Professional Documents
Culture Documents
Short, medium and long term load forecasting model and virtual load
forecaster based on radial basis function neural networks
Changhao Xia a,b,*, Jian Wang b,*, Karen McMenemy c
a
College of Electrical Engineering and Information Technology, China Three Gorges University, Yichang Hubei 443002, China
School of Mechanical and Aerospace Engineering, Queens University, Belfast, Northern Ireland BT9 5AH, UK
c
School of Electronics, Electrical Engineering and Computer Science, Queens University, Belfast, Northern Ireland BT9 5AH, UK
b
a r t i c l e
i n f o
Article history:
Received 26 March 2008
Received in revised form 30 September
2009
Accepted 28 January 2010
Keywords:
Electric load forecasting
Radial basis function
Neural network
Virtual instrument
a b s t r a c t
Articial neural networks (ANNs) can be easily applied to short-term load forecasting (STLF) models for
electric power distribution applications. However, they are not typically used in medium and long term
load forecasting (MLTLF) electric power models because of the difculties associated with collecting and
processing the necessary data. Virtual instrument (VI) techniques can be applied to electric power load
forecasting but this is rarely reported in the literature. In this paper, we investigate the modelling and
design of a VI for short, medium and long term load forecasting using ANNs.
Three ANN models were built for STLF of electric power. These networks were trained using historical
load data and also considering weather data which is known to have a signicant affect of the use of electric power (such as wind speed, precipitation, atmospheric pressure, temperature and humidity). In order
to do this a V-shape temperature processing model is proposed. With regards MLTLF, a model was developed using radial basis function neural networks (RBFNN). Results indicate that the forecasting model
based on the RBFNN has a high accuracy and stability. Finally, a virtual load forecaster which integrates
the VI and the RBFNN is presented.
2010 Elsevier Ltd. All rights reserved.
1. Introduction
Short-term, medium and long-term forecasting of load demand
is necessary for the correct operation of electric utilities. Forecasts
are required for proper scheduling activities, such as generation
scheduling, fuel purchasing scheduling, maintenance scheduling,
investment scheduling, and for security analysis [1].
Generally the forecasting methods found in the literature are
typically based on the following forms of mathematical analysis:
regressive analysis, exponential smoothing, time series, grey box
systems, Kalman ltering, expert systems, wavelet analysis, fuzzy
system modelling, neural network modelling, etc. Many models
for STLF [216,2124] and MLTLF [1,1720] have been proposed
in the literature. In some cases researchers have combined several
methods to develop their own hybrid method. For example in [2] a
fuzzy linear regression method is used for load forecasting weekend power usage whereas weekday loads are forecast using a general exponential smoothing method. The latest developments in
general load forecasting cited in the literature use articial intelli-
* Corresponding authors. Address: School of Mechanical and Aerospace Engineering, Queens University, Belfast, Northern Ireland BT9 5AH, UK. Tel.: +44 0 28
90974181; fax: +44 0 28 90975576.
E-mail addresses: c.xia@qub.ac.uk (C. Xia), j.wang@qub.ac.uk (J. Wang),
k.mcmenemy@ee.qub.ac.uk (K. McMenemy).
0142-0615/$ - see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijepes.2010.01.009
744
Nomenclature
n
a
p
w1
b1
w2
T1
x
La
Lf
Nh
APE
MAPE
Dw
e
I
J
temperature processed
temperature value computed by formula (7)
actual load
forecasted load
number of hours
absolute percentage error
mean absolute percentage error
weight value
error vector
identity matrix
Jacobian matrix
adaptive value
2. Network features
y purelinw2a b2
In this section, a brief background is given regarding the different types of neural networks commonly available. The backward
propagation neural network (BPNN) is an overall approximation
network. It consists of an input layer, a hidden layer and an output
layer. In the approximation, weight adjustment is done according
to the gradient descent method. The algorithm has a slow convergence speed, it easily converges to local but non-global minimum
and can result in over-training to produce uncertain results. Sigmoid neurons in the hidden layer are able to cover a larger range
of inputs, but the number of neurons is xed before training.
