PERGAMON Engineering Applications of Articial Intelligence 11 (1998) 307316

Contributed Paper

A real-time short-term peak and average load forecasting system

using a self-organising fuzzy neural network
P.K. Dash a, H.P. Satpathy b, A.C. Liew a
a
Department of Electrical Engineering, National University of Singapore, Singapore
b
Department of Electrical Engineering, Regional Engineering College, Rourkela, India
Received 15 October 1996; in revised form 1 June 1997

Abstract

This paper presents a self organising fuzzy-neural-network-based short-term electric load forecasting system for real-time
implementation. A learning algorithm is devised for updating the connecting weights as well as the structure of the membership
function of the network. The number of rules in the inferencing layer is optimised; this in turn optimises the network structure.
The proposed algorithm exploits the notion of error back-propagation. The network is initialised with random weights.
Experimental results of the system are discussed from a practical standpoint. The system accounts for seasonal and daily
characteristics, as well as abnormal conditions, holidays and other conditions. It is capable of forecasting load with a lead time
of one day to one week. The adaptive mechanism is used to train the network on-line. The results indicate that the proposed
load forecasting system is robust and yields more accurate forecasts. Furthermore, it allows greater adaptability to sudden
changes, compared with simple neural-network or statistical approaches. Extensive studies have been performed for all seasons,
and some of them are presented in this paper. The new algorithm is tested with a typical load date of the Virginia Utility, and
produces a very robust and accurate forecast, with a Mean Absolute of Percentage of Error (MAPE) mostly less than 1.8% for
24-hours-ahead peak load forecast, and 1.6% for a 168-hours-ahead forecast. # 1998 Published by Elsevier Science Ltd. All
rights reserved.

Keywords: Real-time implementation; Short-term electric load; Forecasting system; Self organising fuzzy neural network

1. Introduction and articial neural-network algorithms (ANN). Time

series [2] and regression [3] are two major classes of
Accurate short-term (one day to one week ahead) load conventional statistical algorithms, and have been
forecasts represent a great saving potential for the applied successfully in this eld for many years.
economic and secure operation of power systems [1]. However, these techniques do not produce a su-
Forecasts of peak and average loads are necessary for ciently accurate forecast, and the accuracy deteriorates
scheduling functions such as hydro-thermal co-ordina- for longer variations of non-stationary load and
tion, Interchange evaluation, fuel allocation, mainten- weather variables.
ance and security assessments. So the development of Over the past few years, articial neural networks
an accurate, fast, simple and robust short-term load (ANN) have been proposed as a powerful tool for
forecasting methodology is of importance to both elec- short-term load forecasting [46] problems. It is
tric utilities and their customers, and in turn intro- known that articial neural networks do not require
duces higher reliability and better management. any explicitly dened relationship between input and
For nearly two and half decades researchers have output variables. The corresponding mapping between
studied this problem of improving load forecast accu- input and output is obtained using a training algor-
racy, and the various models proposed have varied in ithm. The ANN modelling needs only the selection of
the complexity of their functional form and estimation input variables, thus avoiding the diculties associated
procedures. with conventional modelling processes.
The most commonly used techniques include statisti- Peng et al. [5] proposed a minimum-distance-based
cally based techniques, expert-system-based methods strategy to identify the appropriate historical patterns

0952-1976/98/$19.00 # 1998 Published by Elsevier Science Ltd. All rights reserved.

PII: S 0 9 5 2 - 1 9 7 6 ( 9 7 ) 0 0 0 6 9 - 9
308 P. Dash et al. / Engineering Applications of Articial Intelligence 11 (1998) 307316

