Available online at www.sciencedirect.

com

ScienceDirect
Energy Procedia 69 (2015) 1999 – 2008

International Conference on Concentrating Solar Power and Chemical Energy Systems,

SolarPACES 2014

Clear-sky irradiance model for real-time sky imager application

J. Nou1, R. Chauvin1,2, S. Thil1,2, J. Eynard1,2 and S. Grieu1,2,*
1
PROMES-CNRS, Rambla de la thermodynamique, Tecnosud, 66100 Perpignan, France
2
University of Perpignan Via Domitia, 52 Avenue Paul Alduy, 66860 Perpignan, France

Abstract

In a context of sustainable development, enthusiasm for concentrated solar power technologies is developing. So, the CSPIMP
(Concentrated Solar Power efficiency IMProvement) project started in 2013 in order to achieve a better competitiveness for the
CSP plants. Its main target is to develop new procedures to improve steam turbine start up cycles, maintenance activities and
plant control. One challenge in the project is to better forecast the solar resource in order to optimize plants operation. The
development of a clear sky model is therefore an essential step to forecast Direct Normal Irradiance (DNI) because clear sky
represents the nominal operating conditions of a concentrating solar power plant. This paper focuses on estimating DNI under
clear-sky conditions using a knowledge model based on the coefficient of relative optical air mass, DNI and atmospheric
turbidity fluctuations. The satisfactory results obtained validate the proposed methodology.
© 2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
© 2015 The Authors. Published by Elsevier Ltd.
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer review by the scientific conference committee of SolarPACES 2014 under responsibility of PSE AG.
Peer review by the scientific conference committee of SolarPACES 2014 under responsibility of PSE AG

Keywords: Clear sky irrandiance model; sky imager; real-time application; concentrating solar power systems.

1. Introduction

It is widely acknowledged by solar companies and plant operators that cost remains the main drawback of
Concentrating Solar Power (CSP) systems. In that context, the CSPIMP (Concentrated Solar Power efficiency
IMProvement) project has been initiated in 2013 in order to make CSP plants more competitive. The main target of
the project is to improve plant efficiency by developing better procedures for steam turbine start up cycles and
maintenance activities as well as proposing advanced plant control schemes.

* Corresponding author. Tel.: +33468682257; fax: +33468682213.

E-mail address: stephane.grieu@promes.cnrs.fr

1876-6102 © 2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer review by the scientific conference committee of SolarPACES 2014 under responsibility of PSE AG
doi:10.1016/j.egypro.2015.03.208
2000 J. Nou et al. / Energy Procedia 69 (2015) 1999 – 2008

