USING NEURAL NETWORKS TO FORECAST RENEWABLE
ENERGY RESOURCES
Rafael Peña
1
and Aurelio Medina
2
1
Ingeniería en Energía, Universidad de la Ciénega del Estado de Michoacán de Ocampo
Avenida Universidad 3000, Sahuayo, Mexico
2
Facultad de Ingeniería Eléctrica, Universidad Michoacana de San Nicolás de Hidalgo
Avenida Francisco J. Múgica S/N, Morelia, Mexico
Keywords: Neural networks, Forecast techniques, Time series, Renewable energy.
Abstract: This contribution presents the application of feed-forward neural networks to the problem of time series
forecasting. This forecast technique is applied to the water flow and wind speed time series. The results
obtained from the forecasting of these two renewable resources can be used to determine the power
generation capacity of micro or mini-hydraulic plants, and wind parks, respectively. The forecast values
obtained with the neural network are compared against the original time series data in order to show the
precision of this forecast technique.
1 INTRODUCTION
The study of trends and patterns of complex systems
is of great interest since the results obtained from
these studies support the decision-making process in
many activities. In particular, applications such as
electric load forecasting (Wei and Jie, 2008);
(Bunnoon et al., 2009); (Hahn et al., 2010),
economic forecasting (Fang-yuan and Feng-you,
2008), forecasting natural and physical phenomena
(Makarov et al., 2010) have been widely studied.
In this context, due to the neural networks ability
to discover patterns in nonlinear and chaotic
systems, they can be used to predict the behavior of
these systems more accurately than many current
techniques, such as Exponential Smoothing, and
Holt-Winters' methods (Gelper et al., 2009).
Neural network have shown to have great
potential for renewable resources forecasting.
Examples using neural networks in power
generation based on renewable energy like water
(Xinhua and Zhuying, 2010), wind (Chen and Lai,
2011), and solar (Ghanbarzadeh et al., 2009) can be
found in the literature.
In this paper a feed-forward neural network is
applied to know the future behavior of the water
flow and wind speed time series. Results obtained
from the water flow time series can be used to
determine if a micro or mini-hydraulic plant can be
installed, the theoretical power generation and the
technical characteristics of each electro-mechanical
component of the micro-hydraulic generation system
(Peña et al., 2009). On the other hand, with the
results obtained from the wind speed time series, the
power generation capacity of a wind park that will
have in the next days can be determined (Lange and
Focken, 2005).
The rest of this paper is organized as follows:
Section II explains the general structure of a neural
network; Section III presents a case study of the
application of the neural network to the problem of
water flow time series forecasting; Section IV shows
the results obtained in the forecast of a wind speed
time series; finally, Section V draws the main
conclusions of this research work.
2 NEURAL NETWORKS
A neural network is a computational model that is
closely based on the neuron cell structure of the
biological nervous system. Feed-forward neural
networks are composed of layers of neurons in
which the input layer of neurons is connected to the
output layer of neurons through one or more layers
of intermediate neurons, as shown in Figure 1.
401
Peña R. and Medina A..
USING NEURAL NETWORKS TO FORECAST RENEWABLE ENERGY RESOURCES.
DOI: 10.5220/0003684204010404
In Proceedings of the International Conference on Neural Computation Theory and Applications (NCTA-2011), pages 401-404
ISBN: 978-989-8425-84-3
Copyright
c
2011 SCITEPRESS (Science and Technology Publications, Lda.)
The training process of the neural network
involves adjusting the weights until a desired
input/output relationship is obtained. The majority of
adaptation learning algorithms are based on the
back-propagation algorithm (De Jesus and Hagan,
2007). Through back-propagation algorithm, the
neural network forms a mapping between inputs and
desired outputs from the training set by altering
weighted connections within the network.
The training method used in this research is the
Levenberg-Marquardt back-propagation algorithm
(LMBP). The LMBP algorithm can train a neural
network with high degree of efficiency because it
uses a combination of the back-propagation
algorithm, the gradient descent method, and the
Gauss-Newton method (Wenshang et al, 2008).
Figure 1: Neural network general scheme.
3 WATER FLOW TIME SERIES
FORECASTING
In hydraulic plants the water is the raw material.
Thus, it is important to know its behavior with the
time. In order to determine the hydroelectric
potential of the water flowing through a river or
stream, it is necessary to know the average monthly
water flow and the difference in heights to which
water can fall.
The average monthly water flow is defined as the
amount of water passing through a particular point
in each instant time. The difference in heights to
which the water flow can fall is measured from the
level at which the water enters to the canal used to
carry the water to the turbines and the level of the
water at which the water is returned to the river.
