Energy 83 (2015) 387e393

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

Solar radiation forecasting in the short- and medium-term under all

sky conditions
J. Alonso-Montesinos a, b, F.J. Batlles a, b, *
a
Department of Chemistry and Physics, University of Almería, 04120 Almería, Spain
b
CIESOL, Joint Centre of the University of Almería-CIEMAT, 04120 Almería, Spain

a r t i c l e i n f o a b s t r a c t

Article history: Meteorological conditions are decisive in solar plant management and electricity generation. Any in-
Received 14 October 2014 creases or decreases in solar radiation mean a plant has to adapt its operation method to the climato-
Received in revised form logical phenomena. An unexpected atmospheric change can provoke a range of problems related to
22 January 2015
various solar plant components affecting the electricity generation system and, in consequence, causing
Accepted 13 February 2015
Available online 16 March 2015
alterations in the electricity grid. Therefore, predicting atmospheric features is key to managing solar
plants and is therefore necessary for correct electrical grid management. Accordingly, a solar radiation
forecast model is presented, where the three solar components (beam, diffuse and global) are predicted
Keywords:
Forecasting
over the short- and medium-term (up to three hours) for all sky conditions, demonstrating its potential
Solar radiation as a useful application in decision-making processes at solar power plants.
Cloud motion © 2015 Elsevier Ltd. All rights reserved.
Satellite images
CSP plants
PV systems

1. Introduction plant's management and, consequently, the resultant electricity

Electricity production from renewable energy sources is
increasingly competing with conventional electricity generation
methods [1] as a means of avoiding a variety of environmental and
economic problems [2e4]. Over recent years, solar energy has
become more relevant with an expansion in the number of solar
plants in operation [5]. CSP (Concentrated solar power) plants and
another system [PV (photovoltaic) plants] are sophisticated elec-
tricity generation methods which work best under favorable
meteorological conditions. Knowing when clouds or attenuating
atmospheric factors might appear over the solar field helps avoid
various problems, such as thermal stress on concentrator mirrors or
on CSP plants receivers [6]. Other problems are related to electricity
integration into the grid, where electricity generation instability
might appear. Therefore being able to predict meteorological phe-
nomena is important to solar plant operators, as this can affect the
* Corresponding author. Department of Chemistry and Physics, University of
Almería, 04120 Almería, Spain. Tel.: þ34 950 015914.
E-mail address: fbatlles@ual.es (F.J. Batlles).
generation.
A range of technologies have been used to study different at-
mospheric features.
Over the years, satellite images have built up a spatial view of
the Earth's surface. These images have been used by many scientists
to develop a variety of atmospheric studies, such as those for cloud
cover. Cloud cover was examined in Ref. [7], where the authors
presented an estimation method using the thermal-infrared
channel at night time. In addition to this, the sky was classified
into three different sky conditions: cloudless, partially-cloudy and
overcast [8]. Cloud identification also using satellite images, was
carried out in Ref. [9], where the authors classified clouds into three
different heights (high, medium and low) and studied cloud motion
using the maximum cross-correlation method. Cloud cover was
likewise studied using another emerging technology, the total sky
imager where sky cameras are installed at ground level, giving a
different sky vision to the satellite images. A sky classification was
presented in Ref. [10]; the authors used a KNN (K nearest neigh-
bors) classification algorithm to divide the sky into seven different
groups according to cloud presence. Combining sky camera images
with radiometric data, the clouds were classified into four groups,
showing radiation attenuation based on cloud type [11]. in addition

http://dx.doi.org/10.1016/j.energy.2015.02.036
0360-5442/© 2015 Elsevier Ltd. All rights reserved.
388 J. Alonso-Montesinos, F.J. Batlles / Energy 83 (2015) 387e393

to this, a cloud identification system was developed where the sky

camera images were processed according to a sky classification MSG satellite
[12], thus solving the saturation problem in the solar area [13]. By image collection
combining both satellite and sky camera imagery, a cloudiness
forecast was made for the short- and medium-term [14], where a
user interface provided cloudiness forecasting in real-time [15].
Other studies have focused on solar radiation estimation.
Traditionally, solar radiation was estimated using parametric and Cloud detection 3 consecutive
images
empirical models. Parametric models required detailed atmo-
spheric information in order to calculate atmospheric trans-
mittance [16e19]. Empirical models were obtained from different
correlations [20e22]. Consequently, some authors began to use
Yes
satellite images to quantify solar radiation. The model developed in
Ref. [23], was tested using an independent dataset of 1-year hourly
ground measurements in a location of USA [24]. Global and direct Determination of
horizontal irradiance were estimated from the model, whereas that cloud motion vectors
(CMV)
diffuse horizontal irradiance was estimated from the other irradi-
ance components. The rRMSE (relative root mean square error)
values were 21.5%, 40.9% and 54.2% for global, direct and diffuse
irradiance, respectively, and the rMBE (relative mean bias error)
values were 4.9%, 2% and 15.4% in the aforementioned order. Also,
Solar radiation evaluation
the model developed for estimating direct normal irradiance in Ref.
(Heliosat 2)
[25] was tested over eight stations in Saudi Arabia for hourly
satellite-derived DNI (Direct normal irradiance) values and ob-
tained an rRMSE of 36.1% and an rMBE of 4.3% for all sky conditions
[26]. Global solar radiation was also estimated using MSG
(Meteosat Second Generation) satellite imagery [27]. Moreover,
solar radiation was estimated above Turkey, where the authors Applying CMV to
used satellite images and artificial neural network [28]. One of the last MSG satellite image
most consolidated models for estimating solar radiation under
clear sky conditions was been used in the Heliosat-2 method to
estimate the three solar radiation components, this was developed
in the ESRA (European Solar Radiation Atlas) [29]. One study [30]
combined the Heliosat-2 method with digital terrain models to Solar radiation forecasting
obtain a solar radiation estimation, whereas [31] made a global 12 times x 15 min = 180 min
radiation prediction combining two methods: one focused on sat-
ellite images with the Heliosat-2 method; and the other focused on
ground-based lineal model. Fig. 1. Flowchart of the methodology developed for the solar radiation forecasting.