Depending on the CSP configuration (turbine, back-up generator, buffer tank, storage system…), a customized
forecast model must be built to deal with the plant's specific behaviour against solar variability. Different forecast
models exist to assess the solar resource at short-term (i.e. a few minutes), medium-term (i.e. a few hours) and long-
term (i.e. a few days) horizons.
For instance, Numerical Weather Prediction (NWP) models are well adapted to the forecast of changes in
meteorological parameters at long-term horizon [1]. These models are able to provide forecasts of parameters such
as temperature, relative humidity and wind speed but also global and direct irradiance, which is relevant information
for CSP plant strategy improvement and integration into the electricity grid. However, NWP models do not take into
account the cloud cover. So, to refine their spatial and temporal resolutions, such models can be coupled with
satellite imagery. Indeed, satellite imagery is able to provide more comprehensive information about cloud motion
[2]. Images are provided every 15 minutes by geostationary satellites, as Meteosat in Europe or GEOS in the USA,
with a spatial resolution of ~1 km². From these images, it is possible to obtain a daily trend of solar irradiance, what
can be relevant in order to optimize the use of a storage system. Moreover, NWP models and satellite imagery can
be combined with time series forecasting models, based on a historical climate database, with the aim of identifying
cycles and/or trends in the solar resource [3]. However, NWP models, even coupled with satellite imagery, are not
suited to the real-time management of a solar power plant.
Images can also be provided by ground-based cameras, named sky imagers. Such cameras allow solar radiation
to be forecasted at high spatial (~100 m²) and temporal (< 30 min) resolutions. Indeed, sky imagers can provide in
real-time a very short term forecast of the beam attenuation produced by clouds, allowing the power plant owner to
make more accurate bids in the market with less risk of under-generation (leading to penalties) or over-generation
(leading to losses due to curtailment). Sky imagers can also contribute to an increase in equipment lifetime and
reliability using forecasting algorithms developed to take advantage of solar variability. It is especially interesting
for storage systems for which the overall return on investment is directly related to the system's lifetime. A smart
charging and release algorithm can take advantage of the future availability and variability of the solar resource to
better coordinate the backup generator with a buffer tank. It could help to reduce the number of charging/release
cycles and would make storage systems more economically feasible. Finally, sky-imaging systems can also be used
to adjust the Heat Transfer Fluid (HTF) flow in real time. At the moment, CSP plants are working with strong
limitations on the HTF flow in order to avoid bending absorber pipes during fast fluctuations of the solar resource.
Having the HTF flow as a function of the DNI and using a sky-imaging system would allow the plant to operate
with lower limitations on this flow. Consequently, HTF temperature would be more often at its nominal point,
allowing the plant to be more effective.
Among the different approaches and devices available to perform the forecasting of solar resource, we focused on
ground-based cameras which are well-suited to estimate and forecast the clearness index at a very short-term horizon
and a high spatial resolution [4–6]. Usually, this forecasting approach consists in two complementary steps. The first
one is about determining the clearness index (its value ranges between 0 and 1) corresponding to the sun occultation
caused by clouds. The second step is related to the clear sky model used to estimate the theoretical solar irradiance
received at ground level without clouds. Developing a clear sky model is therefore an essential step to forecast direct
normal irradiance. Nowadays, clear sky models developed for satellite imagery [2,7] or ground-based camera
applications [4–6] are often inappropriate to real-time applications. Indeed, at the moment, there are no works
dealing with the clear sky detection in real-time. As an example, in a previous work [8], we proposed a methodology
based on a multi-resolution analysis (using the discrete wavelet transform) in order to detect clear sky days in our
database. This approach is useful for data preprocessing and model development but not adapted to real-time
applications like solar power plants management. Moreover, the existing models dealing with clear sky irradiance
are relatively basic. For instance, an approach is proposed to obtain short-term forecasts of DNI at the ground level
by authors assuming that DNI is constant (900 W.m-2) during the considered period (from 10:00 PM to 14 PM) [6].
More often, clear sky irradiance models use yearly or monthly atmospheric turbidity values or are based on
polynomials. In their work [4], Quesada-Ruiz et al. computed clear-sky DNI from an eighth-order polynomial of the
cosine of the solar zenith angle using a Least Squares Method (LSM). However, the results they obtained show that,
for clear sky conditions, the model is less effective than a basic persistent model. In another study, clear sky
irradiance is computed from the Linke turbidity coefficient. This coefficient has the main drawback to be dependent
on air mass [5]. In order to overcome these different problems, a knowledge model based on the coefficient of
relative optical air mass, DNI and atmospheric turbidity fluctuations is proposed in order to estimate the clear sky
 J. Nou et al. / Energy Procedia 69 (2015) 1999 – 2008 2001

