A NEW NEURAL SYSTEM FOR LOAD FORECAST IN
ELECTRICAL POWER SYSTEMS
A Topological Level Integration of Two Horizon Model Forecasting
Rodrigo Marques de Figueiredo, José Vicente Canto dos Santos and Adelmo Luis Cechin
PIPCA - UNISINOS, Av. Unisinos 950, São Leopoldo, Rio Grande do Sul, Brazil
Keywords: Artificial Neural Networks, Electric Power Systems, Load Forecasting.
Abstract: This work presents a new integrated neural model approach for two horizons of load forecasting. First of all
is presented a justification about the design of a computational neural forecasting model, explaining the
importance of the load forecast for the electrical power systems. Here is presented the design of the two
neural models, one for short and other for long term forecasting. Also is showed how these models are
integrated in the topological level. A neural model that could integrate two forecasting horizons is very
useful for electrical system enterprises. The computational system, here presented, was tested in three
different scenarios, where each scenario has specific electrical load behaviour. At last the results is
commented and explained.
1 INTRODUCTION
Actually the load forecasting is an important tool for
energy enterprises. The forecast for electrical power
systems is subject to internal variables in addition of
external variables, stochastics variables, like
meteorological and macroeconomic variables. The
first one has an imply in residential loads and the
second one has a strong imply in industrial loads
(Ardil et al, 2007). The modern way to develop a
forecaster is by the using of ANN, Artificial Neural
Network, models.
In the literature, there are many papers about the
use of neural modeling for only one forecasting
horizon, examples are the work of Botha (Botha,
1998), Drezga (Drezga, 1999), Saad (Saad, 1999),
Charytoniuk (Charytoniuk, 2000), Fukuyama (Fuku-
yama, 2002), Funabashi (Funabashi, 2002) and
Abdel-Aal (Abdel-Aal, 2004). But neural modeling
for two or more forecasting horizons is scare, one of
the few exmples is the work of Shirvany (Shirvany,
2007).
The present paper propouses a new neural model
for load forecasting by the using of two integrated
models, one for short term and other for long term
load forecasting. The resulting model has the ability
for short and long term load forecasting at the same
time, with better performance, both in response
quality and computational performance.
The electrical power system focused in this
forecast system is located in a large area in the south
of Brazil. All the tests and results showed in this
paper are referred to this system. This area is divided
in seven nodes and each node has one type of the
three electrical consuption behaviour, residential,
industrial or a mixed type. After this introduction,
follows the description of the proposed system, the
tests performed and the results obtained and, finaly,
our conclusions.
2 THE COMPUTATIONAL
FORECASTING SYSTEM
The forecasting system consists in two neural
models, one for short term and other for long term
forecasting. These neural models are given by the
artificial neural network application. The models
were individually designed and validated to later be
integrated. The data base of variables available to be
used to design the models are given by meteorolo-
gical, macroeconomic and electrical variables.
The variable space for an electrical system is too
large, even when it is reduced for the three types
showed above. For a better model response this
363
Marques de Figueiredo R., Vicente Canto dos Santos J. and Luis Cechin A. (2009).
A NEW NEURAL SYSTEM FOR LOAD FORECAST IN ELECTRICAL POWER SYSTEMS - A Topological Level Integration of Two Horizon Model
Forecasting.
In Proceedings of the 6th International Conference on Informatics in Control, Automation and Robotics - Intelligent Control Systems and Optimization,
pages 363-366
DOI: 10.5220/0002199703630366
Copyright
c
 SciTePress

space must be reduced. Variable selection methods
are the best way to reduce the variable space
removing from model most of redundant and
irrelevant variables.
2.1 Variable Selection
The variables were selected by the using of forward
selection. In this method the neural model is
constructed by its interaction, where in each interact-
tion one variable is included in model. The criteria
used to the model construction are the minor
response error for a validation (Seeger, 2003). This
algorithm runs until a stop criteria, in this paper case
an error level minor than fifteen percent. For the two
models, short and long term, this method is applied
by individually manner. In the variable selection in
addition to the inclusion of new variables were also
varied the number of neurons in the hidden layer of
ANNs, seeking for the best system performance.
2.2 Long Term Model
The main objective of this model is to provide the
behaviour information of the electrical system to
short term model, through the topological integra-
tion. In this model the forecasting horizon chosen
was the monthly horizon, because that information is
very important for the business of the electrical
energy sector utilities (Quintanilha et al, 2005).
