1

Solar Power Prediction Based on Satellite

Images and Support Vector Machine
Han Seung Jang, Student Member, IEEE, Kuk Yeol Bae, Student Member, IEEE,
Hong-Shik Park, Member, IEEE, and Dan Keun Sung, Fellow, IEEE

Abstract—Penetration of solar energy into main grid has There have been several studies on prediction technologies
gradually increased in recent years due to a growing number for solar irradiance or PV power generation [4], [5]. We can
of large-scale photovoltaic (PV) farms. The power output of categorize these studies into three major categories. The first
these PV farms may fluctuate due to a wide variability of
meteorological conditions, and, thus, we need to compensate for category includes solar irradiance or cloud index prediction
this effect in advance. In this paper, we propose a solar power technologies based on satellite images [6]–[8]. The variability
prediction model based on various satellite images and a support of solar irradiance is mainly determined by the amount and
vector machine (SVM) learning scheme. The motion vectors of motion of clouds, which can be detected using satellite images,
clouds are forecasted by utilizing satellite images of atmospheric and partly determined by the amount of aerosol [9], [10].
motion vectors (AMVs). We analyze 4 years’ historical satellite
images and utilize them to configure a large number of input Meteorological satellite images are a useful source of detection
and output data sets for the SVM learning. We compare the of cloud-motion and the overall meteorological changes. Solar
performance of the proposed SVM-based model, the conventional irradiance predictions using satellite images based on cloud
time-series model, and an artificial neural network (ANN) model motion vectors (CMVs) were proposed in [6], [11], [12].
in terms of prediction accuracy. The second category includes cloud detection technologies
Index Terms—Solar power, Irradiance, Prediction, Forecasting, for solar irradiation prediction [13]–[16]. Image processing
Photovoltaic, Satellite images, Support vector machine, Machine methodologies using total sky imagers (TSIs) to analyze cloud
learning. movement and generate short-term forecasts of solar irradiance
at the ground level were proposed in [13], [15], [17]. The
I. I NTRODUCTION third category includes solar power prediction technologies
based on machine learning schemes. Time series technologies
HE world has faced a serious depletion problem of
T natural resources and a climate change problem due to
an overuse of fossil fuels, and, thus, we need to take more al-
such as ARMA and ARIMA were used for solar power
prediction in [18], [19] and a spatial-temporal autoregressive
with exogenous input (ARX) model was also proposed in [20].
ternative energy sources, so called ‘renewable energy sources’ However, these time series based prediction technologies have
instead of fossil fuels. Penetration of renewable energy sources a limitation on non-linear parameters such as cloud movement
into main grid has gradually increased in recent years and this and meteorological variables. Hence, artificial neural network
penetration is expected to increase more rapidly until 2030. (ANN) and support vector machine (SVM) schemes have been
Focusing on the solar energy in South Korea, the penetration used in forecasting the global and horizon solar irradiance and
ratio of solar energy among all available energy sources is power generation for PV systems [21]–[23].
expected to gradually increase up to 14.1% until 2035 [1]. There have been several studies on solar irradiance pre-
Large-scale photovoltaic (PV) farms have been widely de- dictions based on the SVM scheme [22], [24]–[26]. These
ployed in the world [2]. Germany, China, and the U.S. have studies mainly utilized meteorological or climate data such
installed a number of large-scale PV farms and South Korea as historical temperature and irradiance data in the SVM-
installed about 467 MW PV plants in 2013 [1]. Due to a based prediction models. However, in this paper, we propose
growing number of large-scale PV farms, the percentage of a solar power prediction model based on various satellite
solar power in total power grid keeps increasing and the images such as atmospheric motion vector (AMV), clouds,
power output of these PV farms may fluctuate due to a and irradiance images, and an SVM learning scheme. The
wide variability of meteorological conditions [3]. Accordingly, advantage of utilizing satellite images on the solar power
large-scale PV farm operators need to obtain more accurate prediction is a macroscopic observation of the amount and
prediction on how much energy can be produced from their movement of clouds. Furthermore, the SVM learning scheme
PV farms so that they can trade solar power on the energy can achieve high performance by minimizing the prediction
market. In this situation, solar power prediction technologies error and maximizing the generalization ability of the predic-
play a very important role since depending on the accuracy of tion model. We first propose an extraction method of AMV
prediction, the amount of economic benefits from solar energy from satellite images, and then forecast the motion of clouds
is significantly different. by utilizing these AMVs. We thoroughly analyze 4 years’
historical satellite images and utilize them to configure a large
The authors are with the School of Electrical Engineering, Korea Advanced
Institute of Science and Technology (KAIST), Daejeon, Korea (e-mail: {jhans, number of input and output data sets for the SVM learning.
kybae, park1507, dksung}@kaist.ac.kr). As a result, the proposed SVM-based prediction model can
2

Wind
Wind speed
direction

1 m/s 2 m/s 5 m/s 25 m/s

Fig. 2. A standard station model for wind direction and speed.