As demonstrated above, the study of atmospheric features is

solar altitude was higher than 10 e this was done to avoid mis-
constantly evolving and expanding. However, to improve solar
takes in cloud detection due to low image brightness. (see Fig. 1)
systems and solar power plants management, predicting solar ra-
Furthermore, diffuse and global irradiance measurements were
diation has proven to be essential information [32], greatly
taken from two CMP11 Kipp & Zonen pyranometers, while beam
contributing to stable electricity generation and its integration
irradiance was measured using a CH1 Kipp & Zonen pyrheliometer;
from solar plants into the public electricity grid.
these instruments were installed on a horizontal two-axes solar
In this work, we have looked at the need to provide future at-
tracker.
mospheric information by developing a system capable of pre-
The testing facility is located at the CIESOL (Solar Energy
dicting the three solar radiation components (beam, diffuse and
Research Center) at the University of Almería, Spain (36.8oN, 2.4oW,
global) over the short- and medium-term (up to three hours) under
at sea level), which has a Mediterranean climate and a high mari-
all sky conditions, and then comparing the results with ground-
time aerosol presence.
level radiometric measurements.

2.2. Image collection and cloud detection

2. Materials and methods
In a general way, the structure of the methodology can be
defined as follow:
2.1. Data collection
Satellite data from the years 2010e2014 were used. For the
forecasting estimation, five MSG satellite channels were collected
(VIS 0.6, VIS 0.8, IR 3.9, IR 10.8 and IR 12.0) out of a total of 12
spectral channels received every 15 min. For each of the five chosen
channels, a 121  161 pixel raster image, centered on the University
of Almería, was saved and then used via an algorithm whenever the
The first step in carrying out solar radiation forecasting is correct
satellite image collection. The images are received and converted
into reflectance and temperature (from visible and infrared chan-
nels, respectively).
When the images are converted, the atmospheric conditions can
be obtained for each satellite pixel, combining spectral channels
with different remote sensing techniques. Clouds are the most
attenuating factor in the solar radiation environment. Cloud iden-
tification in satellite images, is carried out using various spectral
tests propose in a study by Ref. [9], in which the studied area is fully
analyzed, looking for the presence of any clouds in the different
atmospheric layers.
 J. Alonso-Montesinos, F.J. Batlles / Energy 83 (2015) 387e393
2.3. Determination of cloud motion vectors
An important issue in radiation forecasting is understanding the
evolution of atmospheric elements e future scenarios are a
consequence of present and previous situations.
As mentioned above, clouds are the most attenuating factor in
solar radiation. With satellite images, the clouds are identified at
the pixel level, thus, cloud motion can represent general pixel
motion. Since cloud motion defines representative general pixel
motion, the motion of attenuation factors has also been repre-
sented using cloud motion.
In the following subsections, the methodology is presented for
predicting these attenuating factors. Here, the method for studying
cloud motion is laid out, based on the methodology proposed in
Ref. [9]. To do this, three consecutive satellite images were used,
each split into five sectors. In each sector, the maximum cross
correlation method was applied, producing a pixel maximum be-
tween two consecutive images, where the two compared sectors
are most similar, thus obtaining the representative cloud motion.
Subsequently, to apply various quality tests, one motion vector is
obtained for each sector. A total of 12 displacements were carried
out, providing a displacement for the following 180 min (15 min for
each displacement). In this way, the future cloud positions were
determined and, as a consequence, the future pixel positions.
2.4. Solar radiation evaluation
In this section, the solar radiation evaluation methodology is
presented. Generally, the solar radiation is represented by Eq. (1):
Radest ¼ Radcs Kci ;
where Radest is the solar radiation estimation (the beam, diffuse or
global component) expressed in Wm2 under clear skies; and Kci is
an atmospheric attenuation factor known as the cloudiness index.
Therefore, to evaluate solar radiation, it is necessary to obtain a
solar radiation estimation under clear skies. In this work, the ESRA
model [29] has been used to estimate solar radiation under
cloudless skies. The ESRA model was compared with other similar
models, offering more accuracy in most cases and stations where
the models were compared. Consequently, the solar radiation
estimation was carried out on the third consecutive image to have
the latest radiation estimation. Beam radiation was calculated as
shown in Eq. (2):
BESRA ¼ I0 ε sin aTrb ðaÞ;
where I0 is the solar constant equal to 1367 Wm2; ε is the
correction used for the variation of the sun-earth distance from its
mean value; a is the solar altitude angle (in radians) and Trb(a)
represents the beam transmittance (of the beam radiation) under
cloudless skies as a function of the a angle.
The diffuse radiation component also depends on the Linke
turbidity factor and is estimated using the ESRA model. Eq (3)
shows the expression to estimate this component:
DESRA ¼ I0 εTrd ðTL ÞFd ða; TL Þ;
where Trd(TL) is the transmittance at zenith for diffuse radiation
and Fd represents the diffuse angular function.
The last radiation component is the global radiation, which is
calculated using Eq. 4
GESRA ¼ BESRA þ DESRA :
As solar radiation is not only attenuated by clouds, it is necessary
389
to make a complete attenuation study. To do this, an atmospheric
model was implemented to obtain the total atmospheric attenua-
tion (Kci).
2.4.1. Attenuation model e Heliosat 2
The Heliosat 2 model was utilized to study solar radiation using
satellite images. During the 1980s, many authors proposed this
method's first approximations [33,34]. One study focused on the
Heliosat model [35], where the authors presented a Heliosat-2
method approach. Taking this article into account, and consid-
ering this method as one of the most accurate for solar radiation
mapping using Meteosat images, we have used Heliosat-2 to
determine the atmospheric state in the estimation of beam, diffuse
and global solar radiation. Hence, it is possible to determine the
solar radiation attenuation that does not reach the ground.
As with the solar radiation estimation, the attenuation factor is
only calculated for the third consecutive image, following Expres-
sion 5:
8
>
> n <  0:2; Kci ¼ 1:2
<
0:2 < n < 0:8; Kci ¼ 1  n
(5)
>
> 0:8 < n < 1:1; Kci ¼ 2:0667  3:667n þ 1:667n2
:
n > 1:1; Kci ¼ 0:05
where n is the cloud index, obtained using Eq. (6):
  