After the forward selection application the
variables were selected, resulting in the neural
model for long term forecasting. The monthly
information of temperature and residential,
commercial and industrial electrical load as input,
with six neurons in hidden layer and one as output,
indicating the long term total load forecast. This
model uses as input the monthly information of one
year and one day ago. That information give to the
long term model the monthly tendency of each
month of the year with all seasonal influences. This
fact makes the model more robust.
2.3 Short Term Model
This model try, as main objective, mimetizes the
electrical power system load behaviour. As like long
term model, this model uses the forward selection to
choose its variables. In this model case faster varia-
bles behaviour is relevant to it, like meteorological
and electrical variables.
After the use of forward selection the neural
model was constructed with the variables selected.
This uses the daily information, about one day ago,
of temperature, humidity and total electrical load as
input, with six neurons in the hidden layer and one
as output, representing the total load for the shot
term forecast.
2.4 Model Integration
The integration of the short and long term forecast
models is the main step of the computational system
design. Is important keep in mind that this integra-
tion is given in the topological level. With this type
of integration only the tendencies of each model are
passed to the other. In other types of integration the
error also is integrated.
The neurons sharing guarantee the tendencies ex-
change between long and short term models without
polluting yours responses. But this is not a total
share, only a parcel of these neurons is shared.
Using the neural models for short and long term
forecasting with six neuron in hidden layer, a new
neural model are created with merging these
models. There were made tests to verify the number
of shared neurons in hidden layer is needed to
bettering the model response. In this test the number
of shared neurons was varied in one to all (twelve).
Figure 1: Trial with neurons sharing.
Figure 2: Neural model integration.
ICINCO 2009 - 6th International Conference on Informatics in Control, Automation and Robotics
364

The Figure 1 shows that four is the best number
of shared neurons to this application. The Figure 2
shows the final arrange of neural model, in
highlighting the shared neurons in dark color. Also
are showed the inputs and the outputs of final model.
The final model uses twelve neurons in hidden
layer, with four exclusively used by short term
model, four for the long term model and four
neurons being shared by the two models, unifying
these models in only one.
2.5 System Architecture
The architecture of the computational system is gi-
ven by three main parts, or modules. This architect-
ture is showed in Figure 3.
Figure 3: Computational system architecture.
Database contains all information about the elec-
trical power system. For forecasting models is very
important a large database as possible (Swinder et al,
2007). In the data treatment module the data is
synchronized, normalized and separated per type.
This learning occurs throughout the artificial neural
network (ANN) training. The data set is delivered to
the neural model aligned like is showed in Figure 4.
Figure 4: Data set temporal aligning.
In Figure 4 the forecast instant represents the
moment when the computational forecasting system
is executed. This data alignment avoids the need for
not available data. That case occur when two
forecast-ting horizons are used in the same model
and one horizon is overridden by the other.
3 TESTS AND RESULTS
The system proposed was subjected to three
different scenarios of load consumption being that,
Industrial Load Region, Residential Load Region
and Mixed Load Region. The tests outcomes of the
integrated system are compared with the outcomes
of the separated models for short and long term
forecasting. In the tests was used the same number
of sample for each region data set, and the same data
set to individually forecaster (short and long term)
and the integrated proposed system. There are
performed the Ten-Fold Cross Validation method to
prove the benefit of the models integration. As
quantitative metric was used was the Root Mean
Squared Error (RMSE), and all the results presented
in this section were obtained with this metric.
3.1 Industrial Load Region Test
Industrial load has a seasonal behaviour with strong
dependence of macroeconomic factors, that indicates
the production behaviour of the industry and per
consequence it is your electrical power consumption.
The proposed system and individually models,
developed to create the proposed system, results for
this scenario are showed in Table 1.
Table 1: Industrial region test results.
Forecast
Horizon
Propose
Integrated System
Individually
Models
Long Term 4,6% 21,4%
Short Term 13,2% 23,7%
3.2 Residential Load Region Test
The residential load presents a different behaviour, it
is not seasonal. This type of consumer has a
behaviour closely liked to the meteorological condi-
tions. In cold days the residential consumer uses
their heaters, and in the hot days they use their air
conditioners. The system outcome to this type of
load consumption is given in Table 2.
Table 2: Industrial region test results.
