar radiation and sunshine duration in large cities of
China. Energy 82, 589600.
Louche, A., Notton, G., Poggi, P., Simonnot, G., 1991. Correlations for direct normal and global horizontal
irradiation on a French Mediterranean site. Sol. Energy 46, 261266.
Lourakis, M.I.A., 2005. A Brief description of the Levenberg Marquardet algorithm implemented by
LEVMAR Institute of Computer Science Foundation for Research and Technology - Hellas
(FORTH).
Madsen, K., Nielsen, H.B., Tingleff, O., 2004. Methods for nonlinear least squares problems, Technical
Report. Informatics and Mathematical Modeling, Technical University of Denmark.
Mahmood, R., Hubbard, K.G., 2002. Effect of time of temperature observation and estimation of daily solar
radiation for the Northern Great Plains, USA. Agron. J. 94, 723733.
Marquardt, D.W., 1963. An algorithm for least-squares estimation of nonlinear parameters. J. Soc. Ind. Appl.
Math. 11, 431441.
Menges, H.O., Ertekin, C., Sonmete, M.H., 2006. Evaluation of global solar radiation models for Konya,
Turkey. Energy Convers. Manage. 47, 31493173.
Mohammadi, K., Shamshirband, S., Tong, C.W., Alam, K.A., Petkovic, D., 2015a. Potential of adaptive
neuro-fuzzy system for prediction of daily global solar radiation by day of the year. Energy Convers.
Manag. 93, 406413.
13
Mohammadi, K., Shamshirband, S., Anisi, M.H., Alam, K.A., Petkovic, D., 2015b. Support vector regression
based prediction of global solar radiation on a horizontal surface. Energy Convers. Manage. 91, 433
441.
Mousavi, S.M., Mostafavi, E.S., Jaafari, A., Jaafari, A., Hosseinpour, F., 2015. Using Measured Daily
Meteorological Parameters to Predict Daily Solar radiation. Meas. 76, 148155.
Mubiru, J., Banda, E.J.K.B., D'Ujanga, F., Senyonga, T., 2007. Assessing the performance of global solar
radiation empirical formulations in Kumpala, Uganda. Theor. Appl. Climatol. 87, 179184.
Newland, F.J., 1989. A study of solar radiation models for the coastal region of South China. Sol. Energy 43,
227235.
Ododo, J.C., Sulaiman, A.T., Aidan, J., Yuguda, M.M., Ogbu, F.A., 1995. The importance of maximum air
temperature in the parameterisation of solar radiation in Nigeria. Renew. Energy 6, 751763.
Ogelman, H., Ecevit, A., Tasdemiroglu, E., 1984. A new method for estimating solar radiation from bright
sunshine data. Sol. Energy 33, 619625.
Ozgoren, M., Bilgili, M., Sahin, B., 2012. Estimation of global solar radiation using ANN over Turkey.
Expert Syst. Appl. 39, 50435051.
Piri, J., Kisi, O., 2015. Modelling solar radiation reached to the Earth using ANFIS, NN-ARX, and empirical
models (case studies: Zahedan and Bojnurd stations). J. Atmos. Sol. Terr. Phys. 123, 3947.
Prescott, J.A., 1940. Evaporation from water surface in relation to solar radiation. Trans. R. Soc. Australia
46, 114118.
Rahimikhoob, A., 2010. Estimating global solar radiation using artificial neural network and air temperature
data in a semi-arid environment. Renew. Energy 35, 21312135.
Reddy, K.S., Ranjan, M., 2003. Solar resource estimation using artificial neural networks and comparison
with other correlation models. Energy Convers. Manage. 44, 25192530.
Richardson, C.W., 1985. Weather simulation for crop management models. Trans. ASAE. 28, 16021606.
Rietveld, M.R., 1978. A new method for estimating the regression coefficients in the formula relating solar
radiation to sunshine. Agric. Meteorol. 19, 243252.
Sonmete, M.H., Ertekin, C., Menges, H.O., Hacseferogullari, H., Evrendilek, F., 2010. Assessing monthly
average solar radiation models: a comparative case study in Turkey. Environ. Monit. Assess. 175,
251-277.
Swartman, R.K., Ogunlade, O., 1967. Solar radiation estimates from common parameters. Sol. Energy 11,
170172.
Togrul, I.T., Onat, E., 1999. A study for estimating solar radiation in Elazig using geographical and
meteorological data. Energy Convers. Manage. 40, 15771584.
Tymvios, F.S., Jacovides, C.P., Michaelides, S.C., Scouteli, C., 2005. Comparative study of Angstrom and
artificial neural network methodologies in estimating global solar radiation. Sol. Energy 78, 752762.