n ¼ r  rg rc  rg ; (6)
where r is the apparent albedo, rg is the ground albedo and rc is the
cloud albedo.
(1)
2.5. Solar radiation forecasting
When the solar radiation and the attenuation factor have been
calculated, the future scenarios then have be reproduced. Instead of
determining solar radiation in the short and medium-term, we
have determined the cloudiness index in the future scenarios. To do
this, we have applied cloud motion vectors (Section 2.3) to each
pixel of the raster image analyzed.
Specifically, we have moved the cloudiness index values and not
the estimated radiation values, because these indices are the most
representative factors related to solar attenuation, whereas the
solar radiation estimation is simply the pixel solar radiation esti-
mation under clear skies, which can be calculated for a given
moment and position.
(2) Lastly, we need to calculate the resultant solar radiation over the
following three hours. To do this, the cloudiness indices are moved
twelve times, one for each 15 min, according to the cloud motion
vectors. Each time, the solar radiation is estimated for a cloudless
sky in all the pixels, where the product between the radiation
estimated in a pixel and the cloudiness index in the same pixel
generates the final solar radiation estimation for each radiation
component.
As shown above, a novel combination of cloud motion vectors
and solar radiation estimations has been developed to give prac-
tical information about radiation behavior over the short- and
(3) medium-term; thus providing a decisive tool in the solar energy
field.
3. Results
In this section, the solar radiation forecasting results are pre-
sented. A total of four year were analyzed (2010e2014) where the
(4)
beam, diffuse and global radiation component have been predicted
over the short- and medium-term. To compare the predicted
390 J. Alonso-Montesinos, F.J. Batlles / Energy 83 (2015) 387e393
radiation with real radiation values, data from the radiometric
sensor were also collected. As the approximate time spent receiving
a satellite image is 15 min, the predicted values are likewise made
for 15-min periods; thus, these predicted values can be compared
to the mean radiation values over these previous 15-min periods.
For all the results, we have obtained a range of parameters: the
RMSE (root-mean-square error) given by Eq. (7) and expressed in
Wm2; the nRMSE (normalized RMSE) given by Eq. (8) and
expressed in %; the MBE (mean bias error) given by Eq. (9) in
Wm2; the nMBE (normalized MBE) given by Eq. (10) and
expressed in % and the dimensionless correlation coefficient R
given by Eq. (11):
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
u1 X N
RMSE ¼ t ðRest  Rmea Þ2
N i¼1
where N is the total number of estimations; Rest is the radiation
estimation and Rmea is the measured radiation.
nRMSE ¼ ½RMSE=ðRmax  Rmin Þ100
1 XN
MBE ¼ ðRest  Rmea Þ
N i¼1
nMBE ¼ ½MBE=ðRmax  Rmin Þ100
 difference might be representative but, looking at the RMSE values,
R ¼ sRest Rmea sRest sRmea ;
where sRest;mea is the covariance of the two input data sets (the
estimated and measured radiations), sRest is the standard Rest de-
viation and sRmea is the standard Rmea deviation.
The parameters were calculated for all solar radiation compo-
nents (beam, diffuse and global). Moreover, the results were clas-
sified into four groups. The first group includes the results for all
sky conditions. For the second group, we selected days where the
skies were cloudless, obtaining the results for clear-sky conditions.
Another group was created where the selected days included the
presence of cloud transients, with the solar radiation varying very
quickly between low and high values. Finally, the last group was
collated to consider the results under overcast skies. For this, we
selected days when solar radiation was low throughout the day.
3.1. Results for all sky conditions
Under all sky conditions, moments were processed in which
three satellite images were consecutive. For beam radiation fore-
casting, a total of 30,383 registers was analyzed; for diffuse fore-
casting, a total of 31,100 registers were processed whereas global
radiation was forecast 31,094 times. Table 1 shows the forecasting
results for all solar radiation components in the subsequent three
hours. As can be seen, the FCST (Forecast) represents the forecast
