Professional Documents
Culture Documents
1 Introduction
The prediction of hourly solar radiation data has important consequences in
many solar applications. Such data can be regarded as a time series and its
prediction depends on accurate modeling of the stochastic process. The compu-
tation of the conditional expectation, which is in general non-linear, requires the
knowledge of the high order distribution of the samples. Using a nite data, such
distributions can only be estimated or t into a pre-set stochastic model. Meth-
ods like AutoRegressive (AR) [1] prediction, Markov chains [2,4] and ARMA
model [3] for designing the non-linear signal predictors are examples to this ap-
proach. The neural network (NN) approach also provides a good to the problem
by utilizing the inherent adaptive nature. Since NNs can be trained to predict
results from examples, they are able to deal with nonlinear problems. Once
the training is complete, the predictor can be set to a xed value for further pre-
diction at high speed. A number of researchers have used NN for prediction of
hourly global solar radiation data. In these works, the data is treated in its raw
form as a 1-D time series, therefore the inter-day dependencies are not exploited.
This paper introduces a new and simple approach for hourly solar radiation
forecasting. First, the data are rendered in a matrix to form a 2-D image-like
F. Sandoval et al. (Eds.): IWANN 2007, LNCS 4507, pp. 749756, 2007.
c Springer-Verlag Berlin Heidelberg 2007
750 F.O. Hocaoglu, O.N. Gerek, and M. Kurban
where the rows and columns of the hourly solar radiation matrix indicate days
and hours, respectively. Such 2-D representation provides signicant insight
about the radiation pattern with time. The informational insight is apparent
from the sample surface-plots and image visualizations (in gray-scale) presented
in Figures 1 and 2.
600
500
400
Solar Radiation(W/m2)
300
200
100
0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
Hours
600
500
Solar radiation(W/m )
2
400
300
200
100
0
400
300 25
20
200 15
100 10
5
0 0
Day
Hour
350
300
250
200
150
100
50
0 5 10 15 20 25
xi,j xi,j+1
(2)
xi+1,j xi+1,j+1 =?
The linear lter coecients: a1 , a2 and a3 are optimized, and the prediction
result xi+1,j+1 is estimated as
The total error energy corresponding to the whole image prediction can be cal-
culated as:
m n
= 2ij (5)
i=1 j=1
where m and n correspond to the width and height of the image, which are, for
the solar data, 365 and 24, respectively. The lter coecients that minimize this
function can be found from the solution of the minimization derivative equation:
= = =0 (6)
a1 a2 a3
The solution to equation 6 yields the following matrix-vector equation:
R11 R12 R13 a1 r1
R21 R22 R23 a2 = r2 (7)
R13 R23 R33 a3 r3
data. There are several techniques to achieve high speed NN algorithms. Among
these techniques, heuristic techniques were developed from an analysis of the
performance of the standard steepest descent algorithm. Among the category
of fast algorithms, the methods use standard numerical optimization techniques
such as conjugate gradient, quasi-Newton, and Levenberg-Marquard. The basic
back propagation algorithm adjusts the weights in the steepest descent direction.
It turns out that, although the function decreases most rapidly along the negative
of the gradient, this does not necessarily produce the fastest convergence. In the
conjugate gradient algorithms a search is performed along conjugate directions,
which produces generally faster convergence than steepest descent directions.
Newtons method is an alternative to the conjugate gradient methods, which often
converges faster. As a drawback, the method is complex and expensive for its the
Hessian matrix calculation in feed forward neural networks. The computationally
simpler quasi-Newton methods do not require calculation of second derivatives.
Similarly, the Levenberg-Marquardt algorithm was also designed to approach
second-order training speed without having to compute the Hessian matrix [6].
Since Levenberg-Marquardt algorithm supplies faster convergence it is adopted
and used in this study.
5 Experimental Results
x 11 x 12 .x.13. . . x 1n x 11 x 12 . . . . . x 1n
.
. x 22 . x 21 x 22 .
. .
. . . .
. .
x m1 . . . x mn x m1 . . . x mn
Fig. 4. 1-D and 2-D prediction templates used for modeling the image
For the minimum RMSE linear prediction, the optimal coecients are analyt-
ically determined by solving Eq. 8. The 2-D image data is fed to the prediction
system, and error gures are obtained for each hour. The error gure for 2-D
3-tap optimum lter is given in Fig. 5.
