15 views

Uploaded by Ryan Akbar

- Inverse Laplace Transform by p Fe
- ct
- Addressing
- Magneto Hydro
- A Multi-objective Optimal Power Flow Using Particle Swarm Optimization
- Stack Computers
- Reactive Power Control (a.k.a) How to Get B
- DC – DC Converter (Chopper)
- Reduction Matrix and Transient Ana Lysis
- Load Forecasting
- Stage 2
- theses
- Reactive Power Control
- How to Calculate Solar Energy System Production
- R-8 hydro-thermal_coordination
- memory organisation
- Short Term Load Forecasting Using Multi Layer Perceptron
- Memory Organisation
- INTELLIGENT CONTROL TECHNIQUE USING ARTIFICIAL NEURAL NETWORK
- ELECTRICAL PROJECTS USING MATLAB/SIMULINK Multi Machine Power System Stability Enhancement Using Static Synchronous Series Compensator (SSSC)

You are on page 1of 8

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

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

Nomenclature

n

a

p

w1

b1

w2

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

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.

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.

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

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

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.

MAPE 1=N h

APE

745

Nh

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

|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).

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/

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

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

BPNN

GRNN

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

0.1012, 0

1.6393, 508

0.3922, 0

0.8557, 546

0.4037, 0

0.8351, 547

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

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

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

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

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

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

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

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

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

[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.

- Inverse Laplace Transform by p FeUploaded bySyed Danish Fayaz
- ctUploaded byDeepak Tardalkar
- AddressingUploaded byTapan Chowdhury
- Magneto HydroUploaded byTadi Chiranjeevi
- A Multi-objective Optimal Power Flow Using Particle Swarm OptimizationUploaded bySurya Ch Venkata
- Stack ComputersUploaded byjyothibellary4233
- Reactive Power Control (a.k.a) How to Get BUploaded byapi-3747381
- DC – DC Converter (Chopper)Uploaded byyopisukita
- Reduction Matrix and Transient Ana LysisUploaded byEmil Popov
- Load ForecastingUploaded bywikyarif
- Stage 2Uploaded byMohit Kumar Shakya
- thesesUploaded bykartikeysaurabh
- Reactive Power ControlUploaded byKhurram Hashmi
- How to Calculate Solar Energy System ProductionUploaded byJaspal Singh Rattey
- R-8 hydro-thermal_coordinationUploaded bySanjana Samantray
- memory organisationUploaded byGautam Manaktala
- Short Term Load Forecasting Using Multi Layer PerceptronUploaded byIJMER
- Memory OrganisationUploaded bysansure
- INTELLIGENT CONTROL TECHNIQUE USING ARTIFICIAL NEURAL NETWORKUploaded byShivan Biradar
- ELECTRICAL PROJECTS USING MATLAB/SIMULINK Multi Machine Power System Stability Enhancement Using Static Synchronous Series Compensator (SSSC)Uploaded byAPLCTN
- FUNDAMENTAL CONCEPTS of ARTIFICIAL NEURAL NETWORKSUploaded byYağmur Fidaner
- Direct energy conversion Lecture I.pdfUploaded byAftab Ali Afridi
- Fuzzy Ideology Based Long Term Load ForecastingUploaded byCristian Klen
- Short Term Load ForecastingUploaded bySunil Kumar
- Transient Stability Analysis of Nine Bus System with multiple contingenciesUploaded byShashank Veerapaneni
- Electricity Load Forecasting by Artificial Neural Network ModelUploaded byBogdan Vicol
- On-line Power Systems Voltage Stability Monitoring using Artificial Neural NetworksUploaded byLucian Toma
- ETAP 7 InstallationUploaded byHamid Paminto Nugroho
- Programming in RUploaded byJulie Finch
- Transient Stability Analysis with PowerWorld SimulatorUploaded byvitaxe

- KIR Crane 50 TonUploaded byRyan Akbar
- ipi25009Uploaded byRyan Akbar
- Grafik Kualitas Daya Elektrik 2015Uploaded byRyan Akbar
- Tugas Komcer II Jst Dan AgUploaded byRyan Akbar
- Laporan Percobaan Titrasi Asam BasaUploaded bypadmanaba
- 1-s2.0-S0888613X99000201-mainUploaded byRyan Akbar
- 1-s2.0-S0888613X99000201-mainUploaded byRyan Akbar
- akun 4shared.txtUploaded byRyan Akbar

- Ultrasound PDMUploaded byMahesh Kumar
- MHDD ManualUploaded byuniyo_987
- Holland Some Diagnostic Scales for Research in Decision Making and Personality Identity, Information, And BarriersUploaded byMariana Ciceu
- MAX30102.pdfUploaded byEdward Amoyen Abella
- Interview GuideUploaded byMohammad Asif
- Machine Design-II EME 602_NEWUploaded bynitin_johri
- Static ElectricityUploaded byDeepak Gupta
- Mnhug Infrastructurepresentation 140324074758 Phpapp01Uploaded byVenkata Apparao S
- Horiba D-52_53_54_55_manualUploaded bySaray Landeros García
- 2009 BaldridgeUploaded byMigUel Angel Rocha
- Atm Po AdminUploaded bymevrick_guy
- TABMA Technical Note 3 - Pergola Span TablesUploaded byGao Weiquan
- Chubby PaperUploaded bySri Harsha
- Basic Structures in VhdlUploaded byAnimesh Sharma
- Practical Implementation of WSN Based Data Acquisition SystemUploaded byTutor Iqbal
- What is Electrical StressUploaded byLedred1
- 12p5Uploaded byBerry101
- US96 480US BoosterBulletin Tcm67 9543Uploaded byJose Santos
- Bt200 Profibus Cable TesterUploaded by24ava
- 04 (1).pdfUploaded byElenildo Barros
- thermo Term Paper A04Uploaded bymanish_jain5791
- Standard Specification for WASHED AND DRIED SANDUploaded bykkamalakanna
- 1GA48_2eUploaded bykarayev
- Seidel 1998 Qualitative Data AnalysisUploaded byTravis McGee
- completeMcaRevisedSyllabusWef2009-10Uploaded byBapuji Valaboju
- Networking Solutions Products Swtich GuideUploaded byAzahar
- GNU Emacs Lisp Reference ManualUploaded byManuel Vázquez Acosta
- Cheat SheetUploaded bydedzcnaga
- lm56.pdfUploaded bytutesam
- History of MeasurementUploaded byJason Bueta