Application of Neural Networks for
Short-Term Load Forecasting
Reza Afkhami, and F. Mosalman Yazdi
Abstract--This paper presents the development of an
artificial neural network based short-term load forecasting
model. The model can forecast daily load profiles with a load
time of one day for next 24 hours. In this method can divide days
of year with using average temperature. Groups make according
linearity rate of curve. Ultimate forecast for each group obtain
with considering weekday and weekend. 24 hours of a day
divided to 3 groups at 8 hours, network for every each of eight
varieties must interpolate. This paper investigates effects of
temperature and humidity on consuming curve. For forecasting
load curve of holidays at first calculate pick and valley and then
the neural network forecast is re-shaped with the new data. The
networks are trained using hourly historical Load data and daily
historical max/min temperature and humidity data. The results
of testing the system on data from yazd utility are reported.
Index Terms--Artificial neural network, Holiday forecasting,
pick and valley load forecasting , Short-term load forecasting.
I. INTRODUCTION
L OAD forecasting is vitally important for the electric
industry in the deregulated economy. It has many
applications including energy purchasing and generation, load
switching, contract evaluation and infrastructure development.
Scheduling and planning for energy transactions. Numerous
forecasting techniques based on exponential smoothing [1],
Box-Jenkins [2] and expert system [3] have been proposed. In
this technique the functional relationship between the forecast
load (output) and the historical load and/or historical weather
variable (input) must be defined in advance prior to the
building of load models. Recently a number of artificial neural
network approaches for electric load forecasting have been
proposed. In this study, ANN techniques have been used to
forecast the daily peak and valley load, daily load curve and so
on.
The ANN is regarded as a powerful method for handling
nonlinear complex phenomenon, and it is able to develop a
forecasting model automatically only by training with stored
actual data. In the methods used, ANN training data are
observed for a few weeks before the target day. However the
ANN method, using data from the few weeks proceeding the
target day as training data, can not accurately forecast seasonal
trends mainly due to insufficient learning pattern [4]. 1
Reza Afkhami, PRT Inc., Dallas, Texas http://www.prt-inc.com
(email: Saeed@prt-inc.com).
F. Mosalman Yazdi is with the Department of Electrical Engineering,
Islamic Azad University of Mehriz, Iran (e-mail: fmosalman@gmail.com).
0-7803-9525-5/06/$20.00 ©2006 IEEE.
The ANNs described in this paper are multilayer
perceptron structures designed to forecast the hourly load for
24 hours ahead. The back propagation has been for ANN
training. The purpose of this study is to forecast load
accurately, using actual data from the same period of previous
several years are training data in order to expand learning
pattern. With the proposed method, the mean absolute
forecasting error was below 2%. Moreover, the forecasting
errors of the methods in this paper have became smaller than
the errors of the method using data only previous year of the
target year as ANN training data. This confirms the
effectiveness of the training using actual data from several
years. This study used actual power load and weather
conditions during the one at yazd electric power co., yazd,
Iran.
II. ANN
A generic feed forward neural network (FNN) is shown in
fig.1, which has input, hidden and output layer of processing
elements. Processing elements in an ANN are also known as
neurons. These neurons are interconnected by means of
information channels called interconnections. Each neuron can
have multiple inputs, while there can be only one output.
Inputs to a neuron could be from external stimuli or could be
from output of other neurons. Copies of the single output that
comes from a neuron could be input to many other neurons in
the network. It is also possible that one of the copies of a
neuron's output could be input to itself as a feedback. In this
case, the network is called self recurrent neural network.
There is connection strength, synapses or weight associated
with each connection. When the weighted sum of the inputs to
a neuron exceeds a certain threshold, the neuron is fired and
output signal is produced. The network can recognize input
patterns once the weights are adjusted or tuned via some kind
of learning process. The back propagation learning algorithm
is the most frequently used method in training the feed
forward neural networks. In order to train the self recurrent
neural networks, a modified learning algorithm called
dynamic back propagation learning algorithm [5] is used.
 can be noted that if wjd=0, then (3) is for the usual feed
forward network. For the recurrent weight wjd, the error
gradient is represented as fallows
∂E p ∂o pj
− = δ pj ( z −1 (o pj ) + w jd z −1 ) (7)
∂w jd ∂w jd
Where δpj is similarly defined as (4) or (5), and the output
gradient is generated by the dynamic or recurrent equation as
follows.
