See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/319490145

An application of the Multilayer Perceptron: Estimation of Global Solar

Radiation and the Establishment of Solar Radiation Maps of Togo " An
application of the Multilayer Perceptro...

Article · August 2017

CITATIONS READS

0 468

7 authors, including:

Amou Komi Apélété Tchamye Tcha-Esso Boroze

Université de Lomé Université de Lomé
11 PUBLICATIONS 16 CITATIONS 9 PUBLICATIONS 41 CITATIONS

SEE PROFILE SEE PROFILE

S. Ouro-Djobo Sagna Koffi

Université de Lomé Université de Lomé
18 PUBLICATIONS 232 CITATIONS 21 PUBLICATIONS 17 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Numerical study of turbulent flows in cooling channels View project

Identification and measurement of atmospheric pollutants in agglomerations and industrial zones in Togo View project

All content following this page was uploaded by Sagna Koffi on 05 September 2017.

The user has requested enhancement of the downloaded file.

Sustainable Energy, 2017, Vol. 5, No. 1, 6-15
Available online at http://pubs.sciepub.com/rse/5/1/2
©Science and Education Publishing
DOI:10.12691/rse-5-1-2

An application of the Multilayer Perceptron: Estimation

of Global Solar Radiation and the Establishment
of Solar Radiation Maps of Togo
Komi Apélété AMOU*, Tchamye Tcha-Esso BOROZE, Sanoussi OURO-DJOBO,
Koffi SAGNA, Yaovi Ouézou AZOUMA, Magolmèèna BANNA, Kossi NAPO

Solar Energy Laboratory, Department of Physics, Faculty of Sciences, University of Lomé, Lomé, Togo
*Corresponding author: [email protected]

Abstract This paper presents a new neural network approach for the generation of synthetic monthly radiation
data for nine localities in Togo. The neural model employed is the well-known Multi-Layer Perceptron (MLP)
paradigm, in feedback architecture, using a record of historical values for the supervised network training. The
method is based on the MLP ability to extract, from a sufficiently general training set, the existing relationships
between variables whose interdependence is unknown a priori. Simulation results are compared to the measured
values for the three towns where solar irradiation is measured in Togo. The results show that the generated values
are of the real values. The method has been developed using data values from Lomé, Atakpamé and Mango, and is
generalized to generate data of any location for the establishment of solar maps. Indeed, the proposed methodology
is of general applicability to the estimation of highly complex temporal series.
Keywords: estimation, neural model multi, Layer Perceptron (MLP), solar radiation maps
Cite This Article: Komi Apélété AMOU, Tchamye Tcha-Esso BOROZE, Sanoussi OURO-DJOBO,
Koffi SAGNA, Yaovi Ouézou AZOUMA, Magolmèèna BANNA, and Kossi NAPO, “An application of the
Multilayer Perceptron: Estimation of Global Solar Radiation and the Establishment of Solar Radiation Maps of
Togo.” Sustainable Energy, vol. 5, no. 1 (2017): 6-15. doi: 10.12691/rse-5-1-2.