UNEP, 1993. World Atlas of Desertification, The United Nations Environment Programme (UNEP),
London.
Yao, W., Li, Z., Wang, Y., Jiang, F., Hu, L., 2014. Evaluation of global solar radiation models for Shanghai,
China. Energy Convers. Manage. 84, 597612.
14
Zanetti, S.S., Sousa, E.F., Oliveira, V.P., Almeida, F.T., Bernardo, S., 2007. Estimating evapotranspiration
using artificial neural network and minimum climatological data. J. Irrig. Drain. Eng. 133, 8389.
W11
W11
H
Bj=B5
i j k
Fig. 2 Architecture of ANN1 model with the structure of 2-5-1 (two inputs, one output and five neurons in the
hidden layer)
15
Layer 1 Layer 2 Layer 3 Layer 4 Layer 5
A x y
1 w1
x
N
A
2
f
N
B
w2
1
y x y
B
2
Fig. 3 Structure of an ANFIS model with two inputs, one output and two rules
16
Fig. 4 Expression tree of GEP8 model, functions set: +,,*,/,Ln, terminals set: do, d1, d2 denote to S/So, P, Ho,
respectively
17
Table 2a Temperature-based empirical equations
Model No. Equation Reference
2 Richardson (1985)
3 Allen (1997)
( )
12 Rietveld (1978)
( )
16 Gopinathan (1988)
( )
17 ( ) ( ) Newland (1989)
18
19 Ampratwum and Dorvlo (1999)
21 ( ) Bakirci (2009b)
22 Bakirci (2009b)
25 Gariepy (1980)
( )
28 Lewis (1983)
29 Lewis (1983)
30 Lewis (1983)
31 Lewis (1983)
32 ( ) Abdalla (1994)
19
42 ( ) Togrul and Onat (1999)
0.1
8 Inversion rate 0.1 Gene recombination rate
Head size
0.1
3 One-point recombination 0.3 Gene transposition rate
Number of genes
20
19 0.716 0.349
20 0.126 0.242
21 0.530 0.906 -0.256
22 0.723 0.310
23 83.123 0.222 -0.406
24 0.376 0.409 -0.001
25 0.310 0.553
26 0.326 0.443 -0.007
27 0.830 0.003 -0.190
28 141.554 -0.596
29 32.464 0.288 -0.331
30 7.484 0.105
31 33.714 -0.017
32 0.368 0.411 0.0002 -0.0009
33 1.030 0.285 -0.023 -0.088
34 0.286 0.497 0.004 -0.0006 -0.004
21
9 3.184 2.420 17.61 0.812
10 3.940 3.020 25.04 0.707
11 1.942 1.294 10.83 0.933
12 2.876 2.052 13.60 0.906
13 3.655 3.334 18.38 0.932
14 1.895 1.280 10.13 0.935
15 1.895 1.280 10.12 0.935
22
GEP1 Tmin, Tmax 4.245 3.414 26.95 0.678
Temperature-based
GEP2 Tmin, Tmax, Ho 4.130 3.162 22.91 0.676
GEP3 S, Ho 1.941 1.284 10.91 0.923
Sunshine-based GEP4 S/So, Ho 1.921 1.297 10.62 0.933
GEP5 S, So, Ho 2.073 1.396 10.88 0.919
GEP6 S/So, RH, Ho 2.254 1.590 11.47 0.908
GEP7 S/So, T, Ho 2.312 1.719 11.90 0.903
Sunshine-based GEP4 [ ]
Various meteorological ( )
GEP8 [ ] [ ]
parameters-based
23
Table 9 Weights and biases for the best ANN model (ANN11)
Weights between input (i)
Wij
and hidden (j) layers
i Wi1 Wi2 Wi3 Wi4 Wi5 Wi6 Wi7
1 0.480 0.758 -0.709 0.883 -0.208 -1.533 -0.740
2 0.232 1.233 1.100 -0.712 0.555 0.658 -0.494
3 0.360 -0.789 -0.336 -0.018 0.802 1.019 0.248
4 -1.049 1.124 1.305 0.264 -0.005 0.663 -0.688
5 0.583 -0.578 0.740 -0.878 -0.902 -0.357 -1.694
Biases between input (i) B1 B2 B3 B4 B5 B6 B7
and hidden (j) layers -1.841 -1.690 0.810 0.689 0.824 -0.933 -1.600
Weights between hidden (j) W11 W21 W31 W41 W51 W61 W71
and output (k) layers 0.504 -0.268 0.073 0.231 -0.362 -0.180 -0.158
24
ANN11 1.850 1.184 9.58 0.935
ANFIS10 1.842 1.239 9.73 0.936
*
Bold values indicate the statistical indices of the best models
Highlights
25