∂o pj ∂o pj
= f j′ (net j )( z −1 (o pj ) + w jd z −1 ) (8)
∂w jd ∂w jd
If the self recurrent neurons are only in the hidden layer,
not in the output, then the network is called a diagonal
Fig. 1. Artificial neural network.
recurrent neural network [6]. Thus the rule of adjusting
weights is defined as
III. BP ALGORITHM ∂E p
∆w ji (n + 1) = η (− ) + α∆w ji (n) (9)
An error function for a pattern p can be defined as ∂w ji
1
Ep = ∑ (t pj − o pj ) 2 (1) Where η is the learning rate, α is the momentum factor and
2 j ∆ represents the changes in weight.
And E = ∑ E p is the overall measure of the error for all IV. LOAD MODEL DESCRIPTION
p
patterns, p=1, …, P. here tpj is the target output for jth The most of the forecaster used a multilayer perceptron
component of the output pattern for pattern p and opj is the network. The MLP must be trained with historical data to find
corresponding network output. The self recurrent neural the appropriate values for wij and the number of required
network [5] is specified as neurons in the hidden layer. Prior to the training of neural
networks, the weights in the network are set randomly.
o pj = f j (net pj )
Multiple training cases are then applied to adjust the weights
(2)
net pj = ∑ w jk o pk + w jd z −1 (o pj ) using a parameter optimization technique based on back
k propagation. In this technique, the steepest descent algorithm
Where fj is a differentiable and no decreasing function is used to minimize the error square function
(sigmoid type), wjd is the weight for the delayed feedback, wjk 1 k
is input or output weight and z-1is a unit delay operator. To E= ∑ ( L(i ) − Lˆ (i )) 2 (10)
obtain a rule for adjusting weights, the gradient of Ep with
2 i =1
respect to input or output weight wij is used and it is Where k is the number of training cases used. The weights
represented as follows in between the hidden/output layers and input/hidden layers
are adjusted by propagation back from the output layer the
∂E p ∂o pj
− = δ pj (o pi + w jd z −1 ) (3) error difference between the actual load and predicated load.
∂w ji ∂w ji
V. ADAPTIVE SCALING
Where δpj is the equivalent error defined for the output
layer as The adaptive scaling employed in the daily forecaster and
δ pj = (t pj − o pj ) f j (net pj ) (4) temperature forecaster are identical exert for the number of
temperatures considered. In either case, the outputs of the
And for a unit in an arbitrary hidden layer as ANN are scaled so that the network forecast matches the
δ pj = f j′ (net pj )∑ δ pk wkj (5) desired high and low temperatures. Scaling uses a linear
k transformation
Where δpk is the equivalent error back propagated from the TS (k ) = m(k ).TU (k ) + b(k ) (11)
next layer and fj′ is the derivative of the sigmoid function fj.
Where TU(k) and TS(k) are the unscaled and scaled
The output gradient is generated by the dynamic or recurrent
temperature, respectively, at hour k and m(k) and b(k) are
equation as follows
scale and offset parameters, respectively. Note the dependence
∂o pj ∂o pj of the scale parameters (m and b) on time k typically, the scale
= f j′ (net pj )(o pi + w jd z −1 ) (6)
∂w ji ∂w ji parameters change at the hour of the day where a high or low
occurs in the network forecast [7].
For the hidden or output layer, opi is the ith component of
the input, which can be the output from the pervious layer. It
 VI. THE WEATHER FORECASTING The network of weekends has 120 inputs and 8 outputs.
Although temperature is one of the effective climate variety These inputs are 24 hourly loads of the previous day, 24
on consuming curvature, but other climate factors such as hourly temperature of the previous day, 24 hourly relative
relative humidity and wind speed have considerable effect on humidity of the previous day, 24 hourly temperature forecasts
consuming ration. In this research we only affect the humidity for the coming day, 24 hourly relative humidity forecasts for
effect. the coming day, 2 day type indicators for each day of the
week. The outputs are the 8 hourly load of the day to be
A. The Hourly Temperature Forecaster forecast.
One of the input data of neural network for load forecast is
temperature. Therefore it is necessary to forecast temperature VIII. HOLIDAYS
of special day temperature doesn't depend on type of day & it Load curve of holidays differ from a typical weekday, also
is special occasion therefore forecast doesn't need to classified number of these days in historical information in compression
days of year. Other way temperature of each day can be with a typical weekdays is less. Because of unormal load
forecasted from temperature of previous day. The resultant of behavior in these days and not enough samples neural network
considering temperature information of 3 years & test can't trained. For forecasting load curve of holidays at first
different training set are for getting high accuracy must use calculate pick and valley and then the neural network forecast
information of 5 days prior to the day to be forecast. By this is re-shaped with the new data. In this paper we use two
the accuracy will be about 100%. Network has 28 inputs and different neural networks to predict peak and valley load
24 outputs. These inputs are 24 hourly temperatures of the separately.