designers, to be used in a first approach in the solar system

design and sizing. Many procedures have been developed
1. Introduction to draw solar radiation maps. In most cases, to obtain a
solar radiation map it is necessary to know the solar
Due to uncertainty of availability of fossil fuels, increasing radiation of many points spread wide across the real zone
environmental pollution and general awareness amongst of the map where it is going to be drawn. These solar data
the common people, the green sources of energy are being can be available in several time scales. Usually there are
encouraged. The green sources of energy include solar two time scales in which the data are available: daily and
photovoltaic, solar thermal, wind, biomass, small and big hourly scales. Nevertheless, for most of the locations all
hydro, tidal, wave, ocean, etc. Among these sources of over the world the records of solar radiation data are non-
energy, wind, solar and hydro are the common ones in use. existent. Only very few locations have the privilege of
Actually, solar photovoltaic cost is competitive compared having good meteorological stations where records of
to wind and other green sources of energy. Solar solar radiation have being registered. In order to know the
photovoltaic technology is being used commonly for the behavior of solar radiation at the site of interest, long-term
generation electricity for both grid connected and data from a nearby location along with empirical, semi-
standalone power systems. For proper, economical and empirical, physical, neural networks, wavelets, fractals,
efficient development and utilization of solar energy, an etc. techniques are used. Regression type models were
accurate knowledge of the availability and variability of first applied to this problem by Angström [1] based on
solar radiation intensity both in time and special domain is sunshine duration. This model was developed for other
very critical. The availability of information about the locations by Rietveld [2], Newland [3], Gopinathan [4]
solar radiation in the zone where a solar system is going to and Akinoglu and Ecevit [5]. Furthermore, some regression
be installed is necessary for the designer of solar systems. models incorporating trigonometric functions have been
This information, in case it exists, can be available, in proposed by Dorvlo and Ampratwum [6]. One of the
several ways. The most common is in tabular form with a limitations imposed by linear regression models is that
lot of very useful information, usually large solar radiation they will underperform when used to model nonlinear
sequences, but extremely difficult to handle. Another systems. Radiation data being time dependent have also
method can be solar radiation maps of the zone where the been modeled using harmonic analysis methods by Philips
installation is going to be made. This method is usually [7] and Herrero [8]. In recent years, Artificial Neural
more efficient, easy to handle and preferable by the Network (ANN) methods have been employed for the
 Sustainable Energy 7

prediction of global solar radiation both in time and special are mesured every hour from 2002 to 2009.
domains. Instead of complex rules and mathematical Different locations, with different features were taken.
routines mentioned above, ANN model are able to learn In Table 1 these sites, with some of their characteristics
the key information patterns within a multidimensional are shown. Seven different cities have been chosen. The
information domain. Furthermore, the neural approach is election of those seven locations was done due to the
particularly suitable to solve the problem of identification different position they have. The seven places are cities in
in the presence of noisy data (Rumelhart et al. [9]; Hertz et Togo with different climates. Also these seven locations
al. [10]. Gardner and Dorling [11] concluded that neural have different altitudes upon sea level and their latitudes
networks generally give better results compared with vary almost 5, from 6,17° to 10,88° North. The hourly air
statistical linear methods, especially where the problem temperature and relative humidity data of these seven
being analysed includes nonlinear behavior. Al-Alawi and locations were recorded by the National Meteorological
Al-Hinai [12] have used ANN model to predict solar Institute of Togo.
radiation in areas not covered by direct measurement
instrumentation. The input data to the network are the Table 1. Locations in study
location, month, mean pressure, mean temperature, mean Cities Longitude(°) Latitude(°)
vapor pressure, mean relative humidity, mean wind speed
Lomé 1,25 6,17
and mean duration of sunshine. The ANN predicts solar
radiation with an accuracy of 93% and mean absolute Tabligbo 1,50 6,58
percentage error of 7.3%. Togrul and Onat [13] used Kouma-Konda 0,57 6,95
geographical and meteorological parameters along with
Atakpamé 1,12 7,58
the ANN methods for the prediction of GSR for a city in
Turkey. Kalogirou et al. [14] used a recurrent neural Sokodé 1,15 9,00
network method to estimate the maximum solar radiation Kara 1,17 9,55
using measured values of air temperature and relative
Niamtougou 1,08 9,80
humidity as input. The authors found that the correlation
coefficient varied between 98.6% and 98.8%. Sözen et al. Mango 0,47 10,37
[15] used meteorological and geographical parameters as Dapaong 0,25 10,88
input into the neural networks model for the prediction
solar potential for Turkey. In another study, Yang and
Koike [16] utilized upper air humidity values for the
estimation of solar radiation on the surface of the earth
3. Methodology
through ANN method. Other authors used MLP method as
The interest in ANNs is largely due to their ability to
can be seen from these references (Elizondo et al. [17];
mimic natural intelligence in its learning from experience
Al-Alawi and Al-Hinai [12]; Mohandes et al. [18,19];
(Wasserman [24]. They learn from examples by
Togrul and Onat [13]; Hontoria et al. [20,21]; Tasadduq et
constructing an input–output mapping without explicit
al. [22]; Kalogirou et al. [14]; Tymvios et al. [23]). The
derivation of the model equation. ANN have been used in
present paper utilizes the air temperature, relative
a broad range of applications including: pattern classification
humidity and the latitude of Lomé, Atakpamé and Mango
(Lippmann [25]; Bishop [26]), function approximation,
data as input in neural networks for the prediction of
optimization, prediction and automatic control (Pham and
global solar irradiation (GSR) on horizontal surfaces in
Liu) [27] and many others. Additionally, ANNs have been
seven other cities in Togo, and to establish its solar maps.
used extensively for meteorological applications.
The methodology proposed makes use of a neural
2. Apparatus and Data Gathering network system called Multi-Layer Perceptron (MLP)
(Haykin [28]; Hornik et al. [29]; Lippmann [25]). This
neural network has already been presented in previous
Measuring apparatus was installed on the terrace of
works (Hontoria et al. [20,21,30,31]; Zufiria et al. [32]). In
solar laboratory of University of Lomé and meteorological
those works the capabilities of the MLP to generate
station of Lomé (lat. 06°10’ North and long. 01°15’ East ),
synthetic hourly solar radiation series have been demonstrated.
at Lomé, in the terrasse of meteorological stations of
Moreover, in comparison with other methods for generating
Atakpamé (lat. 07°35’ North and long. 01°07’ East ) and
synthetic sequences of solar data, the MLP improves
Mango (lat. 10°22’ North and long. 00°28’ East). All
(Hontoria et al.) [20,30] the results obtained by those
sensors are installed on the roof-top in a position relatively
methods. In this paper, MLP is used for the estimation of
free from any external obstruction, and readily accessible
GSR based on measured temperature, relative humidity
for inspection and general cleaning. Measurements of
and latitude data.
global radiation were carried out using good precision
An ANN consists of many interconnected identical
pyranometer that respond to radiation within a 2π
simple processing units called neurons. Each connection
steradian (hemispherical) field of view, which is sensitive
to a neuron has an adjustable weight factor associated with
in the wavelength range from 0.4 µm to 1.1 µm.
it. Every neuron in the network sums its weighted inputs
The stations are scheduled to measure the global solar
to produce an internal activity level 𝑎𝑎𝑖𝑖 .
radiation (GSR) from 5:00 a.m. to 8:00 p.m. daily. The
GSR is measured every 30 seconds and after one hour, the n
meter averages and records the hourly values. The air =ai ∑wij xij − wi0 (1)
temperature, relative humidity and GSR data respectively j =1
8 Sustainable Energy