The historical water flow measurements can be
organized into a time series. Figure 2 shows the
water flow time series used in this work. This
historical data corresponds to the measurements
taken from the "The Naranjillo" hydrometric station,
in Actopan River, Veracruz, México (CONAE,
2005). The time series has 300 monthly
observations, from the period of January 1961 to
December 1985.
For this case study, a neural network containing
three layers and thirty neurons in each layer was
implemented in the C# language. 200 data from the
time series were used to train the neural network and
the last 100 data were used for the validation of the
obtained forecast.
Figure 2: Water flow time series.
The calculated forecast data obtained with the
neural network and the historic data are compared in
Figure 3. From this Figure it can be observed, that
the forecasted data matches satisfactorily the
original time series.
Figure 3: Neural network forecast data.
In order to evaluate the accuracy of the obtained
forecasts, the absolute error was calculated. The
absolute error (AE) is defined as the magnitude of
the difference between the exact value and the
approximation; it is the difference between the
historical data contained in the time series and the
forecasted data. The AE is determined by (Koller
and Friedman, 2009),
=
−
(1)
where AE is the absolute error; y
h
is the historical
data; and y
f
is the forecasted data.
E
1
E
2
E
3
O
1
O
2
O
3
E
i
O
j
S
W
ji
W
1,1
W
j
0
50
100
150
200
250
300
0 30 60 90 120 150 180 210 240 270 300
Water flow (1 × 10
3
m
3
)
Month
0
50
100
150
200
250
0 102030405060708090100
Water flow (1 × 10
3
m
3
)
Month
Historical Data Forecast Data
NCTA 2011 - International Conference on Neural Computation Theory and Applications
402
The Figure 4 shows the AE obtained for each
one of the 100 forecasted data using the neural
network. The maximum AE obtained is 95.16, the
minimum AE is 0.04, while the average AE is 12.9.
From this graph, it can be also observed that AE is
bigger when the time series presents high values,
especially around the 75 and 85 data from the time
series.
Figure 4: AE obtained with the water flow time series.
4 WIND SPEED TIME SERIES
FORECASTING
Forecasting techniques are frequently used in wind
generation systems to understand the behavior of the
wind in the days ahead. Based on these results, the
power generation capacity of wind turbines installed
in a certain area can be calculated.
This power generation capacity is used to make
technical and economic decisions, e.g., transmission
system operators are interested in know the
production capacity of a wind farm in order to
maintain the balance of power transmitted through
the network, while in a deregulated system, the wind
farm owner is interested in know the production
capacity at least 48 hours in advance to raise the
necessary strategies to compete in the energy market
(Bathurst et al., 2002).
In this case study, the time series used
correspond to wind speed measurements made by
the Comisión Federal de Electricidad (CFE) in the
hybrid wind-photovoltaic generation system, "San
Juanico”, located in Baja California Sur, México.
Figure 5 illustrates the "San Juanico" wind speed
time series; this time series contains 744
measurements recorded on an hourly basis at a
height of 33 m, in the month of March, 2000.
The neural network used in this case study
contains three layers and forty-eight neurons in each
layer. The first 624 data from the time series were
used to train the neural network and the last 120 data
were used for the validation of the obtained forecast.
The forecasts data obtained by applying the
neural network are shown in Figure 6. The
forecasted data satisfactorily matches the original
time series; however, it cannot adequately reproduce
some of the peaks taking place in the original time
series.
Figure 5: Wind speed time series.
The AE obtained from the forecasted data using
the neural network is shown in the Figure 7. The
maximum AE obtained is 6.97, the minimum AE is
4 × 10
-3
, while the average AE is 1.28.
Figure 6: Neural network model response.
Figure 7: AE obtained with the wind speed time series.
0
10
20
30
40
50
60
70
80
90
100
0 102030405060708090100
Absolute error
Data
0
2
4
6
8
10
12
14
0 62 124 186 248 310 372 434 496 558 620 682 744
Wind speed (m/s)
Data
0
2
4
6
8
10
12
14
16
0 102030405060708090100110120
Wind speed (m/s)
Data
Historical Data Forecast Data
0
1
2
3
4
5
6
7
8
0 102030405060708090100110120
Absolute error
Data
USING NEURAL NETWORKS TO FORECAST RENEWABLE ENERGY RESOURCES
403
Visually, the forecast accuracy obtained in this
case study with respect to the previous one seems to
be lower, but the average AE is bigger for the first
case study, this is due to the magnitude of measured
quantities. Also, please consider that the water flow
through rivers tends to be periodic over time, while
wind speed is not and it depends on other physical
factors.