DNI in real time. The paper is organized as follows: section 2 is about the PROMES-CNRS experimental platform
as well as the measured and computed variables used in the clear sky model development and validation. Section 3
focuses first on atmospheric turbidity. In addition, the model development and optimization processes are described.
Section 4 deals with the estimation results we obtained and an overall analysis. The paper ends with a conclusion
and outlook on further work.
Nomenclature
ܾ Multiplicative coefficient based on the altitude of a considered site
݅ Position of an element in a given vector
݇௥௔௡ௗ Sequence of pseudo random values drawn from the standard distribution on [0,1]
݇௧ Clearness index
݉ Altitude-corrected air mass
‫ݎ‬ Sun-Earth distance correction factor
‫ݐ‬௣௔௦௧ Last time atmospheric turbidity is trustworthy
‫ݑ‬ Binary sequence resulting from the visual highlight of clear sky instants
‫ݒ‬ Pseudo Random Binary Sequence (PRBS)
‫ݒ‬Ԣ Sequence of pseudo random values drawn from the standard distribution on [0,1]
‫ݓ‬ Sequence obtained by multiplying the sequences ‫ ݑ‬and ‫ ݒ‬term by term
‫ܥ‬ଵ Model constant 1
‫ܥ‬ଶ Model constant 2
‫்ܥ‬ Coefficient of atmospheric turbidity
‫ܫ‬ Direct normal irradiance (W.m-2)
‫ܫ‬଴ Extra-terrestrial solar irradiance (W.m-2)
‫ܫ‬஼ௌ Broadband direct normal irradiance under clear sky conditions (W.m-2)
‫ீܫ‬ Generated solar irradiance sequence (W.m-2)
ܰ Total number of elements in a given vector
ܴ଴/ଵ Degradation ratio dealing with the proportion of "1" in sequence ‫ݒ‬
ܶ Atmospheric transmittance resulting from both scattering and absorption of the sunlight
ܶ௅ூ Atmospheric turbidity (Ineichen and Perez [9])
ο‫ݐ‬ Elapsed time since the last trustworthy atmospheric turbidity measurement (s)
ο‫ݐ‬௠௔௫ Maximum value of ο‫( ݐ‬s)
ο‫ݐ‬௠௜௡ Minimum value of ο‫( ݐ‬s)
οܶ Measured change in atmospheric turbidity
οܶ௟௜௠ Threshold of atmospheric turbidity change
οܶ௠௔௫ Maximum accepted change in atmospheric turbidity
οܶ௠௜௡ Minimum accepted change in atmospheric turbidity

2. Experimental set-up and database

(a) Experimental platform (b) PROMES sky imager (c) Irradiance sensor (RSI)

Fig. 1. View of the experimental platform.

2002 J. Nou et al. / Energy Procedia 69 (2015) 1999 – 2008

To develop and validate our model, an experimental platform (Fig. 1.a) has been installed at the PROMES-CNRS
laboratory in 2013 (Perpignan, France, latitude is 42.66° N, longitude is 2.91° E and elevation is ~50 m). This part
of the paper provides a hardware overview of the platform and specifies the working conditions of the different
instruments used.

2.1. PROMES Sky Imager

To determine the clear sky index (its value ranges between 0 and 1) corresponding to the sun occultation caused
by clouds, the PROMES-CNRS laboratory has decided to build its own system, fully customizable from both a
hardware and software point of view. This decision has been motivated by the fact that all the existing sky imagers
suffer from drawbacks and PROMES-CNRS believes that a customizable solution would make such systems more
attractive. After a detailed review of the different cameras proposed by manufacturers, a 5-megapixel camera with a
color CMOS sensor has finally been selected. The camera, named 5481VSE-C and provided by IDS, is equipped
with a Fujinon fisheye lens and protected by a waterproof enclosure manufactured by autoVimation (Fig. 1.b).
Images are collected every 20 seconds at a resolution of 1920 x 2560 pixels and with a bit depth of 8 bits per
channel (i.e. 3x8 bits).

2.2. Irradiance sensor

A Rotating Shadowband Irradiometer (RSI) (Fig. 1.c) has been installed next to the sky imager in order to
measure every minute, with a good precision, the solar irradiance components. Equipped with a silicon photodiode
radiometer and a shadow band, this instrument allows Direct Normal Irradiance (DNI), Global Horizontal Irradiance
(GHI) and Diffuse Horizontal Irradiance (DHI) to be measured or deduced from measurements. The data are
collected by the sky imager software and immediately included in the process analysis. In the present study, we only
focus on DNI measurement because it is a key point in the clear sky model development and because in the CSP
technology, the power production is only affected by the direct component.

2.3. Database

The database used to develop and validate the clear sky model is composed of data collected by the experimental
platform and data derived from the sun-earth equations given by the SG2 algorithm [10]. Knowing the Julian date
and the localization of the considered site (latitude, longitude, and altitude), this algorithm is able to give quickly
and with high accuracy all information like air mass, azimuth/zenith angles, and extraterrestrial solar radiation. In
order to avoid missing data, we considered measures from February the 5th to March the 21st, 2014.

3. Clear-sky irradiance model

Clear sky is the nominal operating conditions of a solar power plant. Consequently, the clear sky irradiance model
is a key component in a forecasting approach. This section of the paper deals with the model developed by the
PROMES-CNRS laboratory for our real-time sky imager application. So, a knowledge model based on the
coefficient of relative optical air mass, DNI and atmospheric turbidity fluctuations is proposed in order to estimate
the clear sky DNI in real time (‫ܫ‬஼ௌ ). Before addressing the development phase, let's talk about atmospheric turbidity.