time from 15 min to 180 min. Over this time period, various sta-
tistical indicators are presented for the three radiation components.
Looking at the area for beam radiation forecasting, the mean
RMSE value was about 190 Wm2, supposing an average nRMSE of
21%. In the case of the nMBE, the values decreased from 16% for
15 min to approximately 12% for 180 min whereas the R value had
an average of about 0.80. In the case of diffuse radiation prediction,
the RMSE oscillated with values close to 100 Wm2, and an average
nRMSE value of about 16%. The MBE and the nMBE indicated that
this radiation component was mainly underestimated, with
normalized MBE values approximately 7%. However, the R values
were close to 0, meaning high dispersion between measured and
estimated values. Global radiation presented the best results, with
an average nRMSE value of about 10% over the three hours, an
average nMBE value of about 7% and an R value higher than 0.90 in
all cases.
After analyzing the data carefully, the values did not present
wide variations over time. Mainly, this was because of the accuracy
in the pixel-motion vectors, which had good pixel motion precision,
providing homogeneous results in general. Considering the low
error in the pixel-motion vectors [9], greater importance was
focused on the pixel radiation estimation, which determined the
future radiation values.
3.2. Results under clear-sky conditions
(7) Clear-sky conditions are produced on days where there is no
cloud presence. For beam radiation forecasting, a total of 344 days
were collected where 9604 registers were analyzed. For diffuse and
global radiation forecasting, 347 days were available with a total of
9845 moments processed. Table 2 shows the radiation forecasting
results.
(8)
Cloudless skies represented the greatest model accuracy for
solar radiation forecasting. In the case of beam radiation, the
nRMSE results remained constant over the three hours with an
(9)
average value of approximately 9%, the RMSE varying between 79
and 86 Wm2; whilst the nMBE was about 6% over the total forecast
time with an R value higher than 0.94. The diffuse component had
(10) an nRMSE value varying from 14% at 15 min to 8% at 180 min. This
(11) these only varied in a few units therefore, these differences are not
considered representative; for this radiation component, the nMBE
produced low values (lower than 4%), whilst the average R value
was 0.50. Lastly, global radiation was the best predicted radiation
component, with an average nRMSE value of about 7% (the RMSE
value varied between 61 and 68 Wm2); whereas the nMBE value
went from approximately 8%e5%, with a difference between
maximum and minimum MBE values of 20 Wm2 and with an R
value close to 1 in all cases.
As can be observed, the proposed methodology for predicting
solar radiation using satellite images is highly accurate under
cloudless sky conditions providing significant progress in solar ra-
diation forecasting.
3.3. Results under partially-cloudy sky conditions
Partially-cloudy days were those days when solar radiation
changed quickly due to the presence of cloud transients. These
transients provoke constant variations in solar radiation behavior,
making it difficult to predict when the clouds will not block the
sunbeam. However, in this work, we have predicted solar radiation
under all sky conditions. A total of 354 days of data were collected,
during which, beam radiation component was carried out 7378
times; whereas diffuse and global radiation were predicted a total
of 7532 times over 355 days. Table 3 shows the forecasting results
on these kinds of days.
Here, for the beam radiation component, the nRMSE had an
approximate value of 29% with a greater difference between the
measured and the predicted values (of about 250 Wm2); the nMBE
value varied between 19 and 23%, whilst the R value was approx-
imately 0.70 for the total period predicted. In the case of diffuse
radiation, the nRMSE was practically constant, with a value of
approximately 21%; the nMBE values were always negative, thus
underestimating the solar radiation component, for which the
average value was 13% whereas the R value was very low in all
situations, except in the last cases, where the value was close to 0.5.
Once again, global radiation presented better results, with the
 J. Alonso-Montesinos, F.J. Batlles / Energy 83 (2015) 387e393 391