where wij is the weight of the connection from input j to - The collection of data necessary for learning and
neuron i, xij is input signal number j to neuron i, and wi0 is evaluating neural network performance.
the threshold associated with unit i. The threshold is - Determining the number of hidden neurons necessary
treated as a normal weight with the input clamped at 1. to obtain a satisfactory approximation.
The internal activity is passed through a nonlinear function j - The achievement of learning
to produce the output of the neuron yi. Several forms of - The evaluation of neural network performance at the
differentiable activation functions have been used with the end of learning.
most popular being the logistic function of the form:
1 3.3. Configuration of Our Neural Network
=f ( x) with ai > 0. (2)
1 + exp(−ai ) Configuration of the neural network. According to our
calculations, the annual solar irradiation for a given city
3.1. Implementation of the Backpropagation depends on such parameters as its latitude, its monthly
average relative humidity and average monthly temperature.
Algorithm Thus in the input of the network we have three neurons
Step 1: Initialization of connection weights. and one neuron in the output layer. For the hidden layer
These weights are randomly selected. there is no rule defined for the number of neurons that
Step 2: Spread entries: MLP with one hidden layer constitutes it. Simulating the network for different values
Propagated from the input layer to the hidden layer: of the number of neurons in the hidden layer, the optimum
results are given for a number of neurons equal to two.
p
Figure 1 shows the configuration of our neural network.
y j = f ( ∑xi vi ). (3)
i =0

Then the hidden layer to the output layer:

q
Z k = f ( ∑ y j wkj ). (4)
j =1

Zk is the output of the network.