speed of lower, middle, and upper wind fields [28]. Fig. 1

shows an exemplary satellite image of AMV. The red, green,
and blue colors indicate the lower, middle, and upper wind
fields, respectively. Fig. 2 shows a standard station model for
the wind direction and speed. There exist various AMVs in
three colors on a satellite image of Korean Peninsula and,
Fig. 1. An exemplary satellite image of AMV [27]. thus, we should accurately extract those AMVs in a novel way
of image processing. We will describe the AMV extraction
simultaneously predict the future amount of clouds and solar method later in detail.
irradiance on multiple sites of South Korea in the range of 15
to 300 minutes (intra-day). We show the prediction accuracy B. Cloud analysis images
of the proposed model, compared with that of the conventional
time-series and the ANN models. The NMSC especially provides a variety of cloud analysis
The rest of this paper is organized as follows. In Section II, images, which contain the amount of clouds, cloud shape, and
we introduce satellite images utilized in this paper. In Section cloud thickness images. The image of the amount of clouds
III, we propose an AMV extraction method. We present the shows the amount of clouds in the range of 0 ∼ 100 on each
extraction method of cloud factors from satellite images in pixel. The cloud shape image shows the shape of cloud based
Section IV. In Section V, we address the training method with on the difference in the absorption of cloud particles and cloud
SVM regression. We evaluate the performance of the proposed temperature in the range of lower cloud (1 ∼ 3), middle cloud
SVM-based prediction model in Section VI. Finally, we draw (4), upper cloud (5 ∼ 9), and mixed cloud (10) on each pixel.
conclusive remarks in Section VII. In the last, the cloud thickness image shows the thickness of
clouds in the range of 0 ∼ 100 on each pixel. These cloud
II. S ATELLITE I MAGES analysis images are very important factors for the prediction
Meteorological satellite images for North-East Asia are pro- of future condition of clouds on the target area, and are used
vided by the Korea National Meteorological Satellite Center for input variables of machine learning.
(NMSC) [27]. The NMSC launched a communication, ocean
and meteorological satellite (COMS) on June 27, 2010, which
C. Irradiance images
is the Korea’s first geostationary multi-purpose satellite and
has performed meteorological and ocean observations and Irradiance images are also provided by the NMSC. These
communications services. The major missions of the COMS images are based on the amount of light intensity reflected
include continuous monitoring of imagery and extraction of by the ground and the irradiance is measured in Watt per
meteorological products, early detection of severe weather square meter (W/m2 ) in the range of 0 ∼ 1000 W/m2 . The
phenomena, and monitoring of climate change and atmo- irradiance can be utilized together with a module temperature
spheric environment. The NMSC provides raw images, basic to forecast the photovoltaic (PV) power generation [29], [30].
images, and processed images for North-East Asia and the
Korean Peninsula. Each of satellite images for the Korean
III. P ROPOSED AMV E XTRACTION METHOD
Peninsula has 1024 × 1024 pixels and the ground resolution
of each pixel is 1720.8 m. Most of images are updated In this section, we propose an AMV extraction method to
every 15 minutes, but approximately 1 hour is delayed due gather the information of wind direction and speed. Before
to the procedure of communications and image processing. In we address the proposed method, we define some variables
this section, we introduce main satellite images used for this and make notations. Fig. 3 shows a single AMV vector. The
paper such as atmospheric motion vector (AMV) image, cloud head, body, and tail are shown at the top of the AMV wing
analysis image, and irradiance image. (red circle), at the bottom of the AMV wing (green circle),
and at the end of the AMV line (blue circle), respectively. In
A. Atmospheric Motion Vector (AMV) images addition, the middle points between the head and body, and
Atmospheric motion vector (AMV) provides the overall between the head and tail are marked with orange and purple
information of atmospheric motion and wind direction and diamonds, respectively.
 3

Head

Body

Tail

Fig. 3. A single AMV vector.

A. Procedure of AMV Extraction

The proposed AMV extraction method consists of four
steps. The first step is to separate red, green, and blue pixels
from an AMV satellite image and save their pixel coordinates
into each repository of red AMVs, green AMVs, and blue
AMVs. For better understanding, we explain the remaining
procedure with respect to the blue AMVs.
The second step is to extract each single blue AMV from Fig. 4. The target and search areas on an AMV satellite image.
the repository of blue AMVs. At the beginning, we randomly
select one pixel from the repository of blue AMVs and then
calculate the distances between the selected pixel and the other sub-repository of the single blue AMV. If the number of pixels
pixels in the repository of blue AMVs. After the completion of whose distances from the middle point (orange diamond) are
distance calculations, we store the neighboring pixel with the under a distance threshold is larger than 4 pixels, the wind
minimum distance into a sub-repository of a single blue AMV speed is counted starting from 25 m/s, otherwise 5 m/s. The
and also assign this pixel as the next selected point. Every time wind speed is finally calculated by additionally considering
when we store a new pixel into the sub-repository of the single remaining speed bars (long bar: 5 m/s, short bar: 2 m/s).
blue AMV, the center coordinate of pixels in the sub-repository
of the single blue AMV is calculated. We continue the same IV. E XTRACTION OF C LOUD I MPACT FACTORS F ROM
procedure by considering the following two constraints: 1) the S EARCH A REAS
distance between the pre-selected pixel and the next selected First of all, we specify a target area and a search area. The
pixel should be less than a threshold and 2) the next selected target area is a place where we want to predict solar power.
pixel is close under a threshold to the center coordinate. If It has a target coordinate (xtarget , ytarget ), width wtarget , and
there are no other pixels satisfying the two constraints on the height htarget . The search area is a neighboring zone of the
repository of blue AMVs, the extraction of the single blue target area and the i-th search area has a search coordinate
AMV is finished, and we complete the sub-repository of the (xsearch,i , ysearch,i ), width wsearch , and height hsearch . Depend-
single blue AMV. Continuing this step, all sub-repositories of ing on the size of the target area, the accuracy and locality of
single blue AMVs are obtained from the repository of blue the prediction may be different.
AMVs. In this section, we address how to extract cloud factors from
The third step is to find the head, body, and tail pixels from the search areas and characterize the cloud impact factors of
the sub-repositories of the single blue AMVs. In this step, each search area on the target area. Since we have three colors
we pick up the farthest two pixels from a sub-repository for of AMV, k = R, G, or B, according to the color notation and
assigning the head and tail and calculate their middle point (a the index of search area, the wind direction, speed, and the
purple diamond in Fig. 3). Except the two picked pixels, the angle of the single AMVs with color k on the i-th search area
farthest pixel from the middle point (purple diamond) in the are denoted by di,k = {dx (i, k), dy (i, k)}, Vi,k (m/s), and
sub-repository is assigned as the body of the single blue AMV. θi,k , respectively.
Then, throughout the distance comparison between the body
and the two picked pixels, the shorter-distance pixel is assigned
as the head and the longer-distance pixel is assigned as the A. Calculation of cloud impact factor
tail. At the last of the third step, a wind direction vector d = We assume that the movement of clouds approximately
(dx , dy ) is calculated from the body to tail, and the angle of the follows the wind direction and speed. However, since we can
wind direction is calculated as θ. We continue this procedure conjecture that the speed of clouds and wind are different,
for the remaining sub-repositories. we may find the optimal ratio of the cloud speed to the wind
The fourth step is to calculate the wind speed for each single speed. α∗ denotes the optimal ratio. Another assumption is
blue AMV. The total number of pixels in a sub-repository of that we do not consider a new formation and disappearance
a single blue AMV is counted and then the middle point (an of clouds. Fig. 4 shows the overall AMV image of the target
orange diamond in Fig. 3) between the head and body pixels area (the magenta box) and the search areas (the cyan boxes).
is calculated. After that, we calculate the distances between On the i-th search area with color k, a set of average AMV
the middle point (orange diamond) and each of pixels in the parameters is calculated by a number of iterations and it is
4