Table 1
Radiation forecasting results under all sky conditions.

FCST. time Beam radiation Diffuse radiation Global radiation

RMSE nRMSE MBE nMBE R RMSE nRMSE MBE nMBE R RMSE nRMSE MBE nMBE R

15 min 190.88 21.61 139.41 15.78 0.8339 104.33 16.11 44.82 6.92 0.0427 111.74 10.14 84.86 7.71 0.9479
30 min 192.35 21.77 139.03 15.74 0.8205 105.21 16.25 45.69 7.05 0.0190 113.83 9.85 83.15 7.20 0.9369
45 min 193.76 21.93 138.05 15.63 0.8069 105.88 16.35 46.33 7.16 0.0055 116.72 10.10 81.05 7.02 0.9245
60 min 194.32 22.00 136.25 15.42 0.7972 106.19 16.40 46.71 7.21 0.0074 118.61 10.27 78.70 6.81 0.9165
75 min 194.07 21.97 133.60 15.12 0.7922 106.08 16.38 46.81 7.23 0.0209 120.38 10.42 75.80 6.56 0.9114
90 min 193.77 21.93 130.51 14.77 0.7910 105.74 16.33 46.63 7.20 0.0445 122.64 10.62 72.83 6.31 0.9084
105 min 193.18 21.87 127.77 14.46 0.7960 105.29 16.26 46.22 7.14 0.0805 123.60 10.70 70.49 6.10 0.9115
120 min 191.39 21.66 124.14 14.05 0.8062 104.55 16.14 45.63 7.04 0.1265 123.25 10.67 67.43 5.84 0.9177
135 min 190.09 21.52 120.41 13.63 0.8157 103.21 15.94 44.69 6.90 0.1823 123.54 10.69 64.67 5.60 0.9230
150 min 188.64 21.35 116.24 13.16 0.8256 101.85 15.73 43.60 6.73 0.2412 123.95 10.73 61.69 5.34 0.9279
165 min 186.12 21.07 111.51 12.62 0.8372 100.27 15.48 42.39 6.55 0.3030 123.10 10.65 58.38 5.05 0.9337
180 min 182.88 20.70 106.48 12.05 0.8487 98.38 15.19 40.86 6.31 0.3618 121.41 10.55 55.20 4.87 0.9388

Table 2
Radiation forecasting results under clear-sky conditions.

FCST. time Beam radiation Diffuse radiation Global radiation

RMSE nRMSE MBE nMBE R RMSE nRMSE MBE nMBE R RMSE nRMSE MBE nMBE R

15 min 86.54 9.80 66.09 7.48 0.9522 39.88 14.16 9.24 3.39 0.4946 69.48 8.47 64.81 7.90 0.9931
30 min 86.71 9.82 65.75 7.45 0.9465 39.95 15.30 9.21 3.53 0.4856 69.34 8.77 64.15 8.11 0.9919
45 min 86.56 9.80 65.24 7.39 0.9435 40.03 15.31 9.05 3.46 0.4796 69.79 8.33 63.24 7.66 0.9915
60 min 86.36 9.78 64.25 7.27 0.9427 39.96 15.23 8.83 3.36 0.4789 68.27 7.81 61.80 7.07 0.9911
75 min 85.95 9.73 63.15 7.15 0.9454 39.76 15.07 8.54 3.24 0.4838 67.72 7.55 60.29 6.72 0.9913
90 min 85.38 9.67 61.64 6.98 0.9499 39.53 14.89 8.15 3.07 0.4934 67.72 7.55 58.38 6.56 0.9914
105 min 84.84 9.61 60.13 6.81 0.9556 39.17 14.67 7.73 2.90 0.5117 66.60 7.32 56.52 6.21 0.9922
120 min 83.75 9.48 58.30 6.60 0.9618 38.82 12.78 7.24 2.38 0.5354 65.31 7.10 54.29 5.90 0.9934
135 min 82.72 9.37 56.36 6.38 0.9672 38.46 10.08 6.74 1.77 0.5650 64.33 6.84 52.07 5.53 0.9942
150 min 81.77 9.26 54.30 6.15 0.9714 38.08 8.91 6.28 1.47 0.6008 63.36 6.70 49.74 5.26 0.9947
165 min 80.77 9.15 51.79 5.86 0.9747 37.67 8.58 5.84 1.33 0.6392 63.01 6.57 47.06 4.90 0.9945
180 min 79.51 9.00 49.65 5.62 0.9776 37.02 8.36 5.61 1.27 0.6807 61.83 6.42 44.95 4.66 0.9950

Table 3
Radiation forecasting results under partially-cloudy sky conditions.

FCST. time Beam radiation Diffuse radiation Global radiation

RMSE nRMSE MBE nMBE R RMSE nRMSE MBE nMBE R RMSE nRMSE MBE nMBE R

15 min 253.70 29.34 206.39 23.87 0.7522 137.31 21.21 86.98 13.43 0.1162 146.99 13.34 109.38 9.93 0.8942
30 min 256.34 29.64 206.29 23.85 0.7253 138.34 21.36 88.77 13.43 0.0951 149.55 13.57 106.76 9.69 0.8721
45 min 259.68 30.03 205.67 23.78 0.6942 138.61 21.41 89.79 13.87 0.0890 154.31 14.01 104.35 9.47 0.8452
60 min 261.63 30.25 203.99 23.59 0.8724 138.86 21.46 90.49 13.98 0.0997 137.60 14.30 102.06 9.26 0.8295
75 min 262.09 30.31 201.14 23.26 0.6609 138.47 21.40 90.44 13.98 0.1274 160.25 14.54 99.32 9.01 0.8225
90 min 262.19 30.32 197.53 22.84 0.6589 137.37 21.31 89.82 13.87 0.1603 163.96 14.88 96.49 8.76 0.8196
105 min 262.55 30.36 194.57 22.50 0.6652 137.10 21.17 88.75 13.71 0.2089 166.40 15.10 94.71 8.60 0.8270
120 min 259.83 30.05 189.67 21.93 0.6840 136.01 21.01 87.15 13.46 0.2607 165.76 15.04 91.48 8.30 0.8421
135 min 257.61 29.79 184.85 21.38 0.7009 134.40 20.75 85.17 13.15 0.3210 165.65 15.03 88.61 8.04 0.8550
150 min 256.29 29.64 180.01 20.82 0.7155 132.82 20.51 83.09 12.83 0.3789 166.57 15.12 85.94 7.79 0.8650
165 min 252.50 29.20 174.10 20.13 0.7364 131.17 20.25 80.80 12.48 0.4356 164.31 15.38 82.69 7.74 0.8794
180 min 248.08 28.69 166.75 19.28 0.7533 129.11 19.93 78.12 12.06 0.4879 162.78 14.77 78.38 7.11 0.8886