The activation function f is the sigmoïd.
Step 3: the error backpropagation
For each example of applied learning based network
input, we calculate the error on the output layers.
Figure 1. Neural network configuration used
Ek =zk (1 − zk ) ( sk − zk ) (5)
3.4. Testing and Validation
where Sk is the desired output and the actual output is Zk.
It propagates this error on the hidden layer. The error of While tests involve checking the performance of a
each hidden layer neuron is given by: sample off neural network and its generalization ability,
p validation is sometimes used when learning. Once developed
= (
Fj y j 1− y j ) ∑wkj Ek . (6) neural network, tests are needed to check the forecast
quality of the neural model. This last step is designed to
k =1
estimate the quality of the network obtained by presenting
Step 4: Fixed connection weights examples that are not part of the training set. A rigorous
Between the input and the output layer layer: validation of the developed model results in a significant
η yi Ek .
∆wki = (7) proportion of correct predictions on all of the validation. If
network performance is not satisfactory, it will either alter
Between the hidden layer and output: the architecture of the network or change the learning base.
η xi F j
∆v ji = (8)

ŋ is a parameter to be determined.
Step 5: Completing step 2 until a stopping criterion to define.
In practice, the search for appropriate weight as
possible to approximate the desired results is complex and
very long. To do this, neural network simulator is used.

3.2. Design of a Neural Network

Neural networks perform nonlinear functions set. Their
implementation requires:
- The determination of relevant inputs and outputs, i.e.
the quantities that have a significant influence on the
phenomenon that one seeks to model. Figure 2. Error back propagation algorithm
 Sustainable Energy 9

Figure 3. Learning curve and its validation

To ensure the validity of this method, it is essential to Step 1. Making a grid of the zone that is going to be
monitor the squared error between the desired output and represented in the map. Each point of the grid corresponds
the actual output (Chatelin, C.) [33]. Figure 3 shows the to a certain real site that will have some particular features
curve obtained for learning data in January. This curve is as its latitude, its longitude and its altitude over sea level,
similar to other curves obtained for different months of the besides the solar radiation data.
year. Step 2. For each point of that grid knowledge of long
The black curve corresponds to learning and that red is sequences of solar radiation is needed.
the validation of learning. From 20 periods, both black Step 3. Once the series of solar radiation is available
and red curves are combined. This means that the network for the entire grid, it is necessary to obtain some particular
has performed learning. But at this stage the square error statistical values to draw the maps. For instance, it is
remains high. Therefore learning was continued and at the interesting to summarise all that information in 12 maps,
end of 6000 loops, the two curves while remaining one for each month of the year.
confused, error became weak. This means that the network
is well trained.
5. Results of the Statistical Evaluation
4. Proposed Methodology to Draw Solar The results of learning and simulated values were
Radiation Maps collected and summarized in the following Table 2. To
better compare the measured data and values obtained by
The preparation of solar radiation maps requires the the MLP, we calculated the relative error for each month
previous knowledge of a lot of information, which can be and MPE, MBE and RMSE of the three cities [34,35]
summarised with these steps: (Rude, J., Ammar, M. Y.).

Table 2. Relative error of each month and MPE, MBE and RMSE for each city
Lomé Atakpamé Mango
Mesured Simulated Mesured Simulated Error Mesured Simulated Error
Months Error %
radiation radiation radiation radiation % radiation radiation %
January 3756,9 3732,9 0,64 4377,3 4412,9 0,81 5631,5 5608,9 0,40
February 4407,5 4409,2 0,04 4845,2 4869,6 0,50 6041,1 6008,5 0,54
March 4782,0 4771,9 0,21 5448,1 5483,9 0,66 6273,1 6215,4 0,92
April 4959,7 4960,3 0,01 5370,5 5371,8 0,02 6016,0 6008,8 0,12
May 4615,1 4694,0 1,71 5337,0 5252,3 1,59 5641,3 5694,2 0,94
June 3788,5 3846,7 1,54 4914,5 4776,8 2,80 5149,2 5237,0 1,71
July 3955,1 3895,0 1,52 4003,7 4095,7 2,30 4509,2 4488,2 0,47
August 3908,0 3846,8 1,57 3533,5 3621,0 2,48 4363,9 4368,6 0,11
September 4498,3 4483,3 0,33 4391,9 4427,7 0,81 4926,0 4912,2 0,28
October 4964,5 4941,3 0,47 5140,5 5167,1 0,52 5776,2 5745,5 0,53
November 4853,2 4904,9 1,07 5259,1 5146,8 2,13 5396,2 5452,0 1,03
December 4193,1 4182,9 0,24 4619,1 4655,5 0,79 5235,0 5216,9 0,34
MPE 0,78 1,28 0,62
MBE 1,05 -3,38 0,20
RMSE 42,04 71,49 41,34
10 Sustainable Energy