3.1. Atmospheric turbidity

Under clear sky conditions, the broadband DNI (‫ܫ‬஼ௌ ) is given by (1):

‫ܫ‬஼ௌ = ‫ ݎ‬ή ‫ܫ‬଴ ή ܶ (1)

with ‫ܫ‬଴ the extraterrestrial solar irradiance (‫ܫ‬଴ = 1361.2 W.m-2 [11]), ‫ ݎ‬the sun-earth distance correction factor and ܶ
the atmospheric transmittance resulting from both scattering and absorption of the sunlight. This atmospheric
 J. Nou et al. / Energy Procedia 69 (2015) 1999 – 2008 2003

transmittance can be obtained using radiation transfer models [12] or a broadband turbidity coefficient like the well-
known Linke turbidity coefficient [13]. Although radiation transfer models have proven to be able to forecast clear
sky DNI with unsurpassed accuracy, they require Aerosol Optical Depth (AOD) data which are difficult to measure
or rarely available with high accuracy. On the other hand, models based on a broadband turbidity coefficient give a
lower accuracy than radiation transfer models but can be easily implemented because they only derive from
broadband beam radiation measurement networks. In 2002, Ineichen and Perez [9] proposed an empirical
formulation of the turbidity coefficient for the normal beam clear sky radiation including both scattering and
absorption phenomena. This new formulation is less dependent on air mass than the Linke turbidity coefficient.
Because this coefficient (T LI ) is well adapted for on-site and real-time applications, it has been selected as a starting
point in the development of our clear sky model. It is calculated using equation (2):

11.1 ܾ‫ܫ‬଴
ܶ௅ூ = 1 + ൤ ή ln ൬ ൰൨ (2)
݉ ‫ܫ‬஼ௌ

with ݉ the optical air mass and ܾ a coefficient taking into account the altitude of the considered site.

3.2. Real-time estimation of the atmospheric turbidity

During the real-time DNI acquisition process, each new DNI value ‫ ܫ‬must be classified as clear sky irradiance or
not. It is a key information because it allows the clearness index ݇௧ to be calculated. It is defined as follows (3):
‫ܫ‬
݇௧ = (3)
‫ܫ‬஼ௌ
with ‫ ܫ‬the measured irradiance and ‫ܫ‬஼ௌ the estimated clear sky irradiance. This index is a powerful indicator of the
beam attenuation produced by clouds and is frequently used in solar resource forecasting [4–6]. To classify the last
measured DNI value correctly, an algorithm based on the atmospheric turbidity transient behaviour has been
developed. It takes advantage of the fact that atmospheric turbidity fluctuations are low through the day. The
algorithm works as follows (Fig. 2):

Fig. 2. Diagram of the developed knowledge model.

2004 J. Nou et al. / Energy Procedia 69 (2015) 1999 – 2008

Each time DNI ൫‫)ݐ(ܫ‬൯ is measured (under clear sky conditions or not), a corresponding coefficient related to
atmospheric turbidity ‫ )ݐ( ்ܥ‬is calculated (A) using equation (2). This coefficient ‫ )ݐ( ்ܥ‬is equal to atmospheric
turbidity ܶ௅ூ (‫ )ݐ‬only if the measured DNI ൫‫)ݐ(ܫ‬൯ corresponds to the clear sky DNI ൫‫ܫ‬஼ௌ (‫)ݐ‬൯. So, in order to decide if
‫ )ݐ( ்ܥ‬corresponds or not to atmospheric turbidity, changes in atmospheric turbidity are considered via its maximum
plausible change οܶ௟௜௠ (C) and its measured change οܶ (D). These two variables are computed from the last time
(‫ݐ‬௣௔௦௧ ) atmospheric turbidity ܶ௅ூ is trustworthy (B). So, ‫ݐ‬௣௔௦௧ is the last time the coefficient related to atmospheric
turbidity ‫ )ݐ( ்ܥ‬is used as atmospheric turbidity. The threshold οܶ௟௜௠ has been computed on the basis of the transient
atmospheric turbidity behaviour we highlighted in our database. Indeed, the bigger ο‫ ݐ‬is, the more plausible the
atmospheric turbidity increase is. However, according to its daily behaviour (atmospheric turbidity never increases
all the day), a maximum change is defined. Fig. 3 shows how οܶ௟௜௠ has been computed according to ο‫ݐ‬.