As a second step prediction model, two neural network structures given in
Fig. 6 are applied to the data. In the rst structure, the input is treated as
1-D, and the input network elements are i th , i+1 th and i+2 th elements of the
754 F.O. Hocaoglu, O.N. Gerek, and M. Kurban
400
200
Error
200
400
600
400
300 25
20
200 15
100 10
5
Day 0 0
Hour
data, where the output is the i+3 th element for each sample in the data. In the
second structure, the proposed 2-D image matrix form is used. The inputs of the
networks are i,j th , i+1,j th and i,j+1 th elements of the 2-D data matrix and the
output is i+1,j+1 th element of the data matrix for each i and j. For each case,
1/6 of the hourly solar radiation data (2 months) is used for training.
600
500
Solar radiation(W/m2)
400
300
200
100
0
0
20
40 0
5
60 10
15
20
80 25
Day Hour
400
200
Error
0
200
400
60
25
40 20
15
20 10
5
Day 0 0
Hour
Table 1. RMSE values for proposed structures and Autocorrelation coecients be-
tween actual values and predicted values of solar radiation data
600
R = 0.976
500
400
Actual pixel values(W/m2)
300
200
100
0
0 100 200 300 400 500 600
Predicted pixel values(W/m2)
Fig. 9. Plot of actual pixel values versus predicted pixel values obtained from NN 2-D
The sigmoid function and the gradient descent algorithm with Levenberg-
Marquard modication are used during learning process with three neurons at
the hidden layer. To accelerate the speed of learning process a momentum term
is used and is updated by a fraction of the previous weight update to the current
one. After the learning phase, the network is simulated by the remaining image
data (Fig. 7) and error samples ere obtained (Fig. 8). Root Mean Square Error
(RMSE) values that are obtained from proposed optimum linear prediction l-
ters and neural networks are presented in Table I. The correlation coecients
756 F.O. Hocaoglu, O.N. Gerek, and M. Kurban
between actual data values and predicted data values are also tabulated here.
The correlation coecients are also presented as a plot of actual pixel values
versus predicted pixel values obtained from 2-D NN2 in Fig. 9.
6 Conclusion
In this work, a novel approach is proposed for hourly solar radiation forecasting.
The hourly solar radiation is interpreted and rendered as an 2-D image and
its properties are examined. It is observed that two dimensional representation
gives more insight to the solar pattern than the regular 1-D interpretation. As an
illustration, 1-D and 2-D optimal linear prediction lters with 3 coecients are
designed and compared in the sense of RMSE and correlation coecients. The
RMS energy value of the data and the prediction sequence are around 198. After
applying the prediction, the RMS value of the prediction error reduces down to
44.33 using 1-D prediction. This value also constitutes the standard deviation
of the statistical system. By using 2-D prediction, this value is reduced further
to 41.09. To emphasize the eciency of the proposed 2-D representation, two
feed forward neural network structures, one for 1-D modeling and the other for
the 2-D, are built and trained by the same data. The RMSE values are obtained
as 42.012 and 38.66 for 1-D and 2-D case, respectively. This observation also
justies the eciency of the 2-D data representation that exploits inter-day
dependencies of the solar radiation pattern. Furthermore, it is clear that the
2-D NN structure provides better prediction than the optimum linear lter. The
2-D representation has potential uses for dierent meteorological parameters
and dierent models such as surface matching, clustering based classication,
etc. Dynamical time varying behavior of the model may also be analyzed. Such
analysis can be regarded as future works of this study.
References
1. Maa, A., Adane, A.: A Two State Markovian Model of Global Irradiation Suitable
for Photovoltaic Conversion. Solar and Wind. Technology 6, 247252 (1989)
2. Amato, U., Andretta, A., Bartolli, B., Coluzzi Cuomo, B.V., Fontana, F., Serio,
C.: Markov Process and Fourier Analysis as a Tool to Describe and Simulate Solar
Irradiation. Solar Energy 37, 197201 (1986)
3. Mellit, A., Benghanem, M., Hadj Arab, A., Guessoum, A.: A Simplied Model
for Generating Sequences of Global Solar Radiation Data for Isolated Sites: Using
Articial Neural Network and a Library of Markov Transition Matrices Approach.
Solar Energy 79, 469482 (2005)
4. Aguiar, J., Collares-Perrira, M., Conde, S.P.: Simple Procedure for Generating of
Daily Radiation Values Using Library of Markov Transition Matrices. Solar En-
ergy 49, 229279 (1988)
5. Gonzalez, R.C., Woods, R.E.: Digital Image Processing, pp. 461463. Prentice-Hall,
Englewood Clis (2002)
6. Hagan, M.T., Menhaj, M.B.: Training Feedforward Networks with the Marquardt
Algorithm. IEEE Transactions On Neural Networks 5, 989993 (1994)