denoted by Ωi,k = {di,k = {dx (i, k), dy (i, k)}, V i,k , θi,k }. The input and output data may have a complex nonlinear
With the set of parameters Ωi,k , we can expect the movement relationship with each other. This complex nonlinear relation-
of the i-th search area after t minutes. The center coordinate ship can be formulated by yi = g(xi ) and, thus, our main goal
of the i-th search area is updated as follows. of machine learning (ML) is to find the most similar function
⌈ ∗ ⌉ f (xi ) ≈ g(xi ). In this paper, we will find a function f (xi )
α × V i,k × 60 × t × cos θi,k with the SVM regression.
xi,k (t) = xsearch,i + ,
ε
⌈ ∗ ⌉
α × V i,k × 60 × t × sin θi,k
yi,k (t) = ysearch,i + , A. Support Vector Machine Regression
ε
Support vector machine (SVM) is one of high-performance
where (xi,k (t), yi,k (t)) and ε (meters) are the newly moved ML schemes, which was originally proposed by Vapnik [31].
center coordinate of the i-th search area with color k after t SVM is designed to not only minimize the error, but also
minutes and the ground resolution of one pixel, respectively. maximize the separation margin among different classes. SVM
In addition, if wtarget = wsearch and htarget = hsearch , the can be applied to regression methods in supervised learning
amount of sharing between the target area and the moved [32]. In the SVM regression model, a regression function is
search area is calculated as, formulated as f (xi ) = wT · xi + b with xi , w ∈ RK and b ∈ R,
max{(wtarget − A), 0} × max{(htarget − B), 0} where wT · xi is the inner product between w and xi . Finding
ζi,k (t) = , the optimal w is the main goal of the SVM regression. The
wtarget × htarget
optimization problem for finding the optimal w is formulated
where A = |xi,k (t) − xtarget | and B = |yi,k (t) − ytarget |. as [31]
Cloud factors such as the amount of clouds and the shape of
1 ∑ M
clouds are calculated in both of the target area and the search minimize ∥w∥2 + C (ξi + ξi∗ )
areas. Ci and Si denote the average amount and shape of w 2 i=1
clouds on the i-th area, respectively. The three kinds of wind subject to yi − wT · xi − b ≤ ε + ξi , ∀i,
fields, R, G, and B, and the cloud shapes (1 ∼ 10) should
be matched in terms of altitudes. The matching rule is R = wT · xi + b − yi ≤ ε + ξi∗ , ∀i,
{1, 2}, G = {3, 4}, and B = {5, . . . , 10}, and according to the ξi , ξi∗ ≥ 0, ∀i,
matching rule, only one of the amount of sharing is activated, where (ξi , ξi∗ ), ε, and C are the slack variables, a precision
i.e., if Si = 1, ζi,R (t) is activated, but ζi,G (t) = ζi,B (t) = 0. parameter, and the constant determining the trade-off between
Lastly, we can calculate the cloud impact factor of the i-th the flatness of f and tolerance of ε. With a Lagrange function
search area on the target area after t minutes as follows: and a dual optimization∑problem, we obtain the optimal
∑ coefficient vector w∗ = i=1 (βi − βi∗ )xi , where βi and βi∗
M
γi (t) = Ci × ζi,k (t) (1)
are the Lagrange multipliers. For a non-linear SVM regression
k=R,G,B
model, we introduce a kernel method. Kernel Φ(·) expands
The total cloud impact factor on the target area after t minutes the input data vector xi with K dimensions into the input
is calculated as data vector with K + D dimensions, i.e., Φ(xi ) ∈ RK+D .
∑
I Hence, the non-linear SVM regression function is formu-
Γ(t) = γi (t), (2) lated as f (xi ) = wT · Φ(xi ) + b with Φ(xi ), w ∈ RK+D
i=1 and b ∑ ∈ R, and the optimal coefficient vector is found as
w∗ = i=1 (βi − βi∗ )Φ(xi ). In the last, the optimal non-linear
M
where I is the total number of search areas.
SVM regression function is formulated as