Fig. 3. οܶ௟௜௠ vs. ο‫ݐ‬.

For now, a simple linear interpolation has been performed in order to determine οܶ௟௜௠ . Therefore, οܶ௟௜௠ is given
by equation (4), with ‫ܥ‬ଵ and ‫ܥ‬ଶ the slope and the intercept of the curve (between ο‫ݐ‬௠௜௡ and ο‫ݐ‬௠௔௫ ), respectively (5):

οܶ௟௜௠ = min(‫ܥ‬ଵ . ο‫ ݐ‬+ ‫ܥ‬ଶ , οܶ௠௔௫ ) (4)

‫ܥ‬ଵ = (οܶ௠௔௫ െ οܶ௠௜௡ )Τ(ο‫ݐ‬௠௔௫ െ ο‫ݐ‬௠௜௡ ) and ‫ܥ‬ଶ = οܶ௠௜௡ െ ο‫ݐ‬௠௜௡ ή ‫ܥ‬ଵ (5)

ο‫ݐ‬௠௜௡ and οܶ௠௜௡ mean that atmospheric turbidity cannot increase more than οܶ௠௜௡ for two consecutive DNI
acquisitions ο‫ݐ‬௠௜௡ (ο‫ݐ‬௠௜௡ = 20 seconds and οܶ௠௜௡ = 0.045). ο‫ݐ‬௠௔௫ and οܶ௠௔௫ mean that atmospheric turbidity
cannot increase more than 1 through the day (ο‫ݐ‬௠௔௫ = 10 hours and οܶ௠௔௫ = 1). These four parameters have been
set taking into account the acquisition time step, accuracy in RSI measurements as well as changes in atmospheric
turbidity one can observe in the database. As a result, they are proper to both the considered site and measurement
devices used. οܶ is about the measured change in atmospheric turbidity. It is computed as follows (6):

οܶ = ‫ )ݐ( ்ܥ‬െ ܶ௅ூ ൫‫ݐ‬௣௔௦௧ ൯ (6)

Once these two thresholds are computed ( οܶ௟௜௠ and οܶ), a decision based on their comparison is taken (E) in
order to determine if the computed coefficient ‫ )ݐ( ்ܥ‬is, or not, the "real" atmospheric turbidity ܶ௅ூ (‫)ݐ‬. This decision
is taken according to the following conditions (7):

οܶ > οܶ௟௜௠ or ‫ܶ < ்ܥ‬௠௜௡ or ‫ܶ > ்ܥ‬௠௔௫ (7)

where ܶ௠௜௡ and ܶ௠௔௫ are constants defined empirically (using our database) and derived from the minimum and
maximum possible values of atmospheric turbidity. If (7) is true (i.e. ‫ ்ܥ‬is not plausible), the mean atmospheric
turbidity value ‫ܶۃ‬௅ூ ‫ ۄ‬calculated over the last five minutes is assigned to ܶ௅ூ (‫)ݐ‬. However, the time of the day impacts
the decision process. For example, during sunrise or sunset, ‫ܶۃ‬௅ூ ‫ ۄ‬is consistently used because atmospheric turbidity is
not easy to compute in such periods. This approach provides more stable values. Sunrise and sunset have been
determined from both the air mass ݉ and time, as shown in Table 1. Finally, once atmospheric turbidity ൫ܶ௅ூ (‫)ݐ‬൯ is
computed, the clear sky DNI ൫‫ܫ‬஼ௌ (‫)ݐ‬൯ is obtained (F) by inverting equation (2).
 J. Nou et al. / Energy Procedia 69 (2015) 1999 – 2008 2005

Table 1. Sunrise and sunset determining.

Sunrise Sunset
m > 10 and time < 12:00 p.m. m > 6 and time > 12:00 p.m.

3.3. Model evaluation

In order to validate our model, an evaluation methodology has been defined. This methodology is based on the
different steps presented by Fig. 4. The main target is to quantify model accuracy. The idea is then to generate a
signal similar to DNI and a noise is added to simulate variations caused by clouds. This noise is only added during
some clear sky instants in order to use clear sky DNI measurements as reference during the validation step. The next
section of the paper gives more details about this process.