5 CONCLUSIONS
A feed-forward neural network has been applied to
forecast the future behavior of two different sets of
time series based on the measurements of renewable
energy resources, such as water flow and wind
speed.
In the first case study, the neural network was
used to estimate the future behavior of a water flow
time series. In the second case study presented, the
application of this forecast technique to the problem
of determining the future behavior of the wind speed
at a given site has been illustrated. The obtained
results in both cases show that the neural network
adequately represents the historical data contained in
the time series.
The obtained results are of great value as they
provide insight into the generation capacity that will
have a micro or mini-hydraulic plant and wind
system in the days forecasted, respectively.
ACKNOWLEDGEMENTS
The authors want to acknowledge the Universidad
Michoacana de San Nicolás de Hidalgo (UMSNH)
through the División de Estudios de Posgrado en
Ingeniería Eléctrica, and the Universidad de la
Ciénega del Estado de Michoacán de Ocampo
(UCM) for the facilities granted to carry-out this
investigation.
REFERENCES
Bathurst, G. N., Weatherill, J., Strbac, G., 2002. Trading
wind generation in short term energy markets. IEEE
Transactions on Power Systems, vol. 17, no. 3, pp.
782-789.
Bunnoon, P., Chalermyanont, K., Limsakul C., 2010. A
Computing Model of Artificial Intelligent Approaches
to Mid-term Load Forecasting: a state-of-the-art-
survey for the researcher. International Journal of
Engineering and Technology, vol. 2, no. 1, pp. 94-100.
Chen, L., Lai, X., 2011. Comparison between ARIMA and
ANN Models Used in Short-Term Wind Speed
Forecasting. Asia-Pacific Power and Energy
Engineering Conference 2011, vol. 1, pp. 1-4.
CONAE (Comisión Nacional de Ahorro de Energía).,
2005. Current situation of the national mini-hydro and
the potential available for the states of Veracruz and
Puebla, Mexico. CONAE, México.
De Jesus, O., Hagan, M. T., 2007. Backpropagation
Algorithms for a Broad Class of Dynamic Networks.
IEEE Transactions on Neural Networks, vol. 18, no. 1,
pp. 14-27.
Fang-yuan, L., Feng-you, G., 2008. Economic forecasting
research based on artificial neural network technology.
Control and Decision Conference, vol. 1, no. 1, pp.
1151-1155.
Ghanbarzadeh, A., Noghrehabadi, A. R., Assareh, E.,
Behrang, M. A., 2009. Solar radiation forecasting
based on meteorological data using artificial neural
networks. 7th IEEE International Conference on
Industrial Informatics, vol. 1, pp. 227-231.
Gelper, S., Fried, R., Croux, C., 2009. Robust forecasting
with exponential and Holt–Winters smoothing.
Journal of Forecasting, vol. 29, no. 3, pp. 285-300.
Hahn, H., Meyer-Nieberg, S., Pickl, S., 2009. Electric load
forecasting methods: Tools for decision making.
European Journal of Operational Research, vol. 199,
no. 3, pp. 902-907.
Koller, D., Friedman, N., 2009. Probabilistic Graphical
Models, Principles and Techniques. The MIT Press,
USA.
Lange, M., Focken, U., 2005. Physical Approach to Short-
Term Wind Power Prediction. Springer, USA.
Makarov, Y. V., Etingov, P. V., Huang, Z., Ma, J.,
Chakrabarti, B. B., Subbarao, K., Loutan, C.,
Guttromson, R. T., 2010. Integration of wind
generation and load forecast uncertainties into power
grid operations. IEEE Transmission and Distribution
Conference and Exposition, vol. 1, pp. 1-8.
Peña, R., Medina, A., Anaya-Lara, O., McDonald, J. R.,
2009. Capacity Estimation of a Minihydro Plant Based
on Time Series Forecasting. ELSEVIER - Renewable
Energy, vol. 34, no. 5, pp. 1204-1209.
Wei, S., Jie, Z., 2008. Short time load forecasting based on
simulated annealing and genetic algorithm improved
SVM. CCC 2008, 27th Chinese Control Conference,
vol. 1, pp. 81-85.
Wenshang, X., Zhenbo, Y., Qingming, Y., Yanliang, S.,
Tianwen, D., 2008. Research on the methods of
improving the training speed of LMBP algorithm.
World Congress on Intelligent Control and
Automation 2008, vol. 1, pp. 5281-5286.
Xinhua, C., Zhuying, L., 2010. The application of neural
network technology in floodwater forecast. 2nd
International Conference on Networking and Digital
Society, vol. 2, pp. 419-421.
NCTA 2011 - International Conference on Neural Computation Theory and Applications
404