V. T RAINING W ITH S UPPORT V ECTOR M ACHINE ∑

M

R EGRESSION f (x) = (βi − βi∗ )Φ(xi ) · Φ(x) + b

i=1
We can train our solar power prediction model based on the ∑
M
support vector machine (SVM) regression scheme with a large = (βi − βi∗ )K⟨xi , x⟩ + b, (3)
amount of historical satellite image data. Before we address i=1
the training and test procedure, we define some parameters. We where the radial basis function (RBF) kernel, K⟨xi , x⟩ =
define a given set of N input and output data pairs as D = exp{−∥xi − x∥/2σ 2 }, is used for this paper. A method of
{(x1 , y1 ), (x2 , y2 ), . . . , (xN , yN )}, and X = {x1 , x2 , . . . , xN } finding b can be referred to [32].
and Y = {y1 , y2 , . . . , yN } denote the sets of input and output
data, respectively. Here in X , xi = [xi,1 , xi,2 , . . . , xi,K ]T
is the i-th input column vector with total K variables. B. Training Method
We use Dtraining = {(x1 , y1 ), (x2 , y2 ), . . . , (xM , yM )} which We have two objectives of prediction. The first one is to
denotes the total M input and output data for training. predict the future amount of clouds at the target area. The
The remaining (N − M ) input and output data are used other one is to predict the future irradiance at the target area.
for test, and the set of test data is denoted by Dtest = Depending on the prediction objectives, we need to configure
{(xM +1 , yM +1 ), (xM +2 , yM +2 ), . . . , (xN , yN )}. the input and output data sets in a different way for ML
 5

TABLE I
trainings. In this subsection, we address the training methods PARAMETERS AND VALUES FOR THE EXPERIMENTS
for each prediction model.
Parameters Values
In the prediction of the amount of clouds, we configure
The size of satellite image 1024 × 1024 pixels
the input data vector xC with a set of the current amount of
The ground resolution of one pixel 1720.8 m
clouds at the search areas C, and the total cloud impact factor
wsearch = hsearch pixels 1, 3, 5, 11, 21, 41
Γ(t) for the prediction horizon of t minutes. Therefore, the
[ ]T Prediction horizon, t 15 ∼ 300 minutes
input column vector is configured as xC = C, Γ(t) and the
dimension of the vector is I +1, where I is the total number of
the search areas. The amount of clouds at the target area after set to the location of Korea Advanced Institute of Science
t minutes is set to a single output value, i.e., yC = Ctarget (t). and Technology (KAIST) with a coordinate of (539,581),
The amount of clouds can be used to determine the clear-sky Daejoen, South Korea. In this paper, the MATLAB-based
index η, defined as the ratio of the irradiance Ψtarget (t) at the LIBSVM (LIBrary for Support Vector Machines) [34] is used
target area to the clear-sky irradiance Ψclear (t). The amount for developing the prediction models.
of clouds Ctarget (t) has an approximately linear relationship
with the clear-sky index η as follows [33]: A. Statistics of the amount of clouds and irradiance
Ψtarget (t) Ctarget (t) We observe both the amount of clouds and irradiance [35]
η= ≈1− (4)
Ψclear (t) 100 for 9 hours from 9:00 to 17:00 at the target area. Fig. 5
shows the monthly statistics of the amount of cloud and
Thus, the irradiance of the target area after t minutes can be
b target (t) = Ψclear (t) × (1 − C
btarget (t)/100). irradiance, respectively. We can classify the characteristics
predicted as Ψ
of the cloud amount of each month into three classes such
In the irradiance prediction, we additionally insert the
as small, medium, and large amount of clouds based on the
solar altitude angle ω, which is calculated at the specific
average amount of clouds of each month. January, March, and
date and time, and the current irradiance of the target area
October belong to the small-amount class of clouds. February,
Ψtarget (0) into the input data vector for the prediction of
April, May, September, November, and December belong to
amount of clouds xC explained above since the solar altitude
the medium-amount class of clouds. June, July, and August
angle and the current irradiance are reference data for the
belong to the large-amount class of clouds, which shows a
prediction of the irradiance value. Hence, the input column
very large amount and high variation of clouds in summer
vector
[ ]T of irradiance is configured as xΨ =
for the prediction
seasons. In the statistics of irradiance, May and December
C, Γ(t), ω, Ψtarget (0) , whose dimension is I + 3, and the
show the highest and lowest average values at the target area,
irradiance of the target area after t minutes is set to a single
respectively.
output value, i.e., yΨ = Ψtarget (t).
Fig. 6 shows the hourly statistics of the amount of clouds
and irradiance at the target area, respectively. The hourly
VI. P ERFORMANCE E VALUATION statistics of the amount of clouds shows similar features over
In this section, we first investigate the statistics of the 9 hours. The hourly statistics of irradiance shows a general
amount of clouds and irradiance in South Korea in order hourly irradiance pattern, which around 12 and 13 o’ clocks
to understand the characteristics of both meteorological data. have peak irradiance values.
Then, we evaluate the performance of the proposed SVM-
based prediction models for the amount of clouds and ir- B. Prediction for future amount of cloud
radiance in terms of prediction accuracy metrics such as
root mean square error (RMSE), mean relative error (MRE), Before making a prediction model for the amount of clouds,
and the coefficient of determination, R2 , which are defined, we need to find the optimal ratio α∗ of the cloud speed to wind
respectively, as follows: speed, which is used to compute the cloud impact factors. To
√ do this, we utilize the correlation values between the variation
∑N of the amount of clouds after t minutes and the cloud impact
i=M +1 (yi − f (xi ))
2
• RMSE = , factors, i.e., ρ(α) = corr[{Ctarget (t) − Ctarget (0)}, Γ(t, α)].
N −M
∑N The correlation value ρ(α) increases as the variation of the
1 |yi − f (xi )|
• MRE [%] = × 100, amount of clouds and the cloud impact factor show a similar
N −M i=M +1 y increasing or decreasing pattern. Fig. 7 shows that the optimal
∑N
i=M +1 (yi − f (xi )) ratio α∗ is found at the maximum correlation value.
2
• R2 = 1 − ∑ ,
N Table II summarizes the prediction results for the amount of
i=M +1 (yi − y)
2

