Electrical Power and Energy Systems 32 (2010) 743750

Contents lists available at ScienceDirect

Electrical Power and Energy Systems

journal homepage: www.elsevier.com/locate/ijepes

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

gent-based forecasting (AIBF) techniques, such as ANN [1

4,6,8,13,14,1619], fuzzy logic [2,12,14,15], and genetic algorithms
[9], all of which show promising results for STLF. Less literature is
available regarding the use of AIBF for MLTLF, especially when they
utilise ANN [1719].
Electric load demand is often considered as a function of weather variables and human social activities.
Traditional econometric approaches establish functional relationships between weather variables and current load demand
for the forecasting function, often assuming a linear relationship.
However, as Park et al. (1991) indicate, the econometric approach
may not give sufciently accurate results because of nonlinear and
non-stationary relationships between the load data and weather
variables. Therefore, an adaptable technique is needed [5]. An
ANN can model any complicated nonlinear relationship and since
STLF is a nonlinear problem the ANN forecasting method, which
combines nonlinear and time series forecasting methods, is widely
applied to STLF. However, it is rarely used in MLTLF since the variations within short-term load forecasting can be considered as a
stable random process. The variations within medium to long term
load forecasting are not usually random and can be attributed to
important factors such as governance within a given country.
Therefore, accurate MLTLF is a much more difcult problem
because it is difcult to describe the forecasting pattern with obvious formula because of the different and various factors which

744

C. Xia et al. / Electrical Power and Energy Systems 32 (2010) 743750

Nomenclature
n
a
p
w1
b1
w2

input of Gauss function

output of Gauss function
input vector of neural network
weight vector between the input and the hidden layer
bias of neurons in hidden layer
weight vector between the hidden layer and the output
layer
b2
bias of neurons in output layer
radbas
radial basis function
purelin linear function
normprod a normalized dot product weight function
C
spread density of radial basis function
T
highest temperature and the lowest temperature in a
day (both given by T),

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

can inuence it. However, theoretically MLTLF is also a nonlinear

problem for which the ANN approach could be utilized to identify
any complicated nonlinear relations, assuming enough data can be
collected for training purposes. An ANN model for MLTLF will be
developed in this paper.
Recently VI has emerged as a means of load forecasting. The efforts of this paper will concentrate mainly on establishing an accurate, stable ANN model suitable for use in VI, and also the design of
a VI for load forecasting. The idea is to combine the numerical potentials of MATLAB with the graphical interface of LabVIEW to present an ANN numerical intensive model with a user friendly
interface easy to operate.
The paper is organized as follows: Section 2 presents the structures and features of the ANN used in this research. The pre-processing of the weather data applied to the STLF and the
determination of training parameters used is outlined in Section
3. A comparison of forecasting results is also documented in Section 3. Section 4 then presents data collection, modelling considerations and forecasting results of the MLTLF based on the
RBFNN. The design process of a VI based on the RBFNN model is introduced in Section 5. Section 6 presents a scheme of a multi-purpose electric load forecaster which integrates database, ANN and VI
technology. Finally, the conclusion is given in Section 7.

where radbas is a MATLAB function. When the input of the Gauss

function n is near the central area, the hidden nodes will produce
a larger output. The model of a RBFNN can be expressed by Eqs.
(2)(4).
The input n is the distance between weight vector w1 and input
vector p, multiplied by the bias b1 of neurons in output layer.
When the distance is decreased, the output then increases. Therefore, the network is capable of local approximation.

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.

RBFNN also consists of input layer, hidden and output layers.

The input layer consists of signal nodes. The hidden layer consists
of radial basis function neurons and the output layer is linear. The
neurons in the hidden layer adopt the radial basis function as an
activation function which is generally the Gauss function:

radbasn en

n jjw1  pjjb1

The output a, of the Gauss function is:

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

where, normprod is a normalized dot product weight function in

MATLAB.
RBFNN or the GRNN is the network with an adaptive structure,
therefore the results do not change after repeatedly training them
and has a high stability which is superior to that of the BP network.