nRMSE value varying from 13 to 15%, and an RMSE value oscillating
between 146 and 166 Wm2. The nMBE value was lower than 10%
in all cases with MBE values lower than 110 Wm2 and an R value
higher than 0.80.
As is shown, for these sky conditions the forecast success
decreased. The constant changes which alter the solar radiation
values (1-min variations) were very difficult to model in 15-min
period situations. To offset the satellite's temporal resolution
problem somewhat, the predicted values were compared with 15-
min average radiation values. Moreover, due to the satellite's res-
olution, when Meteosat is 3 km at nadir, it does not always see
cloud breakage whereas the radiometric sensors capture a high
radiation value, creating a significant difference between measured
and predicted values.
3.4. Results under overcast-sky conditions
Under overcast-sky conditions, solar radiation produces low
values throughout the day due to the presence of clouds which can
be distributed over different atmospheric layers. Under this sky
condition, 131 days were collected with a total of 2002 predictions
for the beam radiation component; whereas for diffuse and global
radiation, 132 days were collected with a total of 2073 predictions.
Table 4 shows the forecasting results under overcast skies.
Looking at the table, under these sky conditions, the beam ra-
diation had an nRMSE value of about 42%, producing a significant
difference between that measured and predicted (with an RMSE of
approximately of 290e300 Wm2); the MBE also had high values of
between 29 and 40%, whereas the R value oscillated between 0.41
392 J. Alonso-Montesinos, F.J. Batlles / Energy 83 (2015) 387e393

Table 4
Radiation forecasting results under overcast sky conditions.

FCST. time Beam radiation Diffuse radiation Global radiation

RMSE nRMSE MBE nMBE R RMSE nRMSE MBE nMBE R RMSE nRMSE MBE nMBE R

15 min 296.51 44.00 271.33 40.26 0.6751 174.69 29.45 147.94 24.94 0.5727 146.03 15.53 118.13 12.57 0.8903
30 min 299.88 44.50 274.09 40.67 0.6529 176.69 30.18 150.93 25.78 0.5463 150.63 13.17 116.43 10.18 0.8611
45 min 300.01 44.52 272.34 40.41 0.6225 180.13 30.18 154.20 25.84 0.5057 155.18 13.54 110.45 9.64 0.8182
60 min 301.07 44.68 271.10 40.23 0.5921 181.57 30.25 155.17 25.85 0.4844 159.67 13.90 107.21 9.33 0.7873
75 min 300.73 44.63 267.35 39.67 0.5610 182.07 30.28 154.84 25.76 0.4793 162.79 14.15 102.69 8.93 0.7646
90 min 300.64 44.61 263.01 39.03 0.5299 182.66 29.68 154.13 25.04 0.4896 165.29 14.37 98.71 8.58 0.7526
105 min 299.56 42.12 257.11 36.15 0.5031 183.05 29.57 152.37 24.62 0.5073 167.95 14.56 93.95 8.15 0.7434
120 min 296.71 41.72 248.30 34.91 0.4805 182.65 29.43 150.06 24.18 0.5393 169.52 14.67 87.03 7.53 0.7395
135 min 297.13 41.77 241.64 33.97 0.4485 180.29 29.05 145.69 23.48 0.5789 173.61 15.03 84.61 7.32 0.7357
150 min 293.73 41.30 231.04 23.48 0.4364 178.03 28.67 140.93 22.70 0.6064 175.37 15.18 78.74 6.82 0.7378
165 min 292.08 41.06 221.94 31.20 0.4186 173.79 27.99 134.86 21.72 0.6518 177.12 15.33 75.88 6.57 0.7429
180 min 288.61 40.58 212.59 29.89 0.4172 169.51 27.29 128.43 20.67 0.6851 178.44 15.44 73.40 6.35 0.7480