∑N clouds with a prediction horizon of 60 minutes. The width and

where y = ( i=M +1 yi )/(N − M ). height of the target and search areas are the same as 41 pixels,
The parameters and values for the experiments are summa- each on this result. The proposed SVM-based prediction model
rized in Table I. We utilize 4 years’ (2011.04 ∼ 2015.09) for the amount of clouds outperforms the persistence model,
satellite images from the NMSC in order to configure a a non-linear auto regressive (NAR) model, and an artificial
large number of data sets for the ML training and test. 70% neural network (ANN) model. The proposed prediction model
and 30% of the data are utilized for the training and the yields the RMSE, MRE, and R2 values of 10.8661, 9.9677,
test procedures, respectively. The center of the target area is and 0.9104, respectively. Especially, the persistence-based
6

The amount of clouds

1

The amount of clouds [0∼100]

100
90 0.9

80
0.8

Correlation Value ρ(α)

70
60 0.7

50 0.6
40
30 0.5

20 0.4
10
0 0.3

1 2 3 4 5 6 7 8 9 10 11 12 0.2
Months
0.1
Irradiance
900 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
800 Ratio of the cloud speed to the wind speed: α
700
Irradiance [W/m2 ]

600 Fig. 7. Finding the optimal ratio α∗ of the cloud speed to the wind speed
500

400 TABLE II
300
P REDICTION RESULTS FOR THE AMOUNT OF CLOUDS WITH t = 60
MINUTES AND wtarget = htarget = 41 PIXELS
200

100 Schemes \ Metrics RMSE MRE (%) R2 (0∼1)

0
1 2 3 4 5 6 7 8 9 10 11 12
Persistence 13.8489 12.7296 0.8554
Months NAR 12.2343 12.3699 0.8641
ANN 12.0336 12.6681 0.8774
Fig. 5. Monthly statistics for the amount of clouds and irradiance SVM 10.8661 9.9677 0.9104

The amount of clouds

The amount of clouds [0∼100]

100
90 model. The more the scatter plot converges to the red line,
80
70
the better the prediction model performs. It indicates that a
60 larger amount of clouds (near 100) and a smaller amount of
50
clouds (near 0) are very well predicted, while it is rather hard
40
30
to predict the medium amount of clouds.
20 The distribution of prediction errors approximately follows
10
the non-standardized Student’s t-distribution [36], whose prob-
0
9 10 11 12 13 14 15 16 17
ability density function (PDF) is
Hours
( ) { ( )2 }−( ν+1
2 )
Γ ν+1 1 x − µ
900
Irradiance p(x|ν, µ, σ) = ( µ ) √ 2
1+ ,
Γ νπσ 2 ν σ
800 2
700
where Γ(·), µ, σ, and ν denote the gamma function, the loca-
Irradiance [W/m2 ]

600
tion parameter, the scale parameter, and the shape parameter,
500

400
respectively. The mean and variance of the distribution are µ
300
and σ 2 ν/(ν−2). The µ, σ, and ν values of prediction errors for
200 the amount of clouds with a prediction horizon of 60 minutes
100 are 0.2842, 3.6460, and 1.25532, respectively.
0 Table III summarizes the prediction results of the amount of
9 10 11 12 13 14 15 16 17
Hours clouds based on the proposed SVM-based model with various
prediction time horizons. It is well known that the suitable
Fig. 6. Hourly statistics for the amount of clouds and irradiance forecast time horizon based on satellite images is up to 5
hours [4]. The very-short-term (15 and 30 minutes horizon)
prediction model yields the lowest prediction accuracy for all and short-term (60 minutes horizon) prediction for the amount
metrics. of clouds show high prediction accuracy with R2 = 0.9745,
Fig. 8 compares the predicted amount of clouds of the 0.9444, and 0.9104, while the prediction accuracy of 90 ∼ 300
proposed SVM-based model with the measured values. Even minutes horizons degrades to R2 = 0.8261 ∼ 0.5757, which
though some prediction errors occur, the prediction results of comes from the fact that new formation and disappearance of
the proposed model generally agree with the measured values. clouds may occur and they are not considered in the proposed
prediction model.
Fig. 9 shows the scatter plot of the measured and predicted Another issue is locality in the prediction. Table IV sum-
amount of clouds for the proposed SVM-based prediction marizes the prediction results of the amount of clouds based
 7

TABLE V
P REDICTION RESULTS FOR THE IRRADIANCE WITH t = 60 MINUTES AND
100
wtarget = htarget = 41 PIXELS

Schemes \ Metrics RMSE (W/m2 ) MRE (%) R2 (0∼1)

Amount of Clouds [0 ∼ 100]

80
Persistence 57.5704 11.0767 0.8890
NAR 74.8116 15.1280 0.8172
60
ANN 65.5209 12.1202 0.8564
40
SVM 44.1390 7.7329 0.9420

20

Measured data
0 Predicted data
on the proposed SVM-based model for various radii of the
target area and t = 60 minutes. The radii of the target area,
0 10 20 30 40 50 60 70 80 90 100
Test samples 35 km, 18 km, 9.4 km, 4.3 km, 2.6 km, and 0.9 km correspond
Fig. 8. The measured and predicted amount of clouds based on the proposed to the width or height pixel sizes of 41, 21, 11, 5, 3, and 1
SVM-based model with t = 60 minutes respectively. The prediction accuracy slightly decreases as the
size of target area decreases, which implies that the amount
100
of clouds depend on the specific target area and its size.
Predicted Amount of Clouds [0 ∼ 100]