of load and temperature for training the network. A days like Christmas and New Year are treated as
partially connected network [4] consisting of main and Sundays. Other holidays, including Good Friday,
supporting blocks is also proposed, which makes use Memorial Day, Independence Day, Labour Day, the
of model reference and functional relationships days preceding Christmas and New Year, and the day
between input (e.g., past load, weather, day type and after, are treated as Saturdays.
hour of the day) and output variables. While all these The approach presented in this paper is amenable to
prior works focus on the application of back-propa- real-time implementation, as hourly or daily adap-
gation training algorithms to train and update the tation of model parameters can be undertaken. The
model parameters, Peng et al. [2, 7] explored the eect network is trained and tested using typical load-data
of a (model development) ARMA Box-Jenkins model, from the Virginia Utility, USA. A comparison has also
and make use of the Widrow-Ho delta rule [8] to been made with the ANN and FNN [22] models for
update the model parameters. short-term load prediction, using the above-mentioned
Fuzzy systems have become another research area load data.
that is receiving increased attention. The pioneering
work of Zadeh [9] in fuzzy set theory has inspired
work in many research areas, including load forecast-
ing [10, 11]. Many fuzzy expert systems have been 2. Overview of the proposed approach
developed for short-term load forecasting (STLF), and
the heuristic features of an expert system provide an One of the salient aspects of a fuzzy inference sys-
excellent approach to this. Fuzzy-logic-based expert tem (FIS) is the determination of the knowledge base
systems for load forecasting require the development (KB), which consists of the following subsystems:
of a fuzzy rule base, which relies upon detailed know-
ledge of the parametric variation of the load pattern. . A mechanism for developing the membership func-
Recently, fuzzy-neural-nets (FNN) have been applied tions.
for model building [12, 13]. Usually, one of two . A fuzzy reasoning mechanism.
approaches is adopted for implementing these FNNs. . The number of rules and the rule-base.
One of the approaches handles fuzzied input data. In
the second case the weights of the network are com- This is supported by an adaptive mechanism that
puted, based upon a fuzzy rule base, but without fuzzi- allows learning to take place. This section presents the
fying the input data [14, 15]. fuzzy inference system (FIS), along with the governing
Neural-network-based models for forecasting pro- equations. The network, consisting of input, fuzzica-
blems are less complex than fuzzy-logic-based systems. tion, inferencing and defuzzication layers, is shown in
However, the simplicity is achieved at the cost of an Fig. 1. The network consists of N input variables, with
explicitly dened relationship between the individual N neurons in the input layer and R number of rules,
inputs and the overall model parameters. Fuzzy-logic- with R neurons for inferencing; thus the number of
based systems allow some insight into the model neurons in the fuzzication layer is N  R. The inputs
parameters with the help of membership functions and to the model are chosen as follows:
rules. 2 3
Ya k D 1
The objectives of the present approach is to study a 6 Ya k 2D 1 7
self-organising fuzzy-neural-network (FNN) that com- 6 7
6 Ya k nD 1 7
bines the self-organising capability of neural networks 6 7
X6 6 Ta k 1
7
7 1
with fuzzy-logic reasoning attributes. The network 6 Ta k D 1 7
modelling starts with a random set of weights, and hence 6 7
4 Ta k 2D 1 5
an arbitrary set of fuzzy sets. The network is initialised Ta k nD 1
with a suciently large number of rule-nodes, which sub-
sequently get optimised. An adaptive mechanism for and the output Y = Yd(k), where Yd(k) represents the
weight updating has been devised, together with the load Ta(k) is the temperature at the kth hour and D is
updating of the associated parameters of the fuzzy mem- the time step ahead for which forecasting is desired.
bership function. The training algorithm exploits the The past loads are taken into account to improve
notion of error back-propagation. The approach pre- upon the prediction capabilities; the notion being simi-
sented in this paper is highly exible, and Sundays and lar to that of auto-regression [3, 16]. The temperatures
holidays can easily be included. This is done by treating are included to reect the weather-sensitivity of the
the Sundays and Saturdays separately from weekdays load. It should be noted that the (k + 1)th element of
(Monday to Friday). The load curve on Sundays and the input vector is the a priori information on the tem-
public holidays is similar in nature, with small deviations perature at the hour at which load forecasting is to be
due to load use pattern on holidays. So, important holi- done.
 P. Dash et al. / Engineering Applications of Articial Intelligence 11 (1998) 307316 309

Fig. 1. Load forecasting model using a self-organising fuzzy neural network (SFNN).

The input to the fuzzication layer is a weighted ver- weights signify the strength of each rule in the output of
sion of the input variables, which can be represented as the model. The output Xout, is given as
X
R
Z W1T X W0 :
 2 Xout x1 ; x2 ; . . . ;xN j mj x1 ; x2 ; . . . ; xN
j1
The set of weights between the input and fuzzication
layers is given by: X
R Y
N

j exp jwij1 xi wij0 jgif : 8
W fW0 ; W1 g j1 i
n o
wij0 ; wij1 : i 1; . . . ; N; j 1; . . . ; R 3

alternatively, the individual elements of the fuzzication 3. Training and model adjustment
(2nd) layer are given by
3.1. Training procedures
Zij wij xi wij0 : 4