In RBFNN, the radial basis neurons only produce responses in a
small area. For a large area of input space it is possible to increase
the radial basis neurons to adjust the network in order to reach the
precision needed. The network scale is generally bigger than BPNN,
but it has features such as adaptive structure, the output being independent of the initial weight value, global and optimal approximation, high precision, quick convergence speed, etc. The
generalized regression neural network (GRNN) is an important alternative network of RBFNN, which is applied in function
approximation.
radbasn en
n jjw1 pjjb1
a radbasjjw1 pjjb1
The output neuron uses a linear function. The input is the output of
the hidden layer. That is,
where, pureline is a MATLAB function, w2 is the weight vector between the hidden layer and the output layer and b2 is the bias of
the output neurons. The b1 can be used to adjust the sensitivity
of the function, but normally spread density C is used in actual applications. Generally b1 = 0.8326/C. Before training, the input vector, the target vector and the spread density should be supplied.
The value C which is too large or small will result in over adaptation
or non-adaptation in function approximation. Therefore, when designing the network design it is necessary to use a trial and error
method until the optimum value for C is found.
The structure of a GRNN is similar to that of a RBFNN, with a
RBF hidden layer, but has a different linear output layer to the
RBFNN. The function of the linear output layer is given by:
y purelinnormprodw2; a
MAPE 1=N h
APE
745
Nh
T 1 jT 17j=23
T1 stands for the processed temperature data. The value 17 in formula (6) is the x value computed according to the following formula
(7), which is proposed by us, as we consider that 40 C or 6 C has
the same inuence on the load.
j40 xj j 6 xj
where La and Lf, respectively are the actual and forecasted hourly
loads. The mean absolute percentage error (MAPE) is then computed by:
T1
1
0.8
0.6
0.4
0.2
-6
10
20
30
40
T/
746
Table 1
Forecasting results of BPNN, RBFNN and GRNN models.
Hours
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
754
744.8
734.3
746.1
724.7
711.8
721.5
734.7
737.1
729.5
664.6
687.6
728
741.2
745
730.4
759.3
747
719.6
686.1
670.5
708.6
730.9
688.6
MAPE (%)
BPNN model
GRNN model
RBFNN model
APE (%)
APE (%)
APE (%)
766.06
758.95
752.33
744.35
738.56
747.76
746.29
730.54
735.51
724.94
706.46
705.61
713.78
712.32
715.98
715.67
720.62
721.40
716.29
705.79
704.71
706.54
709.35
706.48
1.60
1.90
2.45
0.24
1.91
5.05
3.44
0.57
0.21
0.63
6.30
2.62
1.95
3.90
3.90
2.02
5.09
3.43
0.46
2.87
5.10
0.29
2.94
2.60
738.65
737.63
736.71
735.54
734.62
736.05
735.84
733.12
734.09
731.84
721.97
720.56
727.96
727.18
728.96
728.83
730.63
730.86
729.09
720.89
718.68
722.09
725.14
722.01
2.03
0.96
0.32
1.41
1.36
3.40
1.98
0.21
0.41
0.32
8.63
4.79
0.01
1.89
2.15
0.21
3.78
2.16
1.31
5.07
7.18
1.90
0.78
4.85
769.50
764.51
760.41
755.59
751.93
757.66
756.77
746.21
749.88
741.30
693.92
685.08
725.33
721.78
729.71
729.16
736.56
737.51
730.26
687.20
672.38
694.65
711.76
694.14
2.06
2.64
3.56
1.27
3.76
6.44
4.89
1.57
1.73
1.62
4.41
0.37
0.37
2.62
2.05
0.17
2.99
1.27
1.48
0.16
0.28
1.97
2.62
0.80
2.56
2.38
2.13
Table 2
Results for Wilcoxon test.
