Sie sind auf Seite 1von 6

Renewable Energy, Vol. 14, Nos. 1-4, pp.

179-184, 1998

~ )

1998 ElsevierScience Ltd. All rights reserved


Printed in Great Britain
P l h S0960-1481 (98) 00065-2
0960-1481/98 $19.00+0.00

Pergamon

ESTIMATION OF GLOBAL SOLAR RADIATION USING ARTIFICIAL NEURAL NETWORKS


M. Mohandes, S. Rehman, and T. O. Halawani
The Research Institute, King Fahd University of Petroleum and Minerals
Dhahran 31261, Saudi Arabia

ABSTRACT
This paper introduces a neural network technique for the estimation of global solar radiation. There are 41
radiation data collection stations spread all over the kingdom of Saudi Arabia where the radiation data and
sunshine duration information are being collected since 1971. The available data from 31 locations is used
for training the neural networks and the data from the other 10 locations is used for testing. The testing data
was not used in the modeling to give an indication of the performance of the system in unknown locations.
Results indicate the viability of this approach for spatial modeling of solar radiation.
1998 Elsevier Science Ltd. All rights reserved.
KEYWORDS
Artificial neural networks; global solar radiation; renewable energy.

~TRODUCTION
The growing population and the fast depleting reserves of fossil fuels have lead researchers in the fields of
engineering, hydrology, meteorology, economy, and industry to pursue the development and use of
renewable energy resources like solar energy. An accurate detailed long-term knowledge of the available
global solar radiation is of a prime importance for the design and development of solar energy conversion
systems. This information can be acquired by installing proper radiation measuring sensors at locations of
interests. However, many countries do not have a sufficient network of radiation data collection stations.
For such places empirical models have to be developed using data from available stations. These models
are then used to estimate radiation values at other locations of interest.
Saudi Arabia has a vast land with an average bright sunshine of 8.89 hours and an average global solar
radiation on horizontal surfaces of 5591 Whrm 2. These averages are based on the data that was collected at
41 solar radiation stations during approximately 10 years (Saudi Arabian Solar Radiation Atlas, 1983). The
geographical locations of these solar radiation stations are shown in Fig. 1. These stations have been under
the supervision of the Ministry of Agriculture and Water of Saudi Arabia since 1971. A significant portion
of the Kingdom's population lives in rural areas far from major cities. Some of these areas are not
connected with an electrical supply grid and depend on independent diesel generating sets. In those areas
there is a great potential for the use of solar based applications.

179

180

M. MOHANDES et al.
......... i'~",~.

.'..

~'AL-O, urlyyi

so~

'
SJ'#

I~te~atloaat B o ~ !

*T y a

e~a~

~q[

I " uq~r

eAI-Madinall

I H u r l l SU(flIt

~,

fief

eAr "RIY~l~h

IAI~.H~Ue (

........
~%

1 ~'/.~

25"1

Kingdom of l
Saudi Arabi~,

...,.,. ,.,, ........


8 a ~ ' ~ 8a~arshi ~ Bis~a~
~ Al*Hedl
dayl

""

~""

As-SuLayyil

"~" Al-Kvash
iSiP~ Li~,

./,'
/

Fig. 1. Solar radiation stations in the Kingdom of Saudi Arabia.


Stations used for training, * stations used for testing
Many empirical models have been developed for different geographical and meteorological conditions
(Angstrom, 1956; Swartman et al., 1971; Reddy, 1971a and 1972b; Barbaro et al., 1978; Goh, 1979; and
Ogelman et aL, 1984). Monthly average daily values of global solar radiation on horizontal and inclined
surface are reported for Abu Dhabi, United Arab Emirates (E1-Nashar, 1991), several sites in Canada
(Maure et aL, 1979 and Hay, 1979), thirteen stations in India (Mani et eL, 1973), two locations in Lesotho,
South Africa (Gopinathan, 1990 and 1991) and many others. Sabbagh et aL (1977) presented an empirical
formula obtained using the daily total solar radiation, sunshine duration, relative humidity, maximum
temperature, latitude, altitude, and the location relative to the water surface. The authors compared their
model with existing models for different locations within and outside Saudi Arabia and found better
estimates with their model.
The present study uses a neural network technique for modeling monthly mean daily values of global solar
radiation on horizontal surfaces. The neural networks utilize latitude, longitude, altitude and the sunshine
duration for the prediction of solar radiation values. The results of the system indicate a relatively good
agreement between the predicted values and the observed ones.
ARTIFICIAL NEURAL NETWORKS
An Artificial Neural Network (ANN) consists of many interconnected identical processing units called
neurons. Each neuron computes a weighted sum of its n input signals, x j , for j = 1 , 2 ..... n, and then applies
a nonlinear activation function to produce an output signal y

A typical nonlinear function is the sigmoid function, defined by

Estimation of global solar radiation

181

q~(x) - 1 + exp(- x)

(2)

The weights of the interconnections between neurons are adjusted during the training process to achieve a
desired input/output mapping. ANNs come in many different forms, some require topologies with total
interconnection among neurons and others require arrangement in layers. A multilayer feedforward
network has its neurons organized into layers with no feedback or lateral connections. Layers of neurons
other than the output layer are called hidden layers. The input layer consists of a set of sensors that only
provide input signals and do not perform any computations. Input signals propagate through the network in
a forward direction, on a layer-by-layer basis until the output layer. Figure 2 shows a typical neuron and a
two-layer feedforward network.