previous day, the high and low temperatures of the day to be The network of peak load forecasting has 23 inputs and 1
forecast, and the high and the low temperatures of two days output. These inputs are daily high temperature and relative
prior to the day to be forecast. The outputs are the 24 hourly humidity on that day, high temperature and relative humidity
temperatures of the day to be forecast. and Peak loads on seven preceding days of the same day type
B. The Hourly Relative Humidity Forecaster as the particular day. The output is peak load of the day to be
Relative humidity data affect the consuming rate therefore forecast.
relative humidity data of forecasting day must be forecasted. The structure of the neural network for valley load
Input to the forecasting network humidity is the same as forecasting is the same as that for peak load forecasting, but
temperature therefore has training set, that is consist low temperature and relative humidity and valley employed.
information of 5 days prior to the day to be forecast. Network
has 26 inputs and 24 outputs. These inputs are 24 hourly IX. PERFORMANCE
relative humidity of the previous day, the high and low The performance of the proposed neural network is tested
relative humidity of the day to be forecast. The outputs are the on real data from yazd power system. Three years of historical
24 hourly temperatures of the day to be forecast. data is used to train the neural network. Actual weather data
is used so that the effect of weather forecast errors do not alter
VII. THE LOAD FORECASTING the modeling error.
For load forecasting must divide days of year to many The following formula has been used to calculate the
groups with considering temperature. Information of 3 years percentage error
used to increase accuracy. 3 samples exist for every day, must T arg etOutput − Pr edictionOutput
% Error = × 100 (12)
calculate average curve of years then obtain average from it's T arg etOutput
24 hours. Therefore there is an average for each day and an Tables I and II illustrates the result of our approach for
average curve for each year. With dividing this curve can forecasting peak and valley load several holidays in yazd.
divide a year to some groups. Groups make according linearity
rate of curve, linear parts have little variety and are in one TABLE I
group. PEAK LOAD FORECAST RESULT FOR MARCH 2005
For load forecasting of ordinary days of a year must divide
days of a week to weekdays and weekend days. For each Day Actual Peak Forecasting Peak %error
group there is one network. 24 hours of a day divided to 3
20 635 630 0.78
groups at 8 hours, network for every each of eight varieties 21 615 614 0.16
must interpolate and finally accuracy increases. 22 620 622 -0.3
The network of weekdays has 124 inputs and 8 outputs. 23 620 617 0.48
24 600 605 -0.83
These inputs are 24 hourly loads of the previous day, 24
hourly temperature of the previous day, 24 hourly relative
humidity of the previous day, 24 hourly temperature forecasts
for the coming day, 24 hourly relative humidity forecasts for
the coming day, 4 day type indicators for each day of the
week. The outputs are the 8 hourly load of a day.
 TABLE II TABLE IV
VALLEY LOAD FORECAST RESULT FOR MARCH 2005 LOAD FORECAST RESULT FOR 15 th SHABAN1

Day Actual Peak Forecasting Peak %error Hour Forecasted Load Actual Load %error

20 330 333 -0.9 1 341.52 341 -0.15

21 330 329 0.3 2 321.93 325 0.94
22 331 333 -0.6 3 330.42 327 -1.04
23 333 332.2 0.24 4 326.84 322 -1.5
24 332 333.5 -0.45 5 323.45 324 0.17
6 330.7 332 0.39
7 330.1 335 1.46
The forecasting results for one normal day and holiday are 8 340.42 338 -0.72
9 352.24 356 0.49
presented in tables III and IV. The results are listed in table 10 356.01 357 0.28
III, IV are depicted in Fig. 2, Fig. 3. 11 352.65 358 1.49
12 339.52 345 1.59
TABLE III 13 351.35 356 1.31
LOAD FORECAST RESULT FOR 13 th JUNE 2005 14 348.39 353 1.3
15 365.95 371 1.36
16 419 421 0.47
Hour Forecasted Load Actual Load %error 17 548.66 559 1.85
18 593.1 602 1.48
1 324.16 323 -0.36 19 562.97 573 1.75
2 310.55 308 -0.8 20 545.04 554 1.62
3 301.64 299 -0.88 21 512.99 506 -1.38
4 294.64 296 0.48 22 440.25 446 1.29
5 299.53 301 0.49 23 373.63 376 0.63
6 276.1 278 0.68 24 345.17 341 -1.2
7 253.04 255 0.77
8 263.44 262 -0.55 Maximum error: 1.85%, minimum error: 0.15%
9 280.41 283 0.91
10 293.19 291 -0.75
11 292.23 294 0.6
12 293.92 296 0.7
13 287.02 289 0.68
14 285.97 287 0.36
15 281.47 282 0.19
16 280.93 282 0.38
17 286.38 284 -0.84
18 290.75 293 0.77
19 300.95 302 0.35
20 346.44 343 -1
21 552.37 555 0.47
22 540.32 543 0.49
23 456.76 458 0.27
24 375.16 377 0.49

Maximum error: 1%, minimum error: 0.19%

Fig. 3. Comparison of forecasted hourly loads (+...) and actual loads (---) for
15 th shaban (20 September 2005).

