Available online at www.sciencedirect.

com

ScienceDirect
Energy Procedia 42 (2013) 33 – 42

Mediterranean Green Energy Forum 2013

Estimation of tilted solar irradiation using Artificial

Neural Networks
Gilles Nottona*, Christophe Paolia, Said Diafb
a
University of Corsica Pasquale Paoli, UMR CNRS 6134, Centre Georges Peri, Route des Sanguinaires, 20000 Ajaccio, France
b
Centre de Développement des Energies Renouvelables, BP 62, 16340 Bouzareah, Algiers, Algeria

Abstract

Calculation of solar global irradiation on tilted planes from only horizontal global one is particularly difficult when
the time step is small. We used an Artificial Neural Network (ANN) to realize this conversion at an hourly and 10-
min time step. The ANN is developed and optimized using five years of solar data. The accuracy is respectively for
hourly and 10-min data of 6% and 9% for the RMSE and 3.5% and 5.5% for the RMAE i.e. similar or slightly lower
than the errors obtained with conventional empirical correlations for hourly data.

©©2013
2010ThePublished
Authors. Published
by Elsevier Ltd.Ltd. Open access under CC BY-NC-ND license.
by Elsevier
Selection and peer-review under responsibility of KES International
Keywords: Solar radiation; Articial neural network; Model; Tilted planes; Global solar radiation.

1. Introduction

If monthly average values allow to realize a preliminary sizing, daily values and even better hourly
values are required if we want to perform a more correct and precise sizing particularly when an energy
storage mean or an electrical power buffer is in the system. In commercial software for solar systems, the
inclination of collectors is requested, the software “inclines” the horizontal global irradiation contained in
the database with generally a low accuracy. It is useless to develop precise physical models if the input
data are too approximated.
As said by Behr [1], three main reasons make it impossible to develop a simple model for converting
horizontal global solar radiation into tilted radiation:

* Corresponding author. Tel.: +33.495524152; fax: +33.495524141

E-mail address: gilles.notton]univ-corse.fr.

1876-6102 © 2013 The Authors. Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
Selection and peer-review under responsibility of KES International
doi:10.1016/j.egypro.2013.11.003
34 Gilles Notton et al. / Energy Procedia 42 (2013) 33 – 42

x the radiations incident on tilted planes include the radiation reflected by the environment;
x when the plane is tilted, only a portion of the sky is “seen” and the sky diffuse component depends on
the inclination and orientation of the collector, on the sun elevation and azimuth but also on the sky
conditions rarely uniform. This induces complex anisotropic effects. There are two main problems for
this sky anisotropy: the circumsolar brightness due to solar radiation diffusion by aerosols (around the
sun) and the horizon brightness near the horizon.
x at last, these data are relatively rare.

2. Solar radiation components

When solar radiation enters the earth’s atmosphere, the incident energy is removed by scattering and
absorption. The scattered radiation is the diffuse radiation. A part goes back to the space and another one
reaches the ground directly in line from the sun, the beam radiation. The solar radiation on a tilted
collector has a beam component, two diffuse ones from the sky and from the ground (Fig. 1).

solar radiation reflected

back to space

Thin
clouds
beam sky diffuse
beam cloud

aerosols Anisotropic diffuse

radiation on horizontal beam normal
surface
E

Air molecules
ground -reflected diffuse

Fig. 1. Solar radiation on an horizontal and tilted surface.

3. The experimental data and validation tests

Measures are realized in our laboratory in the gulf of Ajaccio (41°55’ N; long: 8°55’ E) at about 200
m from the sea at an altitude of 70 m (Fig. 2) with an insular Mediterranean climate. The data are
measured each minute. For the normal beam irradiance, a pyrheliometer Kipp & Zonen at normal
incidence is mounted on an automatic Solar Tracker SMT-3. Every two days, the pyrheliometer is cleaned
and its alignment verified. The global irradiance on 0°, 45° and 60° tilted surfaces is measured by three
Kipp & Zonen (CM11) pyranometers. The experimental data were collected since January 2006. We have
5 years of minute data of horizontal global irradiance, normal beam irradiance and 45° and 60° inclined
global irradiance. Four calculated parameters are added: horizontal extraterrestrial and diffuse irradiance,
solar declination and the zenith angle [2].