XI

X2

Wt

den layer
xJ,"fwi /

--

Ix//
Inputs

Fig. 2 A typical neuron and a two-layer feedforward network.


One of the major advantages of neural networks is their learning ability to perform specific tasks. Learning
is accomplished by adjusting the weights of the connections between neurons. Weights are adjusted to
allow the network to produce outputs as close as possible to the known correct answers of the training data.
During the learning phase, the network captures the underlying rule for associating the inputs with the
desired outputs. Due to the generalization capabilities of the networks, it performs similarly on data that
have not used for training.
The back-propagation algorithm (Rumelhart et al., 1986 and Haykin, 1994) is a supervised iterative
training method for multilayer feedforward networks. It uses training data consisting of P input-output
pairs of vectors that characterize the problem. A sample from the training data is chosen randomly and
provided to the inputs of the network, which computes the outputs on a layer-by-layer basis until the output
layer. The difference between the actual output of the network and the correct output that is provided in the
training data is used to adjust the weights, so that the next time that same input is provided, the network
output will be closer to the correct one. This process is repeated for all other input-output pairs in the
training data. Thus, the back-propagation algorithm minimizes an error function defined by the average of
the sum square difference between the output of each neuron in the output layer and the desired output.
The error function can be expressed as:
1

d
p

(3)

Where p is the index of the P training pairs of vectors, k is the index of elements in the output vector,

dpk

is the kth element of the pth desired pattern vector, and Opk is the kth element of the output vector when
pattern p is presented as input to the network. Minimizing the cost function represented in equation (3)

182

M. MOHANDES

et al.

results in an updating rule to adjust the weights of the connections between neurons. The weight
adjustment of the connection between neuron i in a layer m and neuronj in layer m+l can be expressed as:
Aw/, = rlSjo,
(4)
Where i is the index of units in a layer m, r1 is a small positive constant called the learning rate, o~ is the
output of unit i in the mth layer, and ~j is the delta error term back-propagated from the jth unit in layer
m+l defined by:

=[dj-oj]oj[1-oj],
gj=yj[1--yj]~6kWkj,

if neuronj is in the output layer


if neuronj is in a hidden layer

Where k is index of neurons in the layer (m+2), ahead of the layer that contains neuronj.
Choosing a small learning rate r/ leads to slow rate of convergence, and too large 77 leads to oscillation. A
simple method for increasing the rate of learning without oscillation is to include a momentum term
aAwji(n ) that determines the effect of past weight changes on the current direction of movement in the
weight space, where n is the iteration number, and a is a small positive constant. Thus the weights update
rule is:
Awj,(n + 1) = r/~o, + ~Swji(n )

(5)

The iterative process of presenting an input-output pair and updating the weights continues until the error
function reaches a pre-specified value or the weights no longer change. In that case the training phase is
done and the network is ready for testing and operation. Detailed description of the multilayer feedforward
neural networks and the backpropagation algorithm may be found in (Haykin, 1994).
THE DEVELOPED MODEL AND RESULTS
ANNs have been used in a broad range of applications including: pattem classification (Lippmann, 1987
and Bishop, 1996), function approximation, optimization, prediction and automatic control (Pham et al.,
1995). Here ANNs are used for modeling global solar radiation on horizontal surfaces in the Kingdom of
Saudi Arabia. We used the back-propagation algorithm for training several multi-layer feed-forward neural
networks to estimate the monthly mean daily values of global solar radiation. The best network consists of
4 inputs, 10 neurons in one hidden layer and one neuron in the output layer. The latitude, longitude,
altitude and sunshine duration values have been used as inputs to the network. The latitude and longitude
were provided in degrees and the altitude in meters. The sunshine duration values were provided as a ratio
of the actual values divided by the maximum possible values for each location. The output is the ratio of
monthly mean daily value of the global solar radiation divided by extraterrestrial radiation outside the
atmosphere. The data from 31 stations, 12 for each, was used for training the neural network. This
provided 372 input-output pairs for training. The training process continues until the error function defined
in equation (3) approaches a prespecified minimum value. After the training is completed, the developed
model is used for testing, where the data from the remaining 10 stations is used. Figure 1 shows the
geographical locations of these 41 stations where circles represent the stations used for training while the
stations used for testing are represented by stars. The testing data was not used in the modeling to give an
indication of the performance of the system in unknown locations. The performance of the neural network
model in addition to the observed values for the 10 locations are shown in Fig.3. The Mean Absolute
Percentage Errors (MAPE) for these locations are shown in Tablel. These results indicate the viability of
this method for global solar radiation modeling.
Table 1. Mean Absolute Percentage Errors (MAPE) for ten locations used for testing.
Station Tabuk AI-UIa Unayzah Shaqra Dawdami Yabrin Turabah Heifa Kwash Najran
MAPE 10.7
6.5
14.6
10.5
13.4
10.1
16.4
11.3
19.1
13.5