C. Xia et al. / Electrical Power and Energy Systems 32 (2010) 743750

MAPE 1=N h

3. STLF models using ANN and the forecasting results

APE

745

Nh

3.1. Data pre-processing

In any type of NN, accurate data is directly related to the adaptability of the model and its forecasting accuracy. In electrical power
load forecasting, historical events can have great effects on data. It is
important to reduce the amount of abnormal data within the dataset
as this tends to interfere with historical data patterns and hence the
forecasting accuracy. Therefore, before any data is used to train the
networks it is necessary to eliminate any abnormal data in order to
recover the true features of the electric load.
In order to increase convergence speed of network, before training, input and output data of NN is processed at [0, 1] as follows. The
historical maximum in the local region needs to be considered, e.g.
the load in Yichang is less than 1300 MW. Therefore, the hourly load
data could be divided by 1300. Similarly, wind speed divided by 15,
and precipitation by 100, atmospheric pressure by 1300, humidity
by 100 and day number of the week by seven. The temperature data
consists of the highest and lowest temperatures in a day (both given
by T). Since temperature has a signicant inuence on load variation
and the possible temperature scope in Yichang is 6 to 40 C, the
highest and lowest temperatures are processed as following:

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

The value 23 in formula (6) is the x value computed by |40  x| or

|6x|. The relationship curve between T1and T is shown in Fig. 1
which is known as a V-shape model for temperature processing.
3.2. Training parameters and forecasting results
The specimen data set used was established based on actual
historical hourly load data and weather data of Yichang in September, 2005, which includes wind speed, precipitation, atmospheric
pressure, maximum temperature, minimum temperature, humidity and day of the week. The data of the former 21 d in the above
specimen set are employed for training and the rest are used for
testing, thus, the hourly load forecasting values of the 22nd day
was predicted by means of (i) BPNN, (ii) GRNN and (iii) RBFNN.
Trial and error methods were utilized to determine the network
structure and parameter. The relative errors between the forecasting values and the actual values are shown in Table 1. The forecasting accuracy was evaluated by the average of absolute percentage
errors (APE).

APE jLa  Lf j=La  100

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/

Fig. 1. Relation curve between T1 and T.

where Nh is the number of hours in the forecasting period.

Structures and parameters of the BPNN model: this paper
choose TRAINLM function in MATLAB to train a neural network
with ve sigmoid neurons in the hidden layer and one linear neuron in the output layer at 50 epochs and 0.0001 training error goal.
Structures and parameters of the GRNN model: the function
NEWGRNN was used to build the GRNN model. Trial and error
methods were utilized to determine the spread parameter for the
model. This was then selected to be 0.4.
Structures and parameters of the RBFNN model: the function
NEWRBE (or NEWRB) was used to build the RBFNN model. Trial
and error methods were utilized to determine the spread parameter for the model. This was then selected to be 5.
In order to ensure the statistical signicance of the result and
reliability of proposed model, Wilcoxon rank sum test was taken
in this paper. The function [p, h, stats] = ranksum (x, y, alpha) was
used for the hypothesis test, performed at the 0.05 signicance level. The function returns p, h and stats from the test. The h-value
stands for the test result, and then p-value is the probability of
observing the given result. If h = 0, it indicates there is no evident
difference between the two sets of data x, y. The stats-value has
a structure with two elds. The eld ranksum contains the value
of the rank sum statistic and the eld zval contains the value of
the normal (Z) statistic. The actual load value in Table 1 is used
as x, whereas y is, respectively the forecasting values by BPNN,
GRNN and RBFNN. In addition, the test was also implemented between the three models. The results for Wilcoxon test in Table 2
indicate that all the h-values are 0, then the null hypothesis, i.e.,
medians are equal, cannot be rejected at the 5% level. Theoretically,
all the three models are reliable and valid. The p-value of GRNN vs.
RBF is 0.4037. But it is regretted that p-value of actual values vs.
forecasting values by RBFNN is slightly small and seems lack of signicant superiority. That is probably because the spread parameter
determined by trial and error methods is not best. In order to improve the drawback, perhaps statistical test and comparison for
RBFNN models at different spread parameters should be done before a better RBFNN model be determined. However, the MAPE value in Table 1 which is extensively applied in load forecasting
indicates the model based on the RBFNN is best one and has a higher forecasting accuracy. In addition, we found the results produced
by the BPNN model are not steady because of excessive training.
However, the results produced by the RBF or the GRNN do not
change after repeatedly training them and has a high stability
which is superior to that of the BP network. The forecasting accuracy of the GRNN model is slightly lower than that of the RBFNN
model. Therefore, the RBFNN or GRNN model would seem the better method to utilise when designing the VI for load forecasting.
4. MLTLF model and the forecasting results
4.1. Considerable factors in MLTLF
There are some conventional MLTLF methods such as the output-value-consumption, electric elasticity-coefcient, gross national product (GNP) synthetic-consumption, per capita
consumption, load density, relativity analysis, and total-outputallocation methods. Generally, these methods rely on statistical
analysis and need some other forecasting data, such as gross
domestic product (GDP), population and the area forecasting value.
In reality, many factors (such as time, weather condition, economy and random effects) affect power system load nonlinearly.
However it is difcult to determine what essential factors for a local area should considered for load forecasting. For example, the