As shown in Table 2 and the Figure 4, Figure 5 and relative humidity are available in addition to their latitude.
Figure 6 obtained, learning was done in good conditions After developing new patterns files, data were
with acceptable accuracy. From the weight determined generated by month. The solar radiation monthly average
after the validation of learning, will be calculated solar of localities obtained by neuronal method is reported in
radiation data for cities whose monthly temperature data, the Table 3.

Figure 4. Estimated and simulated data for Lomé

Atakpamé
Monthly solar irradiation (Wh/d)

6000

5500

5000

4500

4000

3500

3000
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Months
Mesured radiation Simulated radiation

Figure 5. Estimated and simulated data for Atakpamé

6500
Monthly solar irradiation (Wh/d)

Mango
6000

5500

5000

4500

4000

3500
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Months
Mesured radiation Simulated radiation

Figure 6. Estimated and simulated data for Mango

 Sustainable Energy 11

Table 3. The monthly average of solar radiation (Wh /m2/day) for different cities in Togo

Localities Lomé Tabligbo Kouma-Konda Atakpamé Sokodé Kara Niamtougou Mango Dapaong
January 3747,8 3901,7 4164,7 4349,6 4938,2 5322,4 5420,9 5531,5 5712,4
February 4427,4 4567,3 4698,3 4957,8 5456,9 5838,0 5920,8 6029,5 6136,3
March 4775,8 4986,1 5091,3 5468,1 5844,5 6097,3 6124,4 6211,2 6264,5
April 4967,9 5122,3 5056,5 5379,1 5652,1 5880,4 5815,4 5986,6 6119,6
May 4683,5 4893,3 5036,4 5255,8 5524,2 5639,2 5641,6 5695,7 5744,8
June 3818,0 4150,2 4673,5 4775,0 5114,6 5174,0 5229,2 5237,1 5307,8
July 3895,1 3960,4 3987,2 4098,1 4300,7 4379,6 4405,0 4487,5 4546,4
August 3781,5 3856,8 2918,9 3592,1 3767,4 3991,7 3791,0 4281,7 4059,3
September 4478,3 4435,6 4532,9 4450,9 4743,8 4645,3 4672,9 4905,2 4802,7
October 4941,6 5027,5 4843,9 5166,5 5418,2 5562,7 5560,2 5735,2 5770,1
November 4905,3 4984,9 5041,3 5155,7 5316,9 5397,3 5426,4 5465,9 5527,8
December 4180,8 4324,0 4379,4 4649,4 4910,7 5070,8 5105,1 5212,8 5215,1

6. Solar Radiation Maps other cities outside Lomé, Atakpamé and Mango. The
geographical coordinates of the cities whose deposit is
As the results show, the learning was done in good estimated in Table 1.
conditions with an acceptable precision. From the weights A MLP has been trained with hourly solar radiation
determined after the validation of the learning, the solar data of sites in Lomé, Atakpamé and Mango [36]
radiation values will be calculated for the cities whose (Hontoria, L.). For those six cities, the National
monthly temperature, relative humidity data are available Meteorological Institute provides us 7 years data. As in
in addition to their latitude. Thus, the data obtained from the previous section the solar global monthly mean values
meteorology concerns only six other cities covering the are obtained for each point on the grid, so twelve solar
Togolese geographical space. Indeed, the National radiation maps (Figure 7 to Figure 18) were drawn (one
Meteorological Service only has complete data from six for each month).