Forecast
Horizon
Propose
Integrated System
Individually
Models
Long Term 6,7% 22,9%
Short Term 13,0% 24,8%
3.3 Mixed Load Region Test
Mixed load consumer regions are areas where there
A NEW NEURAL SYSTEM FOR LOAD FORECAST IN ELECTRICAL POWER SYSTEMS - A Topological Level
Integration of Two Horizon Model Forecasting
365

is a balance between residential and industrial
consumers. In those areas there is no definition
about the load behaviour, because it follows the
trend given by the industrial and residential load.
The system outcome to the mixed type of load
consumption is given in Table 3.
Table 3: Industrial region test results.
Forecast
Horizon
Propose
Integrated System
Individually
Models
Long Term 5,5% 22,1%
Short Term 11,7% 24,6%
In Figure 5 is ploted the results for short term
forecast, comparing the pattern wait with outcomes
of conventional forecasting system and the new
neural system proposed in this paper. Note that the
proposed system (represented by solid black line)
fits perfectly with the pattern waited (grey line), the
conventional neural system, represented by the short
term model (dashed line) before developed has a
worst behaviour.
Figure 5: Short term load forecasting for mixed region.
4 CONCLUSIONS
The results show that integration of long and short
term model is beneficial to the response of the
integrated system. This integration improve the
system accuracy for both forecast horizon and also
turns the resulting model generic. That affirmation
can be proved by the close results for the tree types
of load consumption. A generic forecasting system
has a important advantage for commercial usage,
because they could forecast many instances with
only one model.
Finally, the main contribution of this work is a
new neural model for load forecasting, by the
topological level integration usage. With this
integration, the computational system has proved
flexible and capable to generating excellent results.
Some other aspects of the load forecast in electric
systems, like the expansion of the time horizon, will
be published in future works.
REFERENCES
Abdel-Aal, R. E., Short-Term Hourly Load Forecasting
Using Abductive Networks, 2004. IEEE Transactions
on Power Systems, Vol. 19, No. 1.
Ardil, C., Manjunath, T., Mehta H., Mehta, A., 2007. A
Multi-Layer Artificial Neural Network Architecture
Design for Load Forecasting in Power Systems.
International Journal of Applied Mathematics and
Computer Sciences, Vol. 4, No. 4.
Botha E. C., Vermaak, J., 1998. Recurrent Neural
Networks for Short-Term Load Forecasting. IEEE
Transactions on Power Systems, Vol. 13, No. 1.
Charytoniuk, W.; Chen, M. S.; Very Short-Term Load
Forecasting Using Artificial Neural Networks. IEEE
Transactions on Power Systems, Vol. 15, No. 1,
February 2000.
Drezga, I., Rahman, S.; Short-Term Load Forecasting with
Local ANN Predictors. IEEE Transactions on Power
Systems, Vol. 14, No. 3, August 1999.
Fukuyama,Y.; Matsui, T.; Iizaka, T.; A Novel Daily Peak
Load Forecasting Method using Analyzable Structured
Neural Network. IEEE T&D Asia, October 2002.
Funabashi, T.; Uezato, K.; Takara, H.; Senju, T.; One-
Hour-Ahead Load Forecasting Using Neural Network.
IEEE Transactions on Power Systems, Vol. 17, No. 1,
February 2002.
Quintanilha, P., Souza, A., Leme, R., Carpinteiro, O.,
2005. A Hierarchical Hybrid Neural Model with Time
Integrators in Long-Term Peak-Load Forecasting.
International Joint Conference on Neural Networks,
2005. IJCNN '05. Proceedings. 2005 IEEE. Vol. 5,
Pages: 2960- 2965.
Saad, M.; Sood, V.; Kandil, N.; Use of ANNs for Short-
Term Load Forecasting. IEEE Canadian Confrerence
on Electrical and Computer Engineering, May 1999.
Seeger, M., Williams, C., Lawrence, N., 2003. Fast
Forward Selection to Speed Up Sparse Gaussian
Process Regression. Workshop on AI and Statistics.
Shirvany, Y., Hayati, M., 2007. Artificial Neural Network
Approach for Short Term Load Forecasting for Illam
Region. International Journal of Electrical, Computer,
and Systems Engineering, Vol. 1 No. 2.
Swinder, D., Barth, R., Meibom, P., Weber, C., 2007.
Changes of System Operation Costs Due to Large-
Scale Wind Integration. Business and Policy Track:
Integrating wind in electricity markets, 2007.
ICINCO 2009 - 6th International Conference on Informatics in Control, Automation and Robotics
366