The output of each neuron in the fuzzication layer is a The network is trained to minimise an objective
fuzzy membership, corresponding to a particular choice function. The performance index (PI) to be minimised
of variables. The activation (membership) function used is the Mean-Square-Error (MSE), given by
for this fuzzication layer is given by PI = (yd(k) y(k))2, where yd(k) is the desired output

and y(k) is the model output. The model parameters of
f Zij exp jZij jgij ; 5 the proposed FIS are updated using the notion of
error back-propagation.
where gij is in the range of 1.0 R gijR5.0. During training, The aim is to optimise the performance index
gij is updated in this range, which is initially set at 2. The PI = (yd(k) y(k))2. The Wij1, Wij0 and gij are updated
output of the fuzzication layer can be expressed as till some stopping criterion is reached. The changes in


mij f Zij exp jwij xi wij0 jgij ; 6 Wij1, Wij0 and gij are computed by dierentiating the
PI with respect to the required parameters. Dene
where mij is the value of the fuzzy membership function of
2
e2k yd k yk ; 9
the ith input variable corresponding to the jth rule. The ac-
tivation function in the inference layer (3rd) uses multipli- 2 2
cative inference. The output of this layer is given by ^k1 @ek and r^k0 @ek ;
r
@W1 @W0
Y
N
2
mj x1 ; x2 ; . . . ; xN mij xi : 7 ^k1 @ek 2ek @ek ;
r
i @W1 @W1
@ek ^ @yd k yk @yk
The defuzzication layer (4th) has the connecting weights  rWij1 :
@wij1 @wij1 @wij1
(nj) to the output from the inferencing layer, and these
310 P. Dash et al. / Engineering Applications of Articial Intelligence 11 (1998) 307316

From equation (7)

" # ek yd k yk;
@yk @ XR YN
rwij k gij k:mn k:n k:Qkgij k1 and
gij
j exp jwij1 xi wij0 j
@wij1 @wij1 j1 i1 rgij k mn k:n k:Qkgij k loge Qk:
0 1
B YN
C @ Q(k) is dened as
j @ exp jwij1 xi wij0 jgij A:
i1
@wij1 Qk jwij1 k:xi k wij0 kj:
i6j
2 3
The value of a is also tuned according to the following
X
R

4 exp jwij1 xi wij0 jgij 5 update equation:
j1
ak ak 1 0:2ek 0:1_ek; 17
gij 1
j mj gij
jwij1 xi wij0 j xi : 10
where the dierent error e(k) = e(k) e(k 1).
Now, the update rule for W1 can be written as The training is continued until the following stop-
ping criterion is reached
wij1 k 1 wij1 k 2akek
rWij1 : 11
iX
50

Similarly, it can be shown that PIk Re; or krImax ;

ki
@ek ^
 rWij0 j mj gij
jwij1 xi wij0 jgij 1 12 where e>0 and Imax is the maximum number of ite-
@wij0
rations allowed. In the present implementation
and the update rule for W0 can be written as Imax=2400.
If, after completion of training, the performance of
wij0 k 1 wij0 k 2akek
rWij0 : 13 the model is not found to be satisfactory, the weights
are reinitialised, the number of rules is increased and
The update rule for gij is derived as follows:
the above-mentioned procedure is repeated.


@ek @ ya k yk @yk

@gij @gij @gij
2 3
@ 4 X R YN

j exp jwij1 xi wij0 jgij 5 3.2. Model optimisation
@gij j1 i1
0 1
After achieving a suitable level of performance over
B YN
C @
j @ exp jwij1 xi wij0 jgij A: the entire training range, the following optimisation
i1
@gij1 procedure is followed to optimise the size of the net-
i6j
2 3 work:
X
R
i. The rule-nodes that produce an output close to
4 exp jwij1 xi wij0 jgij 5 zero over the entire training set are sought. Obviously,
j1 these nodes do not contribute signicantly to the


j mj
jwij1 xi wij0 jgij
loge jwij1 xi wij0 j : model output. Hence, these rule-nodes can be safely
omitted without aecting the model performance.
Hence, ii. Again, inputs for which the fuzzy membership is

unity or close to unity over the entire training set, do
rgij j mj
jwij1 xi wij0 jgij
loge jwij1 xi wij0 j 14 not play any role in the actual model. These input
and variables can be traced down, and are omitted from
the network.
gij k 1 gij k 2akek
rgij : 15 The model is tested without any retraining and, if
the performance is still satisfactory without the above-
Hence, the learning algorithm can be summarised as mentioned rule-nodes and inputs, then a model with
wij0 k 1 wij0 k 2ak:ek:rwij k; reduced size has been obtained. This makes the net-
work computationally more ecient. However, if the
wij1 k 1 wij1 k 2ak:ek:xi k:lawij k; performance of the network is not found to be suit-
gij k 1 gij k 2ak:ek:blagij k; able, then the network has to be retrained with the
j k 1 j k 2ak:ek:mj k; 16 reduced set of rules and inputs. In the simulations
described here 5 input-nodes (M) and 16 rule-nodes
where (R) were ultimately retained for the nal model.
 P. Dash et al. / Engineering Applications of Articial Intelligence 11 (1998) 307316 311