Fig. 2. Position of the Meteorological site and the solar measuring station
 Gilles Notton et al. / Energy Procedia 42 (2013) 33 – 42 35

Some quality control tests are imposed on these minute data in order to extract outlier or missing data
[3-4]. Then the 10-min and hourly values are calculated (we allowed 10% of missing data over the given
period). A second quality control [5] is applied to the 10-min and hourly data, and all data that do not
comply with the following conditions are not used. Very often, the sunset and the sunrise induce some
problems in the correlations, because of the mask effect of the environment and due to the bad response
of pyranometers when the zenith angle is high (cosine effect) [5]. Thus, we extracted these sunset and
sunrise periods. After these verifications, over the 5 years, we have 18 615 validated hourly data (1.5% of
missing data) and 101140 validated 10-min data (6.46% of missing data). Thus, we have monthly files for
10-min and hourly data containing : day, hour, declination, zenith angle, horizontal global irradiation,
normal beam irradiation, extraterrestrial irradiation, global irradiations on a 45° and 60° tilted plane.

4. Artificial Neural Networks (ANN)

An ANN is an information processing paradigm that is inspired by the way biological nervous
systems, such as the brain process information. It is composed of a large number of highly interconnected
processing elements (neurons) working in unison to solve specific problems. The goal of this network is
to create a model that correctly maps the input to the output using historical data so that the model can
then be used to produce the output when the desired output is unknown. An ANN has a parallel-
distributed structure and consists in a set of processing elements called neurons. The ANN structure
consists in:
x an input layer which receives data;
x an output layer to send computed information;
x one or several hidden layers lying the input and output layers.
According to the architecture, all or a part of the neurons in a layer are connected with all or a part of
the neurons of the previous and next layer. The number of hidden layers and of neurons of each layer
depends on the specific model, convergence speed, generalization capability, physical process and the
training data that the network will simulate [6]. A typical artificial neuron and the modeling of a multi-
layered neural network are shown in Fig. 3. The signal flow from inputs x k,1, xk,2,…., xk,p of a layer k is
considered here as unidirectional (feed forward). The output signal O to the j th neuron of the following
layer (k+1) is given by [7]:

§ p ·

O f net  f w T x k  f ¨ ¦ w k , jx k , j ¸
¨j 1 ¸
(1)
© ¹
where wk,j is the synaptic weights (k is the layer index, j is the neuron index) and f net  is referred to
as an activation or transfer function. The variable net is the scalar product of the weight and input vectors.
T is the transpose of a matrix. The result of this sum, net, is then transformed by a transfer function f
which produces the output O of a neuron if this sum exceeds a certain threshold. The output is then
distributed to other neurons as inputs.
There are two main problems concerning the ANN implementation: specifying the network size
(number of layers in the network and number of nodes in each layer) and finding the optimal values for
the connection weights. An insufficient number of hidden nodes may be the cause of difficulties in
learning data whereas an excessive number of hidden nodes might lead to unnecessary training time with
marginal improvement as well as make the estimation for a suitable set of interconnection weights more
difficult [8]. To determine the optimal number of hidden nodes; the common method used is trial and
error based on a total error criterion. This method starts with a small number of nodes, gradually
increasing the network size until the desired accuracy is achieved.
36 Gilles Notton et al. / Energy Procedia 42 (2013) 33 – 42

Synaptic
weights xk,1 wk,1

xk,2 wk,2 Transfer Activation

I O function function
N U net
P T xk,3 wk,3 6 f xk+1,j output
U P
. .
T U
. .
T
. . Tk
. . Threshold

xk,p wk,p

Neurons Input Synaptic

Hidden layers signal weights

Fig. 3. Architecture of an artificial neuron and a multilayered neural network.

One of the properties of ANNs is their ability to learn from their environment and to improve their
performance through a learning process called also training process. Learning results in a change in the
weights value wk,j connecting the neurons from one layer to another. The goal is to achieve equality
between the actual output and simulated output. It is therefore necessary first to choose the learning
algorithm and to define the part of the data used for learning in relation to the total amount of data
available. The various steps in the implementation of an optimized ANN consist in selecting:
- ANN structure;
- transfer function type;
- ANN size (number of layers and number of neurons per layer);
- learning algorithm;
- training/test set;
- input data

5. ANN Implementation

From a bibliography review [9], we chose for the ANN implementation :