The results of test at 0.05 signicance
level
p, h
Zval, ranksum
0.7650, 0
0.2990, 603
0.9097, 0
0.1134, 582
0.2977, 0
1.0413, 537
p, h
Zval, ranksum
Table 3
Load data of Hubei in 19891997.
Year
1989
1990
1991
1992
1993
1994
1995
1996
1997
342.4
381.8
405.7
414.4
452.9
528.4
547.9
583.1
640.4
747
Dw J T J lI1 JT e
10
where, e is the error vector and J is the Jacobian matrix of the network error with respect to weights. The Gauss Newton LM algorithm is a transition algorithm between the steepest descent
algorithm and the Newton algorithm. It is a common algorithm
used to solve nonlinear least square problems.
The network structure and the training parameters may greatly
affect the forecasting accuracy. When there are not enough neurons in the hidden layer, the network can not be properly trained.
When they are too many neurons in the hidden layer, the training
time will increase and other problems can be encountered such as
uncoordinated tting. More training epochs will make it possible
to enhance the accuracy of the model, but may result in a longer
training time. The choice of the expected error goal needs to match
the neuron number chosen. If the error goal is too small then more
nodes and longer training time is necessary. A trial and error method was used to determine the nal structure of the updated BP network and training parameters used are as follows: ve neurons in
the hidden layer, one output neuron, 0.0001 training error goal and
2000 epochs. A specimen set was built on the basis of data in Table
3, whilst the forecasting results are shown in Table 4.
4.2.2. MLTLF model based on RBFNN
The RBF network is based on a functional approximation theory
with many excellent features such as adaptive structure, the output being independent of the initial weight value, global and optimal approximation, high precision and quick convergence speed.
The training of a RBFNN begins from zero neurons. By inspecting
output errors, the neurons in the hidden layer are increased automatically, until the error goal is achieved or the maximum number
of neurons in hidden layer has been exceeded. The specimen set for
training is built according to the data in Table 3. After comparison
of the results using different C values, the spread density C was
chosen to be 1.5. Compared with BPNN, the RBFNN does not
encounter the over-training problem, has a high stability and accuracy, and requires less training parameters and less training time.
The forecasting results (shown in Table 4) are close to the annual
targets of the electric power plan in the Hubei province (Table 5)
and the data published by Hubei Electric Power Dispatching Centre
in 2005. For instance, the MAPE values between RBF-based forecasting values in Table 4 and the target values in Table 5 for
2000 and 2005, respectively are 7.17% and 1.94%. These results
show that the forecasting model based on RBFNN is superior to
that based on BPNN.
Notes: data in Table 5 are from ninth ve-year plan for electric
power in Hubei province (corrected edition), March in 1993.
The results show that the neural network forecasting method is
feasible for MLTLF. Even though the available data is insufcient
and only the tendency factors in the historical load data are taken
into account; a higher forecasting accuracy is achieved. As mentioned previously, if all the data representing politics, economy,
population, climate factors in recent years are collected and edited
into a specimen, then further forecasting accuracy will be achieved.
5. Design of a virtual load forecaster
LabVIEW is graphical programming software which can be used
to establish a VI. MATLAB, however, is still thought to be superior
with regards numerical analysis and processing. Therefore, MATLAB is integrated into software package LabVIEW to design a VI. Library functions are available within LabVIEW that makes it
convenient to connect to software standards such as TCP/IP, SQL
748
Table 4
Forecasting values of peak load of Hubei (104 KW).
Year
1998
1999
2000
2001
2002
2003
2004
2005
2006
BP
RBF
686.16
691.22
735.19
746.34
784.82
804.84
834.6
866.71
884.11
931.95
932.91
1000.6
980.6
1072.5
1026.8
1147.9
1071.1
1226.6
Table 5
Annual targets of electric power plan for Hubei.