90

80

70
C. Prediction for future irradiance
60
Table V summarizes the prediction results for the irradiance.
50 The proposed SVM-based prediction model yields the best
40 performance among the prediction models in terms of RMSE,
30 MRE, and R2 values.
20 The RMSE, MRE, and R2 values are obtained as 44.1390,
10 7.7329, and 0.9420, respectively, when t = 60 minutes and
0
wsearch = hsearch = 41 pixels. Especially, the second best
0 20 40 60 80 100
Measured Amount of Clouds [0 ∼ 100]
prediction model is the persistence-based model, which was
the worst model in the prediction for the amount of clouds.
Fig. 9. Scatter plot of the measured and predicted amount of clouds based
on the proposed SVM-based model with t = 60 minutes Figs. 10 and 11 show the measured and predicted irradiance
values and the scatter plot of the measured and predicted
TABLE III
P REDICTION RESULTS FORTHE AMOUNT OF CLOUDS FOR VARIOUS
irradiance, respectively, based on the proposed SVM-based
PREDICTION TIME HORIZONS WITH wtarget = htarget = 41 PIXELS prediction model. Even though some prediction errors occur,
the prediction results of the proposed model generally agree
Horizons \ Metrics RMSE MRE (%) R2 (0∼1)
with the measured values, and thus, the scatter plot well
15 minutes 5.7367 5.2825 0.9745
converges to the red line, which represents the high prediction
30 minutes 8.4336 7.1020 0.9444
60 minutes 10.8661 9.9677 0.9104
accuracy of the proposed SVM-based prediction model for
90 minutes 14.9205 13.4699 0.8261 future irradiance.
120 minutes 16.8566 17.1327 0.7977 The distribution of prediction errors approximately follows
150 minutes 17.5458 17.8165 0.7646 the non-standardized Student’s t-distribution, and the µ, σ,
180 minutes 19.5643 18.1852 0.7021 and ν values of prediction errors for the irradiance with a
240 minutes 23.7025 25.0641 0.6049 prediction horizon of 60 minutes are -1.3945, 34.9578, and
300 minutes 24.7855 26.6654 0.5757 3.9295, respectively.
Table VI and Table VII summarize the prediction results of
TABLE IV irradiance based on the proposed SVM-based model for vary-
P REDICTION RESULTS FOR THE AMOUNT OF CLOUDS WITH VARIOUS
RADII OF THE TARGET AREA AND t = 60 MINUTES
ing the prediction time horizons and radii of the target area,
respectively. The irradiance of 15 ∼ 150 minutes horizons
Radius \ Metrics RMSE MRE (%) R2 (0∼1)
can be well predicted with R2 = 0.9770 ∼ 0.8641, while the
35 km 10.8661 9.9677 0.9104
prediction accuracy degrades from the 180 minutes horizon to
18 km 12.3090 10.8526 0.8880
9.4 km 13.4196 12.1586 0.8728
R2 = 0.8338. In addition, the irradiance prediction of t = 60
4.3 km 14.4074 13.1201 0.8487
minutes shows significant prediction accuracy of R2 = 0.9420,
2.6 km 15.0402 13.7125 0.8435 0.9273, and 0.9042 in the range of 35 km ∼ 9.4km radii of
0.9 km 17.4361 15.1909 0.7952 the target area in Table VII. However, the prediction accuracy
for irradiance on smaller target areas (4.3 km ∼ 0.9 km radii)
slightly decreases to R2 = 0.8596 ∼ 0.8277.
8

900 VII. C ONCLUSION

Measured data
800 Predicted data We proposed a solar power prediction model based on a
Irradiance [ 0 ∼ 1000 W/m2 ]

700
large amount of various historical satellite images and an SVM
learning scheme. We thoroughly analyzed 4 years’ historical
600
satellite images and utilized them to configure a large number
500 of input and output data sets for the SVM learning. The
400 proposed SVM-based prediction model can simultaneously
predict both the future amount of clouds and solar irradiance
300
on multiple sites of South Korea in the range of 15 to 300
200
minutes (intraday). Throughout the experiments, the proposed
100 SVM-based prediction model shows the highest prediction
0
accuracy, compared to other prediction models such as the
0 10 20 30 40 50 60 70 80 90 100
Test samples conventional time-series and the ANN models. As a result,
these prediction data can be effectively utilized for grid
Fig. 10. The measured and predicted irradiance based on the proposed SVM-
based model with t = 60 minutes operation (load following) and energy management system
(EMS) in smart grid.
Predicted Irradiance [ 0 ∼ 1000 W/m2 ]

700 ACKNOWLEDGMENT
600 The authors would like to thank the Korea National Me-
500
teorological Satellite Center for providing the high quality
satellite images.
400