1. Structure: a MultiLayer Perceptron (MLP) using feed-forward back-propagation;
2. Transfer function : the combination of a sigmoid and a linear transfer functions allows a good
approximation of several types of functions.
3. Learning algorithm: the Levenberg–Marquardt algorithm (LM) is an approximation to the Newton
method used during the training;
4. Training/test set : among the available data, we must use a percentage for the training and the rest for
the testing. As proposed by Dreyfus [10] we must use more than 50% of data for the training phase.
Thus, we decided to take 3 whole years of data for training (60%) and 2 whole years (40%) for
testing;
5. ANN size: whatever the number of inputs, the ANN has to be as simple as possible. These
configurations give the best results with a small error. The compromise between the number of hidden
layers and the number of neurons in each layer allows to obtain a fast and robust network giving the
best results. A large sized architecture requires more data and easily becomes over-trained, i.e unable
to generalize efficiently. There is no mathematical method allowing the optimal sizing of the ANN;
the choice of the optimal structure is done only after testing different configurations and estimating
their performance.
 Gilles Notton et al. / Energy Procedia 42 (2013) 33 – 42 37

6. Input data: the declination represents the position of the Earth from the Sun and depends on the day
number. The sun position influences the quantity and the quality of the sun radiation: when the sun is
high in the sky (low zenith angle), the solar radiation is maximal (in clear skies) and as the optical
path is minimal, therefore the incident radiation is less absorbed. This position is characterized by the
zenith angle. Depending on the season and sky conditions, several values of inclined solar irradiation
correspond to the same horizontal irradiation. In diffuse radiation models, the clearness or the diffuse
index are used to take account of sky condition which is an important parameter in the repartition of
diffuse and beam radiations. The higher the clearness index is, the clearer the sky is and the more is
global radiation composed of beam radiation. The extraterrestrial irradiation is used as reference and
is useful as an input parameter. Thus, the input parameters will be: i) the declination, ii) the zenith
angle, iii) the horizontal global and iv) the extraterrestrial horizontal solar irradiation. We developed
three ANNs for hourly and 10-min time step (Fig. 4):
x the first one for the estimation of global horizontal on a 45° tilted plane;
x the second one for the estimation of global horizontal on a 60° tilted plane;
x the third one for the estimation of global horizontal on a 45° or 60° tilted plane;
In case 3, the number of input data is doubled because each value of hourly global horizontal
irradiation is entered twice, with E=45° and with E=60°.
We use a two hidden layers model and we optimize it in varying the number of neurons in each hidden
layer. We varied the number of neurons in the first hidden layer from 15 to 50 and in the second one from
5 to 50; we limited the number of hidden layers to two because the ANN must be as simple as possible.
Input layer Hidden layers Output layer
Declination G
Model 1

Zenith angle Tz
Global irradiation on
a 45° tilted plane I45°
Extraterrestrial
horizontal irradiation I0
Horizontal global
irradiation I
Input layer Hidden layers Output layer
Declination G
Model 2
Zenith angle Tz
Global irradiation on
a 60° tilted plane I60°
Extraterrestrial
horizontal irradiation I0
Horizontal global
irradiation I
Input layer Hidden layers Output layer
Declination G
Model 3
Zenith angle Tz
Global irradiation
on a E° tilted plane
Extraterrestrial IE
horizontal irradiation I0
Horizontal global
irradiation I
Inclination angle E
(E=45° 60°)
Input measured Output researched data Input calculated
data -Starting data Incoming data data

Fig. 4. The three ANN models.

38 Gilles Notton et al. / Energy Procedia 42 (2013) 33 – 42

We tested various configurations of monolayer and double layer ANN for the three models presented
above and the best configurations are shown in Table 1. The decision criteria are applied to the whole
data (training set and test set) using the RRMSE, the RMAE and the CC values defined by equation (2).
The models being nonlinear, the training results depend on the initial values of the model inputs. Thus,
each configuration has been tested 6 times.