Year
4
1990
1995
1997
2000
2005
2010
381.8
547.9
640.4
867
1126
1450
database, DDE, and Active X. MATLAB also has a script node to link
LabVIEW. Using this script node, MATLAB programs were connected into the block diagrams of LabVIEW. The node mode was
used to call MATLAB programs in LabVIEW, in order to implement
ANN-based computations, data processing and graphical display
for forecasting results.
In order to create the VI, it was necessary to design a user front
end panel. Seven numeric controls are used to input wind speed,
precipitation, atmospheric pressure, maximum temperature, minimum temperature, humidity and day of the week, respectively.
Three dialog controls are also used to input day, month and year.
Finally two graph indicators are used to indicate a daily load curve
and an error curve whilst one numeric indicator is used to output
the peak load. The block diagrams are created using standard
scripts. The MATLAB code used to implement the previous RBFNN
was connected to LabVIEW via a node. The real historical hourly
load data and the weather data of Yichang in September in 2005
are contained in the M le. Once the VI is created, the weather data
for the forecasting period (from ofcial weather forecasting department) can be inputted. After training the network using data from
the former 21 d and inputting the weather forecasting data of the
22nd day, the hourly load forecasting values of the 22nd day was
obtained. The output from the VI is a daily load curve for the forecast day (the 22nd day), in addition to the corresponding relative
error curve and daily peak load. The results show that the VI is able
6. Discussion
The main purpose of a forecasting VI is to obtain multiple types
of forecasting results (e.g. gure, curve and chart). The owchart of
a multi-purpose virtual load forecaster is shown in Fig. 4, where
the data for building a database includes historical daily load,
power consumption of each industry, weather data and climate
features, GDP, population, industry output values, electricity price
and so on. After the database has been established, considering
some expert advice, the main factors which have close relationships with the load in local area are used. Then by using the VI
the operator can select different forecasting periods and select data
from the database according to load features for different areas to
build up a relevant specimen set for ANN training. In this paper
only actual historical daily load and weather data such as wind
speed, precipitation, atmospheric pressure, maximum temperature, minimum temperature and humidity are considered to illustrate a STLF. Concerning MLTLF, even though the tendency factors
in the historical load data were only used, higher forecasting accuracy was achieved. The forecasting accuracy of the model is generally evaluated by error between forecasting values and actual
loads. In the illustration for STLF, MAPE for hourly load is 2.13%.
As for MLTLF, the growing trend was better estimated and an
acceptable MAPE (7.171.94%) was also obtained. If we have a
multi-purpose virtual load forecaster as mentioned above, it would
749
Data Renewal
Data Collection
Data Ordering
Results Application
N
Database
Meet
Accuracy?
Experts
Advice
Data Choice for related
Factors
Test
Data Pre-process
Specimen Set
VI
Forecasting Time
and its Related
Factors
Train
ANN
the inuence of weather factors on load into account. Since a higher temperature or a lower temperature has greater inuence on
load, a V-shape model to process temperature is proposed. Three
neural network models for STLF and two methods in implementing
MLTLF are introduced. By comparing results of short-term, medium and long term load forecasting, the result demonstrates that
forecasting model based on RBFNN is effective and has high stability. Therefore, RBFNN is more suitable to the applications in design
of load forecasting instruments. The virtual load forecaster is easy
to implement and simple to operate. In addition, the forecaster is
intuitive. Many useful curves, such as daily load curves, their corresponding relative-error curves and the value of daily peak loads
can be displayed in the virtual load forecaster.
References
[1] Ranaweera D, Karady G, Farmer R. Economic impact analysis of load
forecasting. IEEE Trans Power Syst 1997;12(3):138892.
[2] Song KB, Ha SK, Park JW, Kweon DJ, Kim KH. Hybrid load forecasting method
with analysis of temperature sensitivities. IEEE Trans Power Syst
2006;21(2):86976.
[3] Aguirre LA, Rodrigues DD, Lima ST, Martinez CB. Dynamical prediction and
pattern mapping in short-term load forecasting. Int J Electr Power Energy Syst
2008;30:7382.