and 0.67. In the case of diffuse radiation, the nRMSE also reached
their highest values, with an average of 28% the difference in the
estimated period is not highly significant because the RMSE value
varied by only a few Wm2; the negative nMBE values indicated
that the diffuse component was underestimated with an average
value of about 22%, while the R value was always higher than 0.47.
Lastly, for the global radiation component, the average nRMSE
value was 14%, the nMBE value decreased from 12 to 6% and the R
value was higher than 0.73 in all cases, being 0.89 for the first
15 min.
Under these kinds of atmospheric conditions, the success rate
decreases. The problem is mainly caused by the satellite vision.
Usually, when these conditions exist, different cloud layers are
present. The problem is that the satellite only sees the higher cloud
layers whereas other layers below go unseen. For the satellite, the
atmosphere is modeled from the satellite's position to the highest
cloud layer, omitting the lower cloud layers from the total attenu-
ation estimation. Consequently, the attenuation factor is obtained
with only limited information. For this reason, there is an over-
estimation for beam and global radiation, regardless of other clouds
below (normally medium and low clouds [9]) that attenuate the
solar radiation to a higher degree than the high clouds alone.
Therefore, it is assumed that in overcast skies, where the sky is
fully-covered by clouds, there are different cloud layers. In
conclusion, under these sky conditions, the radiation is predicted
according to satellite limitations, which can be solved by using
another ground-level technique as a contrast tool.
4. Conclusions
In this work, a solar radiation forecasting method is proposed
for the short- and medium-term, where beam, diffuse and global
radiations are predicted under all sky conditions, presenting the
results in a general way and classifying them into cloudless,
partially-cloudy and overcast conditions.
The ESRA model was used to estimate the three solar radiation
components under clear skies, whereas the Heliosat-2 method was
utilized to model total atmospheric attenuation.
Remote sensing techniques were applied to satellite images,
determining the cloud motion vectors for moving the attenuation
factors calculated for each pixel, thus obtaining the future position
of the atmospheric attenuation factors.
Under all sky conditions, beam radiation was predicted over
three-hours periods with an nRMSE value of about 21% and an R
value higher than 0.79. Diffuse radiation was usually under-
estimated, with nRMSE values close to 16%, whereas global radia-
tion achieved better results with an average nRMSE value of about
10% and a higher R value, close to 1.
Solar radiation predictions under cloudless sky conditions ach-
ieved the best results, where beam and global radiation presented
an nRMSE value lower than 9% and an R value close to 1 for the
three hours predicted, whereas the diffuse component presented
lower RMSE values, having an nMBE value lower than 4% in all
cases.
Generally, the rapid appearance of cloud transients in partially-
cloudy skies, are not seen by satellite images thus provoking a
decrease in the solar radiation forecasting success rate. Mainly,
these results in an average nRMSE value of about 29% for beam
radiation, 20% for diffuse radiation and 14% for global radiation.
Under overcast skies, global radiation presents better results
than the other radiation components, where the nRMSE has an
average value of about 14% and an R value higher than 0.73. Beam
and diffuse radiation reach an average nRMSE value of 42% and 28%,
respectively. The decrease in accuracy is caused by the presence of
different cloud layers. The satellite only sees the higher cloud
layers, whereas other lower cloud layers go unseen. Therefore, the
attenuation factor is modeled according to the higher clouds
regardless of others below, which, nonetheless, attenuate the solar
radiation. In this case, a ground-based radiation prediction method
can contrast the satellite's prediction for improving the results
accuracy.
The motion vectors presented a good level of accuracy, because
the statistical evaluation parameters followed the same trend
throughout the period forecast.
To conclude, the presented methodology is a useful tool for
technologies related to solar resource management, such as in CSP
or PV plants, where solar forecasting can be included in operational
decision systems for the correct adaptation of technological sys-
tems to atmospheric conditions e thus improving solar energy
management systems and the quality of electricity generation.
Acknowledgments
This project has been funded by Torresol Energy Investments,
S.A., a collaboration which we wish to acknowledge. Also, the au-
thors wish to acknowledge the project (CGL2011-30377-C02-02),
which was funded by the Ministerio de Economía y Competitividad
 n Tecnolo
and the project (13/639) financed by the Corporacio gica
de Andalucía.
References
[1] Oak N, Lawson D, Champneys A. Performance comparison of renewable
incentive schemes using optimal control. Energy 2014;64:44e57.
[2] Huanan L, Hailin M, Ming Z, Shusen G. Analysis of regional difference on
impact factors of China's energy e related CO2 emissions. Energy 2012;39:
319e26.
 J. Alonso-Montesinos, F.J. Batlles / Energy 83 (2015) 387e393