Figure 7. Solar radiation map of Togo in January Figure 8. Solar radiation map of Togo in February
12 Sustainable Energy

Figure 9. Solar radiation map of Togo in March Figure 11. Solar radiation map of Togo in May

Figure 10. Solar radiation map of Togo in April Figure 12. Solar radiation map of Togo in June
 Sustainable Energy 13

Figure 13. Solar radiation map of Togo in July Figure 15. Solar radiation map of Togo in September

Figure 14. Solar radiation map of Togo in August Figure 16. Solar radiation map of Togo in October
14 Sustainable Energy

These results show that the solar radiation increases as

the latitude of the place increases. The months of June
July and August are very rainy in the central regio of Togo
and it is a mountainous zone. Also the relative humidity
is high during the same period compared to its value in
the south and north of the country. This results show
that solar radiation decreases in the central part of the
country during this period compared to southern cities like
Lomé, Tabligbo and the cities of Niamtougou, Mango and
Dapaong.

7. Conclusion
The measured data were used to model the solar
radiation for the various cities in Togo. Although the first
models applied to Togo used only calculate solar radiation
values at any time in the cities of measurement, the last
method allowed us to determine with good precision the
value of the solar radiation in several cities covering the
Togolese geographical space. These values were used to
establish the country's solar radiation maps for all months
of the year. These maps will allow the technicians and
engineers to have a precise idea of the solar radiation of
Togo for an optimization of its exploitation. Besides, the
classical methods are unable to generate solar radiation
series in places where no solar information is available.
Nevertheless, the methodology proposed is able to do this,
it is more versatile than the classical methods, and so is
able to draw maps of the zone. This methodology is easily
Figure 17. Solar radiation map of Togo in November extendible to other places. The only requirement is the
knowledge of the hourly solar radiation from only 3 sites
of the country where the map is going to be drawn.

References
[1] Angström, A., 1956. On the computation of global solar radiation
from records of sunshine. Arkiv Geophysik 3(23), 551-556.
[2] Rietveld, M.R., 1978. A new method for estimating the regression
coefficients in the formula relating solar radiation to sunshine.
Agriculture Meteorology 19, 243-252.
[3] Newland, F.J., 1989. A study of solar radiation models for the
coastal region of south China. Solar Energy 43(4), 227-235.
[4] Gopinathan, K., 1988. Computing the monthly mean daily diffuse
radiation from clearness index and percent possible sunshine.
Solar Energy 41 (4), 379-385.
[5] Akinoglu, B., Ecevit, A., 1990. Construction of a quadratic model
using modified Angström coefficients to estimate global solar
radiation. Solar Energy 45 (2), 85-92.
[6] Dorvlo, A., Ampratwum, B., 1999. Modeling of weather data for
Oman. Renewable Energy 17, 421-428.
[7] Philips, W.F., 1984. Harmonic analysis of climatic data. Solar
Energy 32 (3), 319-328.
[8] Herrero, A.C., 1993. Harmonic analysis of monthly solar radiation
data in Spain. Ambient Energy 14 (1), 35-40.
[9] Rumelhart, D.E., Hinton, G.E., Williams, J.R., 1986. Learning
Internal Representation by Error Propagation Parallel Distributed
Processing: Explorations in the Microstructure of Cognition, vol. 1.
MIT Press, MA, Cambridge (Chapter 8).
[10] Hertz, J., Krogh, A., Palmer, R., 1995. Introduction to the Theory
of Neural Computation. Addison-Wesley, Canada.
[11] Declan Conway, Emma Archer van Garderen, Delphine Deryng,
Steve Dorling, Tobias Krueger, Willem Landman, Bruce Lankford,
Karen Lebek, Tim Osborn, Claudia Ringler, James Thurlow,
Tingju Zhu, Carole Dalin, 2015. “Climate and southern Africa's
water-energy-food nexus”, in Nature Climate Change of Nature
Figure 18. Solar radiation map of Togo in December Publishing Group, 5 (9), 837.
 Sustainable Energy 15