N ª º
RMAE ¦ yi  x i Nx RRMSE « yi  xi 2 N» x
i 1 ¬ ¼
12
ªN º ­ª N 2 ºª N 2 º½
CC « ¦ yi  y x i  x » ®« ¦ yi  y  » « ¦ x i  x  » ¾ (2)
¬i 1 ¼ ¯¬i 1 ¼ ¬i 1 ¼¿
yiN and x i are the
N estimated and measured values and the respective average values are given by
x ¦ x i and y ¦ yi where N is the data number.
i 1 i 1

Table 1. Best configurations for the ANN and adequacy parameters

Model Inclination RRMSE RMAE

Neurons per layer CC
number E % %
Hourly data
1 45° 25 15 5.28 2.79 0.998
2 60° 30 20 6.24 3.42 0.998
3 45°/60° 25 20 6.62 3.93 0.997
10-min data
1 45° 12 0 8.10 5.02 0.996
2 60° 24 0 10.11 5.75 0.994
3 45°/60° 12 0 9.46 6.01 0.995

Several remarks can be made:

x The optimal structures are not the same for hourly and 10-min data;
x For the 10-min data, the best results are obtained for a monolayer ANN composed of 12 neurons in its
hidden layer for model 1 and 3 and with 24 neurons for model 2 with an RRMSE between 8.10% and
10.11% and a RMAE between 5.02% and 6.01%;
x For hourly data, the best results are obtained for a double-layer ANN with an RRMSE between 5.28%
and 6.62% and a RMAE between 2.79% and 3.93%;
x The errors are higher for 60° than 45° and in the literature, this observation is also realized for
empirical model and is due to the greater influence of the sky anisotropy on the diffuse radiation when
the inclination angle is high;
x The model 3 proposing to generalize the first two models, gives the worst results.
To illustrate the good reliability of the developed models, we show in Figs. 5 to 6 the estimated data
versus the experimental data for the model 1 for hourly data and the model 2 for 10-min data. We
randomly selected periods of days for which we plotted the experimental and calculated data (Figs. 7-8).
 Gilles Notton et al. / Energy Procedia 42 (2013) 33 – 42 39

1000

800

Estimated data [Wh/m²]

600

400

200

0
0 200 400 600 800 1000
Measured data [Wh/m²]

Fig. 5. Validation of model 1: estimation of hourly global irradiation on a 45° tilted plane.

200

180

160

140
Estimated data (Wh/m²)

120

100

80

60

40

20

0 20 40 60 80 100 120 140 160 180 200

Experimental data (Wh/m²)

Fig. 6. Validation of model 2: estimation of 10-min global irradiation on a 60° tilted plane.

Fig. 7. Validation of the model 1 over few days for hourly data.
40 Gilles Notton et al. / Energy Procedia 42 (2013) 33 – 42

180
Experimental data
160 Modelled data

10-min solar irradiation on 60° plane (Wh/m²)

140

120

100

80

60

40

20

0
5.83 9.83 13.83 17.83 9.17 13.17 18.50 9.67 13.67 18.50 10.00 14.00 6.50 10.50 14.50
Time (hour)

Fig. 8. Validation of the model 2 over few days for 10-min data.

6. Comparison with models of the literature

There are a number of models available to estimate global radiation on inclined surface from
horizontal radiation. But, these models require information at the same time on the global and the beam or
diffuse radiation on a horizontal surface. Olmo et al [11] developed a model only requiring the global
radiation, the sun’azimuth and elevation. Another method consists in coupling two types of models as
illustrated on Fig. 9: model for the estimation of horizontal diffuse solar radiation from the horizontal
global one and model for computing the global solar radiation on tilted planes from horizontal global and
diffuse radiations.
Normal beam Horizontal global Global Global
radiation radiation radiation 45° radiation 60°

Horizontal
diffuse radiation
First step Global
radiation
45°
Second step Global
Horizontal
radiation
diffuse
60°
radiation
Experimental
verification
Measured Calculated Estimated by correlation

Fig. 9. Conventional method in two steps.

94 combinations were tested in Ajaccio by Notton et al [12] using 7 horizontal diffuse solar irradiation
models [13] and 15 tilted diffuse solar irradiation models [14]. The RRMSE obtained with these
combinations are around 10% and the best combination conduces to a RRMSE of 8.11% for 45° and
10.71% for 60° [12]. With the Olmo method [11], RRMSE was only 12.14% for 45° and 17.01% for 60°.
A comparison between our ANN methods with conventional one is shown in Fig. 10.
 Gilles Notton et al. / Energy Procedia 42 (2013) 33 – 42 41