Estimation of global solar radiation

183

Real Values
Neural Networks Estimates

gl

1.0

:~ 0.8

"~ 0.6
'~ 0.4

AI-Ula

"N 0.2
0.0
m

1.0

0.8
w

" 0.6
"~ 0.4

Shaqra

Unayzah

"~ 0.2
__o 0.0

1.0

~ 0.8
~" 0.6
~ 0.4

Yabrin

"d 0.2

~
m

0.0
1.0

;~ 0.8
*" 0.6
"~ 0.4

ah

"~ 0.2
0.0

1.o

:~ 0.8
" 0.6
"~ 0.4
"~ 0.2
o 0.0

N ajran

Kwash
I

4 5 6 7 8 9 10 11 12
Months of the year

4 5 6 7 8 9 10 11 12
Months of the year

Fig. 3. Performance of Neural Networks as compared to real values of monthly mean daily
values of global solar radiation on horizontal surafce for ten cities of Saudi Arabia.

M. MOHANDES et al.

184
CONCLUSION

This paper introduced neural networks technique for modeling the spatial variation of global solar
radiation. The available data from the 41 data collection stations is divided into 31 locations for training
the neural networks and 10 locations for testing. The testing data was not used in the modeling to give an
indication of the performance of the system in unknown locations. The results on these 10 locations
indicate a relatively good agreement between the observed and predicted values.
ACKNOWLEDGMENT
The authors wish to acknowledge the support of the Research Institute of King Fahd University of
Petroleum & Minerals, Dhahran-31261, Saudi Arabia.
REFERENCES
Saudi Arabian Solar Radiation Atlas, SANCST, Riyadh, Saudi Arabia, (1983).
Angstrom, A. (1956). On the computation of global solar radiation from records of sunshine. Arkiv.
Geophysik, 3(23), 551-556.
Swartman, R. K. and O. Ogunladeg (1971). Correlation of solar radiation with common parameters in
Toronto, Canada. Solar Energy, 13, 345-347.
Reddy, S. J. (1971). An empirical method for the estimation of the total solar radiation. Solar Energy, 13,
289-290.
Reddy, S. J. (1971). An empirical method for the estimation of net radiation intensity. Solar Energy, 13,
291-292.
Barbaro, S., S. Coppolino, C. Leone, and E. Sinagra (1978). Global solar radiation in Italy. Solar Energy,
20, 431-435.
Goh, T. N. (1979). Statistical study of solar radiation information in an equatorial region (Singapore).
Solar Energy, 22, 105-111.
Ogelman, H., A. Ecevit, and E. Tasdemiroglu (1984). A new method for estimating solar radiation from
bright sunshine data. Solar Energy, 33(6), 619-625.
E1-Nashar (1991). Solar radiation characteristics in Abu Dhabi. Solar Energy, 47(1), 49-55.
Maure and N. Galanis (1979). Solar radiation data for Quebec. Solar Energy, 23, 309-314.
Hay, J. E. (1979). Calculation of monthly mean solar radiation for horizontal and inclined surfaces. Solar
Energy, 23(4), 301-307.
Mani, A. and O. Chacko (1973). Solar radiation climate oflndia. Solar Energy, 14, 139-156.
Gopinathan, K. K. (1990). Solar radiation on inclined surfaces. Solar Energy, 45(1), 19-25.
Gopinathan, K. K. (I 991). Solar radiation on variously oriented sloping surfaces. Solar Energy, 47(3), 173179.
Sabbagh, J. A., A. A. M. Sayigh, and E. M. A. E1-Salam (1977). Estimation of total solar radiation from
meteorological data. Solar Energy, 19, 307-311.
Rumelhart, D. E., J. L. McClelland, and PDP Research Group (1986). Parallel Distributed Processing.
The MIT Press.
Haykin, S. (1994). Neural networks:A comprehensivefoundation. Macmillan College Publishing.
Lippmann, R. P. (1987). An introduction to computing with neural nets. IEEEASSP Mag., 4-22.
Bishop, C. M. (1996). Neural networksfor pattern recognition. Oxford University Press.
Pham, D. T. and Liu X. (1995). Neural networks for identification, predication and control SpringerVerlag, London.

Das könnte Ihnen auch gefallen