[12] Al-Alawi, S.M., Al-Hinai, H.A., 1998. An ANN-based approach [24] Wasserman, P., 1993. Advanced Methods in Neural Computing.
for predicting global solar radiation in locations with no Van Nostrand Reinhold, New York.
measurements. Renewable Energy 14(1-4), 199-204. [25] Lippmann, R.P., 1987. An Introduction to Computing with Neural
[13] Togrul, I.T., Onat, E., 1999. A study for estimating the solar Nets. IEEE ASSP Magazine, pp. 4-22.
radiation in Elazıg using geographical and meteorological data. [26] Bishop, C.M., 1996. Neural Networks for Pattern Recognition.
Energy Conversion and Management 40, 1577-1584. Oxford University Press, Oxford.
[14] Kalogirou, S., Michanelides, S., Tymbios, F., 2002. Prediction of [27] Pham, D.T., Liu, X., 1995. Neural Networks for Identification,
maximum solar radiation using artificial neural networks. In: Predication and Control. Springer, London.
Proceedings of the WREC VII, Germany. [28] Haykin, S., 1994. Neural Networks. A Comprehensive Foundation.
[15] Sözen, A., Arcaklıog˘ lu, E., O¨ Zalp, M., 2004. Estimation of Macmillan Publishing Company.
solar potential in Turkey by artificial neural networks usinage [29] Hornik, K., Stinchcombe, M., White, H., 1989. Multilayer
meteorological and geographical data. Energy Conversion and feedforward networks are universal approximators. Neural
Management 45, 3033-3052. Networks 2 (5), 359-366.
[16] Yang, K., Koike, T., 2002. Estimating surface solar-radiation from [30] Hontoria, L., Riesco, J., Zufiria, P., Aguilera, J., 1999. Improved
upperair humidity. Solar Energy 72 (2), 177-186. generation of hourly solar radiation artificial series using neural
[17] Elizondo, D., Hoogenboom, G., McClendon, R., 1996. networks. EANN99.
Development of a neural network to predict daily solar radiation. [31] Hontoria, L., Riesco, J., Zufiria, P., Aguilera, J., 2000. Application
Agricultural and Forest Meteorology 71, 115-132. of neural networks in the solar radiation field. Obtainment of solar
[18] Mohandes, M., Rehman, S., Halawani, T.O., 1998. Estimation of radiation maps. 16th European Photovoltaic Solar Energy
global solar radiation using artificial neural networks. Renewable Conference and Exhibition. Glasgow, 1-5 May.
Energy 14 (1-4), 179-184. [32] Zufiria, P., Vazquez, A., Riesco, J., Aguilera, J., Hontoria, L.,
[19] Mohandes, M., Balghonaim, A., Kassas, M., Rehman, S., 1999. A neural network approach for generating solar radiation
Halawani, T.O., 2000. Use of radial basis functions for estimating artificial series. In: Proceedings of the (IWANN099). 4-5 June.
monthly mean daily solar radiation. Solar Energy 68 (2), 161-168. Alicante, Spain.
[20] Hontoria, L., Aguilera, J., Riesco, J., Zufıria, P., 2001. Recurrent [33] Chatelin, C., Les Multi Layer Perceptron (MLP) Nov 2003.
neural supervised models for generating solar radiation synthetic [34] Rude, J., Développement d’un modèle statistique neuronal pour la
series. Journal of Intelligent and Robotic Systems 31, 201-221. description fine de la pollution par le dioxyde d’azote : application
[21] Hontoria, L., Aguilera, J., Zufı ´ria, P., 2002. Generation of hourly à la région parisienne, Université Paris XII - Val de Marne, Thèse
irradiation synthetic series using the neural network multilayer de doctorat, 2008.
perceptron. Solar Energy 72, 441-446. [35] Ammar, Y. M., mise en œuvre de réseaux de neurones pour la
[22] Tasadduq, I., Rehman, S., Bubshait, K., 2002. Application of modélisation de cinétiques réactionnelles en vue de la
neural networks for the prediction of hourly mean surface transposition batch/continu, Thèse de Doctorat de l’I.N.P. –
temperature in Saudi Arabia. Renewable Energy 25, 545-554. Toulouse – 17 juillet 2007.
[23] Tymvios, F.S., Jacovides, C.P., Michaelides, S.C., Scouteli, C., [36] Hontoria, L., Aguilera, J., Zufiria, P., 2005. An application of the
2005. Comparative study of Angström’s and artificial neural multilayer perceptron: Solar radiation maps in Spain. Solar Energy
networks’ methodologies in estimating global solar radiation. 79, 523-530.
Solar Energy 78, 752-762.

View publication stats