300 R EFERENCES
200 [1] “Renewable energy statistics 2013,” Korea Ministry of Trade, Industry
and Energy, November, 2014, http://www.motie.go.kr.
100 [2] “List of photovoltaic power stations larger than 50 megawatts (MW),”
http://en.wikipedia.org/wiki/List of photovoltaic power stations.
0 [3] J. Marcos, L. Marroyo, E. Lorenzo, D. Alvira, and E. Izco, “Power
0 100 200 300 400 500 600 700 output fluctuations in large scale PV plants: one year observations
Measured Irradiance [ 0 ∼ 1000 W/m2 ] with one second resolution and a derived analytic model,” Progress in
Fig. 11. Scatter plot of the measured and predicted irradiance based on the Photovoltaics: Research and Applications, vol. 19, no. 2, pp. 218–227,
proposed SVM-based model with t = 60 minutes 2011.
[4] J. Kleissl, Solar energy forecasting and resource assessment. Academic
TABLE VI Press, 2013.
P REDICTION RESULTS FOR IRRADIANCE WITH VARIOUS PREDICTION TIME [5] S. Pelland, J. Remund, J. Kleissl, T. Oozeki, and K. De Brabandere,
HORIZONS AND wtarget = htarget = 41 PIXELS “Photovoltaic and solar forecasting: state of the art,” IEA PVPS, Task,
vol. 14, 2013.
Horizons \ Metrics RMSE MRE (%) R2 (0∼1)
[6] A. Hammer, D. Heinemann, E. Lorenz, and B. Lückehe, “Short-term
15 minutes 28.2970 4.6743 0.9792 forecasting of solar radiation: a statistical approach using satellite data,”
30 minutes 33.9877 5.6679 0.9670 Solar Energy, vol. 67, no. 1, pp. 139–150, 1999.
60 minutes 44.1390 7.7329 0.9420 [7] A. Hammer, D. Heinemann, C. Hoyer, R. Kuhlemann, E. Lorenz,
R. Müller, and H. G. Beyer, “Solar energy assessment using remote
90 minutes 56.0455 9.3932 0.9155 sensing technologies,” Remote Sensing of Environment, vol. 86, no. 3,
120 minutes 62.3437 10.9566 0.8830 pp. 423–432, 2003.
150 minutes 70.2559 12.0128 0.8641 [8] Z. Peng, S. Yoo, D. Yu, and D. Huang, “Solar irradiance forecast
180 minutes 74.7676 13.9665 0.8338 system based on geostationary satellite,” in proc. of IEEE International
Conference on Smart Grid Communications (SmartGridComm), 2013,
240 minutes 76.2283 16.2605 0.7906 pp. 708–713.
300 minutes 94.8836 17.9500 0.7309 [9] H. Breitkreuz, M. Schroedter-Homscheidt, T. Holzer-Popp, and S. Dech,
“Short-range direct and diffuse irradiance forecasts for solar energy
applications based on aerosol chemical transport and numerical weather
TABLE VII modeling,” Journal of Applied Meteorology and Climatology, vol. 48,
P REDICTION RESULTS FOR IRRADIANCE WITH VARIOUS RADII OF THE no. 9, pp. 1766–1779, 2009.
TARGET AREA AND t = 60 MINUTES
[10] Y. Kaufman, D. Tanré, H. Gordon, T. Nakajima, J. Lenoble, R. Frouin,
Radius \ Metrics RMSE MRE (%) R2 (0∼1) H. Grassl, B. Herman, M. King, and P. Teillet, “Passive remote sensing
of tropospheric aerosol and atmospheric correction for the aerosol
35 km 44.1390 7.7329 0.9420 effect,” Journal of Geophysical Research: Atmospheres (1984-2012),
18 km 50.2472 8.4920 0.9273 vol. 102, no. D14, pp. 16 815–16 830, 1997.
9.4 km 56.0706 9.3361 0.9042 [11] H. G. Beyer, C. Costanzo, and D. Heinemann, “Modifications of the
4.3 km 69.9710 12.4461 0.8596 heliosat procedure for irradiance estimates from satellite images,” Solar
Energy, vol. 56, no. 3, pp. 207–212, 1996.
2.6 km 74.7278 13.1583 0.8505 [12] E. Lorenz, A. Hammer, and D. Heinemann, “Short term forecasting of
0.9 km 82.3364 14.2098 0.8277 solar radiation based on satellite data,” in EUROSUN2004 (ISES Europe
Solar Congress), 2004, pp. 841–848.
[13] R. Marquez and C. F. Coimbra, “Intra-hour DNI forecasting based on
cloud tracking image analysis,” Solar Energy, vol. 91, pp. 327–336,
2013.
 9