746

C. Xia et al. / Electrical Power and Energy Systems 32 (2010) 743750

Table 1
Forecasting results of BPNN, RBFNN and GRNN models.
Hours

Actual values (MW)

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

Forecasting values (MW)

APE (%)

Forecasting values (MW)

APE (%)

Forecasting values (MW)

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

Actual values vs. forecasting values by

BPNN

Actual values vs. forecasting by values

GRNN

Actual values vs. forecasting values by

RBFNN

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

BPNN vs. GRNN

0.1012, 0
1.6393, 508

BPNN vs. RBFNN

0.3922, 0
0.8557, 546

GRNN vs. RBFNN

0.4037, 0
0.8351, 547

Table 3
Load data of Hubei in 19891997.
Year

1989

1990

1991

1992

1993

1994

1995

1996

1997

Peak load (104 KW)

342.4

381.8

405.7

414.4

452.9

528.4

547.9

583.1

640.4

Chinese economy has developed very rapidly in recent years, and

as such relevant policies are in the process of exploration or
improvement and there are many indeterminate factors which will
affect load forecasting. Therefore, forecasting results produced
using the time series model and a relativity analysis to t historical
data, are not satisfactory. It is important therefore to analyze the
load features for the local area, the power consumption structure,
the trend of historical load, the development of society and economy, the climate changes, etc., in order to design a forecasting
model that adapts well to the local area. The ANN forecasting
method can consider enough of these factors to give it a good prospect for accurate prediction.
4.1.1. The classications of factor
Generally, the factors which affect load can be classied into
tendency, season and climate, and random factors. Tendency factors refer to historical load data that has varying tendency, e.g.
yearly varying long-term power demand has obvious increasing

trend. When considering the tendency factors, it is necessary to

analyze historical load data in order to make suitable adjustments
at special situations, e.g. a new building project will have an affect
on power consumption, whilst a new heat and power plant may
lead to abnormal load uctuations. In such cases, it becomes necessary to adjust historical data to eliminate the inuence on a normal trend. In addition, since the tendency item is inuenced by the
macro economy, it is possible to transversely compare the power
consumption of each different type of industry. If industry, commerce, daily life, and urban public utilities are considered as the
primary factors, gross industrial output value, total volume of retail
sales, level of residential consumption and GDP data from recent
years could be collected and edited into a specimen set of neural
networks. Since the power consumption of a large power network,
which changes both monthly and yearly, is inuenced by many
factors, in addition to the trend of historical load, other synthetic
factors should be considered. The primary factors and their trends
should be analyzed in detail. Other tendency factors should also be