[3] Go mez A, Zubizarreta J, Dopazo C, Fueyo N. Spanish energy roadmap to 2020:
socioeconomic implications of renewable targets. Energy 2011;36:1973e85.
[4] Gonçalves C. Renewable energies: choosing the best options. Energy 2010;35:
3179e93.
[5] Tverberg GE. Oil supply limits and the continuing financial crisis. Energy
2012;37:27e34.
pez-Martínez M, Rubio FR. Cloud detection system for a solar power tower
[6] Lo
plant. In: IECON 02 [Industrial Electronics Society, IEEE 2002 28th Annual
Conference of the], 3; 2002. p. 2560e5.
[7] Sospedra F, Caselles V, Valor E, Bella CD, Coll C, Rubio E. Night-time cloud
cover estimation. Int J Remote Sens 2004;25:2193e205.
[8] Ghosh A, Pal NR, Das J. A fuzzy rule based approach to cloud cover estimation.
Remote Sens Environ 2006;100:531e49.
[9] Escrig H, Batlles FJ, Alonso J, Baena FM, Bosch JL, Salbidegoitia IB, et al. Cloud
detection, classification and motion estimation using geostationary satellite
imagery for cloud cover forecast. Energy 2013;55:853e9.
[10] Heinle A, Macke A, Srivastav A. Automatic cloud classification of whole sky
images. Atmos Meas Tech 2012;3:557e67.
[11] Martínez-Chico M, Batlles FJ, Bosch JL. Cloud classification in a mediterranean
location using radiation data and sky images. Energy 2011;36:4055e62.
[12] Alonso J, Batlles FJ, Ternero A, Lo pez G. Sky camera imagery processing based
on a sky classification using radiometric data. Energy 2014. http://dx.doi.org/
10.1016/j.energy.2014.02.035.
[13] Alonso J, Batlles FJ, Villarroel C, Ayala R, Burgaleta JI. Determination of the sun
area in sky camera images using radiometric data. Energy Convers Manage
2014b;78:24e31.
[14] Alonso J, Batlles FJ. Short and medium-term cloudiness forecasting using
remote sensing techniques and sky camera imagery. Energy 2014;73:890e7.
[15] Alonso J, Ternero A, Batlles FJ, Lo pez G, Rodríguez J, Burgaleta JI. Prediction of
cloudiness in short time periods using techniques of remote sensing and
image processing. Energy Procedia 2013;49:2280e9.
[16] Bird RE, Hulstrom RL. Direct insolation models. US Solar Energy Research
Institute Tech Report. 1980. SERI/TRe335e344, http://www.nrel.gov/docs/
legosti/old/344.pdf.
[17] Bird RE, Hulstrom RL. A simplified clear-sky model for the direct and diffuse
insolation on horizontal surfaces. US Solar Energy Research Institute Tech
Report. 1981. 38:SERI/TRe642e761, http://rredc.nrel.gov/solar/pubs/pdfs/tr-
642-761.pdf.
[18] Iqbal M. An introduction to solar radiation. Academic Press; 1983.
[19] Bird RE, Riordan CJ. Simple solar spectral model for direct and diffuse irra-
diance on horizontal and tilted planes at the Earth's surface for cloudless
atmospheres. J Clim Appl Meteorol 1986;25(1):87e97.
393
[20] Orgill JF, Hollands KGT. Correlation equation for hourly diffuse radiation on a
horizontal surface. Sol Energy 1977;19(4):357e9.
[21] Erbs DG, Klein SA, Duffie JA. Estimation of the diffuse radiation fraction for
hourly, daily and monthly-average global radiation. Sol Energy 1982;28(4):
293e302.
[22] Reindl DT, Beckman WA, Duffie JA. Diffuse fraction correlations. Sol Energy
1990;45(1):1e7.
[23] Perez R, Ineichen P, Moore K, Kmiecik M, Chain C, George R, et al. A new
operational model for satellite-derived irradiances: description and valida-
tion. Sol Energy 2002;73(5):307e17.
[24] Vignola F, Harlan P, Perez R, Kmiecik M. Analysis of satellite derived beam and
global solar radiation data. Sol Energy 2007;81:768e72.
[25] Schillings C, Meyer R, Mannstein H. Validation of a method for deriving high
resolution direct normal irradiance from satellite data and application for the
arabian peninsula. Sol Energy 2004a;76:485e97.
[26] Schillings C, Mannstein H, Meyer R. Operational method for deriving high
resolution direct normal irradiance from satellite data. Sol Energy 2004b;76:
475e84.
[27] Zarzalejo LF, Polo J, Martín L, Ramírez L, Espinar B. A new statistical approach
for deriving global solar radiation from satellite images. Sol Energy
2009;83(4):480e4.
[28] Şenkal O, Kuleli T. Estimation of solar radiation over Turkey using artificial
neural network and satellite data. Appl Energy 2009;86(7e8):1222e8.
[29] Rigollier C, Bauer O, Wald L. On the clear sky model of the ESRA e European
Solar Radiation Atlas e with respect to the Heliosat method. Sol Energy
2000;68:33e48.
 pez G. Solar resources estimation combining
[30] Bosch J, Batlles F, Zarzalejo L, Lo
digital terrain models and satellite images techniques. Renew Energy
2010;35(12):2853e61.
[31] Rusen S, Hammer A, Akinoglu BG. Estimation of daily global solar irradiation
by coupling ground measurements of bright sunshine hours to satellite im-
agery. Energy 2013;58:417e25.
[32] Tina GM, Fiore SD, Ventura C. Analysis of forecast errors for irradiance on the
horizontal plane. Energy Convers Manage 2012;64:533e40.
[33] Cano D, Monget JM, Albuisson M, Guillard H, Regas N, Wald L. A method for
the determination of the global solar radiation from meteorological satellite
data. Sol Energy 1986;37:31e9.
[34] Diabate  L, Demarcq H, Michaud-Regas N, Wald L. Estimating incident solar
radiation at the surface from images of the Earth transmitted by geostationary
satellites: the Heliosat project. Int J Sol Energy 1988;5:261e78.
[35] Rigollier C, Lefvre M, Wald L. The method Heliosat-2 for deriving shortwave
solar radiation from satellite images. Sol Energy 2004;77:159e69.