18%
17.00%

16%
RMAE 10-MIN DATA
HOURLY DATA
RRMSE
14%

12%
10.71%
10.00%
10%

8.11%
8%

5.75% 6.01%
6%
5.02%
4.36%
3.93% 3.45%
4% 3.42% 3.08%
2.79% 2.82% 2.69%
2.49%
2%

0%
Olmo et al [11] Notton et al [12] Notton et al [12] This study This study This study This study This study This study
(hourly data) (hourly data) (hourly data) (hourly data) (hourly data) (hourly data) (10-min data) (10-min data) (10-min data)
(45°) (60°) (45°) (60°) ( 45°+60°) (45°) (60°) ( 45°+60°)

Fig. 10. Comparison of the adequacy of conventional methods with ANN methods.

7. Conclusion

We developed three models of solar global irradiation on tilted plane from horizontal ones using
Artificial Neural Networks successively for hourly and 10-min data. This study was performed by using
five years of 10-min solar and hourly radiation data collected in the Mediterranean site of Ajaccio,
France. The choice of the ANN type and transfer functions were realized from a bibliographical study.
Using a sensitivity analysis, we optimized the ANN architecture. 4 input parameters were used in these
models (5 for the third model).
The adequacy of the ANN models is very satisfying, better than that obtained with empirical models
for hourly time-step. For 10-min data estimation, the ANN methodology shows a good adequacy. This
ANN approach out-performs the traditional methods for this type of estimation.

References

[1] Behr, HD. Solar radiation on tilted south-oriented surfaces: Validation of transfer-models. Sol Energy 1997;61(6):399–413.
[2] Iqbal M. An introduction to solar radiation. Canada: Academic Press; 1983.
[3] Li DHW. Lam JC. Evaluation of slope irradiance and illuminance models against measured Hong Kong data. Building Environ.
2000;35:501–509.
[4] Lam JC. Wan KKW. Yang L. Solar radiation modelling using ANNs for different climates in China. Energy Conversion and
Management 2008;49:1080–1090.
[5] De Miguel A. Bilbao J. Aguiar R. Kambezidis H. Negro E. Diffuse solar irradiation model evaluation in the North
Mediterranean belt area. Solar Energy 2001;70(2):143–53.
[6] Krishnaiah, T, Srinivasa Rao, S, Madhumurthy, K, Reddy, RS. Neural Network Approach for Modelling Global Solar Radiation.
J of Applied Sciences Research 2007;3(10):1105-1111.
[7] Abraham, A. Artificial Neural Networks. Handbook for Measurement Systems Design, Peter Sydenham and Richard Thorn
(Eds.), John Wiley and Sons Ltd., London, 901-908, 2005.
[8] Alam, S, Kaushik SC, Garg, SN. Assessment of diffuse solar energy under general sky condition using artificial neural network.
Applied Energy 2009;86:554–564.
42 Gilles Notton et al. / Energy Procedia 42 (2013) 33 – 42

[9] Notton, G, Paoli, C, Vasileva, S, Nivet,ML, Canaletti, JL, Cristofari, C. Estimation of hourly global solar irradiation on tilted
planes from horizontal one using artificial neural network. Energy 2012. 39, 166-179.
[10]Dreyfus, G, Martinez, J, Samuelides, M, Gordon, M, Badran, F, Thiria, S. Apprentissage statistique: Réseaux de neurones -
Cartes topologiques - Machines à vecteurs supports. 3 éd. Eyrolles; 2008.
[11] Olmo, FJ. J.Vida, I. Foyo, Y. Castro-Diez, L. Alabos-Arboledas. Prediction of global Irradiance on inclined surfaces from
horizontal global irradiance. Energy 1999;24:689–704.
[12] Notton, G, Poggi P, Cristofari, C. Predicting hourly solar irradiations on inclined surfaces based on the horizontal
measurements: Performances of the association of well-known mathematical models. Energy Convers Manage 2006;47:1816–
29.
[13]. Notton, G, Cristofari, C, Muselli M, Poggi, P. Calculation on an hourly basis of solar diffuse irradiations from global data for
horizontal surfaces in Ajaccio. Energy Convers Manage 2004; 45:2849-2866.
[14] Notton, G, Cristofari C, Muselli ,M, Poggi P. Performance evaluation of various hourly slope irradiation models using
Mediterranean experimental data of Ajaccio. Energy Convers Manage 2004, 47-2, 147–173.