Fig. 4. Overview of the five steps used for model evaluation.

From the measured DNI (‫ )ܫ‬and thanks to pictures taken by our sky imager (in order to be sure that the sun disk
appears clearly), clear sky instants have been visually highlighted (a). ‫ ݑ‬is the resulting signal (b). It is a binary
sequence where 0 corresponds to a sun occulted by clouds while 1 is about clear sky. On Fig. 4, only five
consecutive days are plotted. However, our study considers 45 days of the database, which is about 150 hours of
clear sky data. In order to evaluate our clear sky model, we degraded a part of these 150 hours of data (c).
Considering this noisy signal, we compared the clear sky values given by our model and measured clear sky data
(without noise). Firstly, we generated a random sequence denoted ‫ݒ‬ƍ (in which each element ‫ݒ‬ƍ௜ ‫[ א‬0,1]). From ‫ݒ‬ƍ
we generated ‫ݒ‬, the Pseudo Random Binary Sequence (PRBS), by using (8), according to the ratio ܴ଴/ଵ dealing with
the proportion of "1" in such a sequence:

if ‫ݒ‬௜ ƍ > ܴ0/1

then ‫ݒ‬௜ = 0 (8)
else ‫ݒ‬௜ = 1

In our study, this ratio ranges between 0.1 and 0.9 (because it can be very different, depending on the season of
the year). As a consequence, the initial signal ‫ ܫ‬has been degraded according to ܴ଴/ଵ . For instance, a ratio ܴ଴/ଵ equal
to 0.7 means that the initial clear sky DNI signal is degraded by 70 %. Then, multiplying the sequences ‫ ݑ‬and ‫ ݒ‬term
by term, we obtained the sequence ‫( ݓ‬d). This sequence was determined by using the following expression (9):

‫ݓ‬ଵஸ௜ஸே = (‫ݑ‬௜ × ‫ݒ‬௜ )ଵஸ௜ஸே (9)

2006 J. Nou et al. / Energy Procedia 69 (2015) 1999 – 2008

where ݅ is about the position of an element in a given vector and ܰ is the total number of elements in that vector.
Finally, from this signal, we generated another signal ‫ ீܫ‬, used as input of our model (e). It is defined as follows:
when ‫ ݓ‬is equal to 0, ‫ ܫ = ீܫ‬and when ‫ ݓ‬is equal to 1, ‫݇ × ܫ = ீܫ‬௥௔௡ௗ . ݇௥௔௡ௗ is a sequence of pseudo random values
drawn from the standard distribution on the interval [0,1]. From this methodology, it is possible to test our model
and evaluate its efficiency knowing the real values of the clear sky DNI (when ‫ݑ‬௜ = 1, with 1 ൑ ݅ ൑ ܰ). In order to
quantify the model accuracy, we considered both the Mean Absolute Error (MAE) and the Normalized Root-Mean-
Square Error (NRMSE), defined as follows (10):

ට1 σ൫‫ܫ‬஼ௌ െ ‫ܫ‬መ஼ௌ ൯
ଶ
1 ݊ (10)
‫ = ܧܣܯ‬෍ห‫ܫ‬஼ௌ െ ‫ܫ‬መ஼ௌ ห ܴܰ‫ = ܧܵܯ‬100 ×
݊ ‫ܫ‬஼ௌ೘ೌೣ െ ‫ܫ‬஼ௌ೘೔೙

with ‫ܫ‬஼ௌ೘೔೙ , ‫ܫ‬஼ௌ೘ೌೣ , ‫ܫ‬஼ௌ and ‫ܫ‬መ஼ௌ the minimum, maximum, measured and estimated values of the clear sky DNI,
respectively. The results we obtained using the proposed knowledge-based model are presented in the following
section of the paper (section 4).

4. Results and discussion

Using our knowledge model and the noisy signal ‫ ீܫ‬, one can obtain clear sky DNI values. To obtain more
relevant and reliable results, we generated ten times the pseudo random binary sequence ‫ݓ‬. The impact of ܴ଴/ଵ on
both the MAE and NRMSE is highlighted by Table 2.