[4] Truong NV, Wang L, Wong PKC. Modelling and short-term forecasting of daily
peak power demand in Victoria using two-dimensional wavelet based SDP
models. Int J Electr Power Energy Syst 2008;30:5118.
[5] Bhattacharyya SC. Short-term electric load forecasting using an articial neural
network: case of northern Vietnam. Int J Energy Res 2004;28:46372.
[6] Taylor JW, Buizza R. Neural network load forecasting with weather ensemble
predictions. IEEE Trans Power Syst 2002;17(3):62632.
[7] Nengling T, Stenzel J, Wu HX. Techniques of applying wavelet transform into
combined model for short-term load forecasting. Electr Power Syst Res
2006;76:52533.
[8] Senjyu T, Takara H, Uezato K, Funabashi T. One-hour-ahead load forecasting
using neural network. IEEE Trans Power Syst 2002;17(1):1138.
[9] Ling SH, Leung FHF, Lam HK, Tam PKS. Short-term electric load forecasting
based on a neural fuzzy network. IEEE Trans Ind Electron 2003;50(6):130516.
[10] Papalexopoulos AD, Hesterberg TC. A regression-based approach to short-term
system load forecasting. IEEE Trans Power Syst 1990;5(4):153550.
750
[11] Yang JF, Stenzel J. Short-term load forecasting with increment regression tree.
Electr Power Syst Res 2006;76:8808.
[12] Pandian SC, Duraiswamy K, Rajan CCA, Kanagaraj N. Fuzzy approach for short
term load forecasting. Electr Power Syst Res 2006;76:5418.
[13] Zhang T, Zhao DF, Zhou L, Wang XF, Xia DZ. Short-term load forecasting using
redial basis function networks and expert system. J Xian Jiaotong Univ
2001;35(4):3314.
[14] Tao WQ, Shi WQ. Short-term electric load forecasting based on radial basis
function networks and fuzzy logic. J Hefei Univ Technol 2005;28(6):6314.
[15] Al-Kandari AM, Soliman SA, El-Hawary ME. Fuzzy short-term electric load
forecasting. Int J Electr Power Energy Syst 2004;26:11122.
[16] Kandil N, Wamkeue R, Saad M, Georges S. An efcient approach for short term
load forecasting using articial neural networks. Int J Electr Power Energy Syst
2006;28:52530.
[17] Kermanshahi B, Iwamiya H. Up to year 2020 load forecasting using neural nets.
Int J Electr Power Energy Syst 2002;24:78997.
[18] Ghiassi M, Zimbra DK, Saidane H. Medium term system load forecasting with a
dynamic articial neural network model. Electr Power Syst Res
2006;76:30216.
[19] Carpinteiro OAS, Lemeb RC, Souza ACZ, Pinheiro CAM, Moreira EM. Long-term
load forecasting via a hierarchical neural model with time integrators. Electr
Power Syst Res 2007;77:3718.
[20] Al-Hamadi HM, Soliman SA. Long-term/mid-term electric load forecasting
based on short-term correlation and annual growth. Electr Power Syst Res
2005;74:35361.
[21] Dillon TS, Sestito S. Neural networks applications in power systems. Leicester
(UK): CRL Publishing Ltd.; 1996. p. 1105.
[22] Dillon TS, Sestito S, Leung S. Short term load forecasting using an adaptive
neural network. J Electr Power Energy Syst 1991;13(4):18692.
[23] Dillon TS, Phua KS. Power system load models suitable for use in generation
planning, maintenance scheduling and stochastic control. In: Proceedings IEEE
control power systems conference 1977, Texas A&M University; 1977.
[24] Dillon TS, Morsztyn K, Phua K. Short term load forecasting using adaptive
pattern recognition and self-organizing techniques. In: Proceedings of the fth
world power system computation conference (PSCC-5), September 1975,
Cambridge, paper 2.4/3, p. 115.
All in-text references underlined in blue are linked to publications on ResearchGate, letting you access and read them immediately.