[14] J. G. Silva Fonseca, T. Oozeki, T. Takashima, G. Koshimizu, Y. Uchida, Han Seung Jang received the B.S degree in Elec-
and K. Ogimoto, “Use of support vector regression and numerically tronics and Computer Engineering from Chonnam
predicted cloudiness to forecast power output of a photovoltaic power National University, Gwangju, Korea, in 2012, and
plant in kitakyushu, japan,” Progress in Photovoltaics: Research and the M.S. degree in Electrical Engineering from Ko-
Applications, vol. 20, no. 7, pp. 874–882, 2012. rea Advanced Institute for Science and Technology
[15] H. Huang, J. Xu, Z. Peng, S. Yoo, D. Yu, D. Huang, and H. Qin, (KAIST), Daejoen, Korea in 2014. Since Mar. 2014,
“Cloud motion estimation for short term solar irradiation prediction,” in he has been a Ph.D. student in the School of
proc. of IEEE International Conference on Smart Grid Communications Electrical Engineering, KAIST. His research inter-
(SmartGridComm), 2013, pp. 696–701. ests include renewable energy forecasting, electric
[16] T. M. Hamill and T. Nehrkorn, “A short-term cloud forecast scheme vehicle (EV) charging management, and machine-
using cross correlations,” Weather and Forecasting, vol. 8, no. 4, pp. to-machine (M2M) communications.
401–411, 1993.
[17] C. W. Chow, B. Urquhart, M. Lave, A. Dominguez, J. Kleissl, J. Shields,
and B. Washom, “Intra-hour forecasting with a total sky imager at the
UC San Diego solar energy testbed,” Solar Energy, vol. 85, no. 11, pp.
2881–2893, 2011. Kuk Yeol Bae received the B.S. degree in Elec-
[18] L. Mora-Lopez and M. Sidrach-de Cardona, “Multiplicative ARMA tronics Engineering from Ajou University, Suwon,
models to generate hourly series of global irradiation,” Solar Energy, Korea, in 2010, and the M.S. degree in Electri-
vol. 63, no. 5, pp. 283–291, 1998. cal Engineering from Korea Advanced Institute for
[19] W. Ji and K. C. Chee, “Prediction of hourly solar radiation using a novel Science and Technology (KAIST), Daejoen, Korea
hybrid model of ARMA and TDNN,” Solar Energy, vol. 85, no. 5, pp. in 2012. Since Feb. 2012, he has been a Ph.D.
808–817, 2011. student in the School of Electrical Engineering,
[20] C. Yang, A. Thatte, and L. Xie, “Multitime-scale data-driven spatio- KAIST. His research interests include renewable
temporal forecast of photovoltaic generation,” IEEE Transactions on energy forecasting, energy management systems,
Sustainable Energy, vol. 6, no. 1, pp. 104–112, Jan 2015. electricity markets, and machine-to-machine (M2M)
[21] R. Marquez and C. F. Coimbra, “Forecasting of global and direct solar communications.
irradiance using stochastic learning methods, ground experiments and
the NWS database,” Solar Energy, vol. 85, no. 5, pp. 746–756, 2011.
[22] J. Shi, W.-J. Lee, Y. Liu, Y. Yang, and P. Wang, “Forecasting power
output of photovoltaic systems based on weather classification and
support vector machines,” IEEE Transactions on Industry Applications, Hong-Shik Park received the B.S. degree from
vol. 48, no. 3, pp. 1064–1069, 2012. Seoul National University, Seoul, Korea in 1977, and
[23] H.-T. Yang, C.-M. Huang, Y.-C. Huang, and Y.-S. Huang, “A weather- the M.S. and Ph.D. degrees from Korea Advanced
based hybrid method for 1-day ahead hourly forecasting of PV power Institute of Science and Technology (KAIST), Dae-
output,” IEEE Transactions on Sustainable Energy, vol. 5, no. 3, pp. jeon, Korea all in Electrical Engineering in 1986 and
917–926, 2014. 1995, respectively. In 1977 he joined Electronics and
[24] J.-L. Chen, H.-B. Liu, W. Wu, and D.-T. Xie, “Estimation of monthly Telecommunications Research Institute (ETRI) and
solar radiation from measured temperatures using support vector had been engaged in development of the TDX digital
machines–a case study,” Renewable Energy, vol. 36, no. 1, pp. 413– switching system family including TDX-1, TDX-1A,
420, 2011. TDX-1B, TDX-10, and ATM switching systems. In
[25] Z. Ramedani, M. Omid, A. Keyhani, S. Shamshirband, and B. Khosh- 1998 he moved to Information and Communications
nevisan, “Potential of radial basis function based support vector regres- Univ., Daejeon, Korea as a faculty. Currently he is a professor of the School of
sion for global solar radiation prediction,” Renewable and Sustainable Electrical and Electronics Engineering, KAIST, Daejeon, Korea. From 2004
Energy Reviews, vol. 39, pp. 1005–1011, 2014. to 2012, he was a Director of BcN Engineering Research Center sponsored
[26] B. B. Ekici, “A least squares support vector machine model for predic- by KEIT, Korea. His research interests are network architecture and protocols,
tion of the next day solar insolation for effective use of pv systems,” traffic engineering, and performance analysis of telecommunication systems.
Measurement, vol. 50, pp. 255–262, 2014. He is a member of the IEEE, IEEK and KICS, Korea.
[27] “Korea National Meteorological Satellite Center,”
http://nmsc.kma.go.kr/.
[28] K. Holmlund, “The utilization of statistical properties of satellite-derived
atmospheric motion vectors to derive quality indicators,” Weather and
Dan Keun Sung received the B.S. degree in Elec-
Forecasting, vol. 13, no. 4, pp. 1093–1104, 1998.
tronics Engineering from Seoul National University
[29] C. Paoli, C. Voyant, M. Muselli, and M.-L. Nivet, “Forecasting of
in 1975 and the M.S. and Ph.D. degrees in Electrical
preprocessed daily solar radiation time series using neural networks,”
and Computer Engineering from the University of
Solar Energy, vol. 84, no. 12, pp. 2146–2160, 2010.
Texas at Austin, in 1982 and 1986, respectively.
[30] Y. Riffonneau, S. Bacha, F. Barruel, and S. Ploix, “Optimal power
Since 1986, he has been with the faculty of the
flow management for grid connected PV systems with batteries,” IEEE
Korea Advanced Institute of Science and Technology
Transactions on Sustainable Energy, vol. 2, no. 3, pp. 309–320, 2011.
(KAIST), Daejon, Korea, where he is currently a
[31] V. Vapnik, The Nature of Statistical Learning Theory. Springer, 2000.
Professor with the School of Electrical Engineer-
[32] A. J. Smola and B. Schölkopf, “A tutorial on support vector regression,”
ing. From 1996 to 1999, he was the Director of
Statistics and computing, vol. 14, no. 3, pp. 199–222, 2004. 

the Satellite Technology Research Center (SaTReC),

[33] D. Dumortier, “Modelling global and diffuse horizontal irradiances
KAIST. He had served as Division Editor of the Journal of Communications
under cloudless skies with different turbidities,” Daylight II, jou2-ct92-
and Networks from 1998-2007. He also had served as Editor of IEEE
0144, final report vol, vol. 2, 1995.
Communications Magazine from 2002-2011. His research interests include
[34] C.-C. Chang and C.-J. Lin, “LIBSVM: a library for support vector
mobile communication systems and networks, with special interest in resource
machines,” ACM Transactions on Intelligent Systems and Technology
management, smart grid communication networks, energy networks, machine-
(TIST), vol. 2, no. 3, p. 27, 2011.
to-machine communications, WLANs, WPANs, traffic control in wireless &
[35] H. S. Jang, K. Y. Bae, H.-S. Park, and D. K. Sung, “Effect of aggregation
wired networks, performance and reliability of communication systems, and
for multi-site photovoltaic (PV) farms,” in proc. of IEEE International
microsatellites. Dr. Sung is a member of the National Academy of Engineering
Conference on Smart Grid Communications (SmartGridComm), 2015,
of Korea. He was the recipient of the 1992 National Order of Merits, the
pp. –.
Dongbaek Medal for successfully developing, launching, and operating the
[36] S. Jackman, Bayesian analysis for the social sciences. John Wiley &
first Korean satellite in Korean history, the 1997 Research Achievement
Sons, 2009, vol. 846.
Award, the 2000 Academic Excellence Award, the 2004 Scientist of the Month
from the Ministry of Science and Technology and the Korea Science and
Engineering Foundation, and the 2013 Haedong Academic Grand Award from
the Korean Institute of Communications and Information Sciences.