Table 2. Impact of ܴ଴/ଵ on both the MAE and NRMSE.

ܴ଴/ଵ 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
MAE (W.m-2) 16.1 17.5 19.2 20.3 21.4 22.7 24.6 27.2 33.4
STD MAE (W.m-2) 4.77 4.68 3.44 2.95 3.05 2.84 1.83 1.90 2.46
NRMSE (%) 1.01 1.26 1.56 1.63 1.72 1.79 1.90 2.14 2.77
STD NRMSE (%) 0.63 0.82 0.65 0.58 0.58 0.50 0.35 0.31 0.32

Taking a look at this table, one can notice that the more the ratio ܴ଴/ଵ increases (i.e. the higher the degradation of
the signal) the quicker the MAE and NRMSE increase. Indeed, considering a high degradation, clear sky data are
less available to facilitate a good computation of atmospheric turbidity and, therefore, the clear sky DNI. With a
ratio ܴ଴/ଵ set to 0.1, the MAE and NRMSE are 16.1 W.m-2 and 1.01%, respectively. With a ratio ܴ଴/ଵ set to 0.9, the
MAE and NRMSE are 33.4 W.m-2 and 2.77%, respectively. Moreover, we clearly notice that the results are very
dependent on the operating conditions considered for model evaluation ( ܴ଴/ଵ values). According to the DNI
measurements realized for this study, the mean ratio of clear sky data during a day is close to 32 %. This
corresponds to a ratio ܴ଴/ଵ equal to 0.68. So, the results obtained for the on-site measured ratio (ܴ଴/ଵ = 0.7) are
satisfactory because of a MAE equal to 24.6 W.m-2 and a NRMSE equal to 1.9% (Table 2). Fig. 5 presents the
results obtained using the proposed clear sky model for three consecutive days of March 2014. These three days
have been selected in order to highlight how the model is able to work with different DNI behaviours. ܴ଴/ଵ has been
set to 0.7 in order to be representative of on-site measurements.
The three selected days (March 1 to March 3) have different solar irradiance profiles. For each day, we can note
sharp and frequent changes in DNI (i.e. cloudy periods of limited duration) and periods during which DNI is
insignificant (overcast). Indeed, when the coefficient related to atmospheric turbidity ൫‫)ݐ( ்ܥ‬൯ does not correspond to
atmospheric turbidity (i.e. condition (7) is true), the mean value of atmospheric turbidity is used (if it is not available
during the last five minutes, like during sunrise, the last value of the day before is used). However, when ‫ )ݐ( ்ܥ‬can
be considered as equal to atmospheric turbidity (i.e. condition (7) is false), the values of ܶ௅ூ (‫ )ݐ‬are updated. These
updates produce more or less important changes in the red curve, depending on the time between two measurements
or the magnitude of the update.
Generally, the clear sky DNI values given by the model form an upper limit envelope (the red curve on Fig. 5)
and clear sky DNI fluctuations, caused by the atmospheric turbidity updates, are relatively slight. It was a key point
 J. Nou et al. / Energy Procedia 69 (2015) 1999 – 2008 2007

to obtain slight fluctuations on clear sky DNI estimation in order to avoid problems during the short-term DNI
forecasting process. Indeed, as previously explained, the development of a reliable and accurate clear sky model is
an essential step in solar resource forecasting. Thus, the more reliable the forecast is, the more efficient the real-time
management of a solar power plant will be.

Fig. 5. Measured clear sky DNI and clear sky DNI estimated by the model, with ܴ଴/ଵ set to 0.7.