C. Xia et al. / Electrical Power and Energy Systems 32 (2010) 743750

taken into account, such as season, climate and random factors. As

for season and climate factors, the climate features in a local area
should also be researched, e.g. the summer in Hubei has often continual high temperatures and there is even a trend towards hotter
summers. Meanwhile, the winter temperatures have become even
colder. In this case, the highest and lowest temperature should be
edited into the specimen set of the ANN. Some important festivals,
such as May 1st, National Day and New Years Day are at a xed
time, and do not have a large inuence on medium and long term
load forecasting. But the spring festival changes yearly, although it
is generally in January or February. If the spring festival is in January or February, the load of this month is very different. Additionally, the load of February also depends on whether it is a leap year.
Therefore, in order to predict the load during important festivals,
daily load data corresponding to the festivals from recent years
could be collected into a historical database as well as the load data
from several days before the festivals. It should also be remembered that electric power load is also easily affected by accidental
or indeterminate factors. To take into account random factors, normally some leeway can be given according to some expert advice.
Thus it is possible to separate the stable random processes from
the unstable random processes and then describe them using the
auto regress moving average (ARMA) model, respectively.
4.1.2. Data collection and its pre-processing
Data collection is a very important process for accurate load
forecasting. The data that can be collected includes: (1) the statistical data of social and economic development including GDP, total
population, output value of 1st, 2nd & 3rd industry; (2) power consumption data of each industry, including power consumption of
1st, 2nd & 3rd industry and the whole society; (3) daily load records, including 8760 hourly load data in each year; (4) forecasting
data of the developing trends with regards local GDP and population during the plan year; (5) signicant strategy intentions, industrial structure adjustment, major objectives of scientic
developments relating to the national economy; (6) electricity
price; (7) seasonally varying situations of power consumption or
load; (8) weather and climate data; (9) power consumption situations of big consumers and (10) increasing situations of power consumption in other districts at home and abroad.
Accurate data is related to the model adaptability and the forecasting accuracy. Historical accidents or other special factors may
affect the data. That is, they may interfere with the historical tendency and inuence the forecasting accuracy. Therefore, as previously mentioned it is necessary to eliminate the abnormal data
as much as possible in order to have a clearer picture regarding
load trends, e.g. when network faults result in power loss in a large
area, it is possible to estimate the effect of this and then amend the
power consumption data.
4.1.3. Forecasting evaluations
A forecasting evaluation is used to evaluate the accuracy of the
forecasting model before an actual forecast is used. Historical data
can be used to test the model. A reasonable forecast not only relies
on scientic theory, reliable data, and an advanced approach, but
also on forecasting experience, logical reasoning ability, capacity
of comprehensive analysis and, of course, the judgment ability of
the forecasting personnel.
4.2. MLTLF models using ANN methods
4.2.1. MLTLF model based on LM algorithm
As mentioned previously, a BP algorithm usually requires a long
training time and quite often it can become xed at a non-optimized solution. An improvement to the BP algorithm is based on

747

LevenbergMarquardt (LM) algorithm. The regulation of weight

adjustment in LM algorithm is as follows:

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

C. Xia et al. / Electrical Power and Energy Systems 32 (2010) 743750

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

Peak load (10 KW)

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

to display the steady outcomes. Therefore, this test has illustrated

the superiority of the RBFNN model. The front panel and block diagram are shown in Figs. 2 and 3.
If we have an ANN forecasting model for MLTLF and follow the
example of STLF, we can also establish a relevant VI for MLTLF.

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

Fig. 2. The front panel of a virtual load forecaster.

C. Xia et al. / Electrical Power and Energy Systems 32 (2010) 743750

749

Fig. 3. The block diagram of a virtual load forecaster.

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

Fig. 4. Flowchart for a multi-purpose virtual load forecaster.

be possible to acquire a model with higher accuracy. Of course, this

should be a very fruitful area for future research.
7. Conclusion
This paper introduces a method that combines neural network
models of load forecasting with virtual instrument technology in
order to build a virtual forecaster. It also presents the scheme of
a multi-purpose virtual load forecaster, to establish a united database for load forecasting and to combine short-term load forecasting with medium and long term load forecasting. When
establishing a short-term load forecasting model, we should take

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

C. Xia et al. / Electrical Power and Energy Systems 32 (2010) 743750

[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.