X. CONCLUSIONS
The system load forecasting model is a critically important
decision support tool for operating electric power systems. An
artificial neural network-based approach has been applied to
forecast the hourly electric load of the power system of yazd.
The ANN multi-layer structure has been trained using the
back propagation technique. For load forecasting must divide
days of year with using average temperature, Groups make
according linearity rate of curve. Ultimate forecast for each
group obtain with considering weekday & weekend. 24 hours
Fig. 2. Comparison of forecasted hourly loads (+...) and actual loads (---) for
1
13 th June 2005. Shaban is a lunar month and 15th shaban is the one of the religious
holiday in Iran.
of a day divided to 3 groups at 8 hours, for each groups there
are one network. For forecasting load curve of holidays, high
and low load is first estimated then the neural network forecast
is re-shaped with the new data.
XI. REFERENCES
[1] W. R. Christiannse, "Short-term load forecasting using general
exponential smoothing," IEEE Transactions PAS, vol. 90, no. 2, pp.
900-911, March/April, 1971.
[2] J. H. Park, Y. M. Park, and K. Y. Lee, "Composite Modeling for
Adaptive Short-Term Load Forecasting," IEEE Transactions on Power
Systems, vol. 6, pp. 450-457, May. 1991.
[3] S. Rahman, and R. Bhatnagar, "An Expert System Based Algorithm for
Short-Term Load Forecast," IEEE Transactions on Power Systems, vol.
3, pp. 392-399, May. 1988.
[4] T. S. Dillon, S. Sestito, and S. Leung, "An Adaptive Neural Network
Approach in A Power System," IEEE Proc. of 1991 ANNPS, pp. 17-21,
July. 1991.
[5] K. Y. Lee, Y. T. Cha, and C. C. Ku, "A Study on Neural Networks for
Short-Term Load Forecasting," Proc. of Applications of Neural
Networks to Power Systems 1991, pp. 26-30, Seattle, WA, July. 1991.
[6] C. C. Ku, and K. Y. Lee, "Diagonal Recurrent Neural Network Based
Control Using Adaptive Learning Rates," Proc. 31st IEEE Conf. on
Decision and Control Tucson, Arizona, pp. 3485-3490, Dec. 16-18,
1992.
[7] A. Khotanzad, M.H. Davis, A. Abaye, and D.J. Martukulam, "An
Artificial Neural Network Hourly Temperature Forecaster with
Applications in Load Forecasting," IEEE Trans. PWRS, vol. 11, pp. 870-
876, May. 1996.
XII. BIOGRAPHIES
Reza Afkhami-Rohani received his BS and MS
degrees in Electrical Engineering from Sharif
University of Technology, Tehran, Iran, in 1989 and
1992. He was an instructor in Sharif University in
1992. Reza received his Ph.D. in the same field from
Southern Methodist University, Dallas, Texas with
outstanding award in 1998. He has over 12 years of
experience in the field of intelligent systems and
their applications to energy forecasting problems.
Currently, he is a director as well as senior R&D
engineer at PRT Inc. at Dallas. His research interests are in signal processing
and neural networks and their applications in modeling and forecasting.
Fatemah Mosalman-Yazdi was born in yazd, Iran,
in 1980. She received the B.S. degree in Electrical
Engineering from Islamic Azad University of Yazd,
Iran, in 2002, the M.S. degree in Electrical
Engineering from Nagaf abad University, Esfahan,
Iran, in 2005. She has been working at Islamic Azad
University of Mehriz. Her main research interests are
in the area of artificial intelligent methods and their
application in power systems.