5. Conclusion and outlook

The CSPIMP (Concentrated Solar Power efficiency IMProvement) project started in 2013 in order to achieve a
better competitiveness for the CSP plants. One challenge in the project is to improve forecast of the solar resource in
order to optimize plants operation. The development of a clear sky model is therefore an essential step to forecast
Direct Normal Irradiance (DNI) because clear sky represents the nominal operating conditions of a concentrating
solar power plant. So, this paper focuses on estimating DNI under clear-sky conditions using a knowledge model
based on the coefficient of relative optical air mass, DNI and atmospheric turbidity fluctuations. This model takes
advantage of the lower variability of atmospheric turbidity, in comparison to the DNI variability.
The empirical formulation of atmospheric turbidity proposed by Ineichen and Perez has been selected in this
study because it proved to be less dependent on air mass than the widely-used Linke turbidity coefficient; it is also
well adapted for on-site and real-time applications. A validation procedure has been proposed and allowed the
model accuracy to be highlighted. With a ratio ܴ଴/ଵ equal to 0.7 (see section 3.3), the MAE and NRMSE are
25 W.m-2 and 2 %, respectively. In addition, the structure of the proposed knowledge-based model makes it well
adapted to real-time applications. As a result, it will be soon implemented and validated on site (Palma del Rio CSP
plant, Spain).
Future work will focus on evaluating the error distribution in the course of the day (error vs. air mass) as well as
new options to deal with ‫( )ݐ( ்ܥ‬i.e. the coefficient related to atmospheric turbidity) not corresponding to atmospheric
turbidity. Finally, in order to forecast at short-term horizon the clear sky DNI and then DNI for all types of sky
conditions, a persistent model based on atmospheric turbidity will be developed and evaluated. For such a forecast
horizon, this approach has proven to be adequate, due to the low variability of atmospheric turbidity.

References

[1] Lara-Fanego V, Ruiz-Arias JA, Pozo-Vázquez D, Santos-Alamillos FJ, Tovar-Pescador J. Evaluation of the WRF model solar irradiance
forecasts in Andalusia (southern Spain). Solar Energy 2012;86:2200–17. doi:10.1016/j.solener.2011.02.014.
[2] Mefti A, Adane A, Bouroubi MY. Satellite approach based on cloud cover classification: Estimation of hourly global solar radiation from
meteosat images. Energy Conversion and Management 2008;49:652–9. doi:10.1016/j.enconman.2007.07.041.
[3] Reikard G. Predicting solar radiation at high resolutions: A comparison of time series forecasts. Solar Energy 2009;83:342–9.
doi:10.1016/j.solener.2008.08.007.
[4] Quesada-Ruiz S, Chu Y, Tovar-Pescador J, Pedro HTC, Coimbra CFM. Cloud-tracking methodology for intra-hour DNI forecasting. Solar
Energy 2014;102:267–75. doi:10.1016/j.solener.2014.01.030.
[5] Chow CW, Urquhart B, Lave M, Dominguez A, Kleissl J, Shields J, et al. Intra-hour forecasting with a total sky imager at the UC San
Diego solar energy testbed. Solar Energy 2011;85:2881–93. doi:10.1016/j.solener.2011.08.025.
2008 J. Nou et al. / Energy Procedia 69 (2015) 1999 – 2008

[6] Marquez R, Coimbra CFM. Intra-hour DNI forecasting based on cloud tracking image analysis. Solar Energy 2013;91:327–36.
doi:10.1016/j.solener.2012.09.018.
[7] Perez R, Kivalov S, Schlemmer J, Hemker Jr K, Renné D, Hoff TE. Validation of short and medium term operational solar radiation
forecasts in the US. Solar Energy 2010;84:2161–72. doi:10.1016/j.solener.2010.08.014.
[8] Nou J, Chauvin R, Traoré a., Thil S, Grieu S. Atmospheric Turbidity Forecasting using Side-by-side ANFIS. Energy Procedia
2014;49:2387–97. doi:10.1016/j.egypro.2014.03.253.
[9] Ineichen P, Perez R. A new airmass independent formulation for the Linke turbidity coefficient. Solar Energy 2002;73:151–7.
doi:10.1016/S0038-092X(02)00045-2.
[10] Blanc P, Wald L. The SG2 algorithm for a fast and accurate computation of the position of the Sun for multi-decadal time period. Solar
Energy 2012;86:3072–83. doi:10.1016/j.solener.2012.07.018.
[11] Gueymard C. An Introduction To Solar Radiation. Encyclopedia of Sustainability Science and Technology 2012.
[12] Gueymard CA. REST2: High-performance solar radiation model for cloudless-sky irradiance, illuminance, and photosynthetically active
radiation – Validation with a benchmark dataset. Solar Energy 2008;82:272–85. doi:10.1016/j.solener.2007.04.008.
[13] Kasten F. The linke turbidity factor based on improved values of the integral Rayleigh optical thickness. Solar Energy 1996;56:239–44.
doi:10.1016/0038-092X(95)00114-7.