3.3. Simulation results and evaluation 2 31=2

1 X0
Tt

While operating in a real-time environment, it is im- SD 4 %errt MAPE 5

2
T tt0
perative that the load forecasting system should be
able to adapt to changing conditions. This objective is where T is the number of time samples, and the
achieved by the following methodology.
MAPE is computed over the same range of t. The err1
For daily adaptation, the optimised weights are used
is the absolute error relative to the peak load for the
from the training set to forecast the rst day's load.
day.
After the end of the day of the forecast model, par-
Abnormal weather and system conditions, such as
ameters are updated until the error becomes insigni-
thunderstorms or transmission outages, are treated as
cantly small. The number of iterations required for
abnormal events with bad real-time readings, and are
this purpose is extremely small. Once the new weights
have been established, the forecast for the next day is not considered in the forecasting models. The inuence
attempted, using this weight vector and the new set of of standard holidays is also considered in the real-time
input data for the day. forecast, and is treated separately, along with week-
For identifying the parameters of the load model, ends.
data related to weekdays (Monday through Friday) The special holiday data occurring in the past and
are treated separately from the weekends (Sundays or the weekend data are used to train the model before a
Saturdays). For this purpose the data related to week- prediction for a holiday is attempted. The collection of
ends are separated from the load database for the pur- similar days from the past ensures that the character-
pose of identication of the parameters of the weekday istics of only that type of holiday are reected in the
load model during network training. In a similar way, data set.
the parameters of the weekend load model are ident- The performance of the proposed model has been
ied by using the appropriate load database. Similar compared with those of some existing neural tools and
load patterns are used during the training of the net- time-series approaches [1, 1822]. It has been found
work to obtain a faster convergence. that this method performs better than the other
As the weekends are excluded from the rst set of models, especially those that use back-propagation as
the database, the weight vector obtained after the fore- the training algorithm. To evaluate the performance of
cast of a Friday's load is used to predict the load on the proposed network architecture, it was used to fore-
Monday. For one-week-ahead forecasts, the adap- cast both one-day and one-week-ahead peak and aver-
tation is done once a week, i.e. at the end of the week, age loads.
when the entire load prole for the whole week is This section also describes an example of the
available. authors' early work, which uses a fuzzy neural network
However, until some appreciable change in the load (FNN) modelled as a ve-layer feed-forward neural
pattern is encountered, the network is left to operate network [22]. The input vector to the FNN consists of
in prediction mode. Suppose that at the kth time-step it dierences in weather parameters between the present
is observed that the error in prediction is showing a
and forecast instant. The output of the FNN gives the
gradual increasing trend; this indicates a substantial
load correction which, when added to the past load,
change in the load pattern. In order to reduce the
gives the forecast load. Fig. 2 shows the proposed
error within the desired performance level, the network
model of the FNN using the ANN architecture. The
is retrained using available data up to the (k 1)th
FNN has a total of ve layers. Nodes at layer one are
time-step. After training, the network is again used for
the input linguistic nodes. Layer 5 is the output layer
forecasting.
and consists of two nodes, one for the actual load cor-
The mean absolute percentage error (MAPE) is used
to test the performance of the model, and is dened as rection eLC and the other the desired load correction
follows: eLC. Nodes in layers two and four are term nodes
which act as membership functions to represent the
 X
1 N term sets of the respective linguistic variables. Each
MAPE
N i1 node in layer three represents the preconditions of the
  rule nodes, and layer-four links dene the conse-
jforecasted load actual loadj  100 =actual load;
quences of the rules.
where N is the number of patterns in the date set used In the simulations, data from a utility in the state of
to evaluate the forecasting capability of the model. Virginia, USA are used. A mathematical software
The standard deviation (SD) of the absolute relative package, MATLAB, is used for obtaining load fore-
error is as follows: casts.
312 P. Dash et al. / Engineering Applications of Articial Intelligence 11 (1998) 307316

to abrupt variations in the weather pattern in the sum-

mer season.
In some utilities a week-ahead forecast (168-hours)
is also desired. The 168-hours-ahead predictions over
four weeks in winter and summer are therefore pre-
sented in Fig. 4(a) and (b), respectively. These results
are marginally better than those from the 24-hours-
ahead forecasts. This reinforces the fact that the load
patterns of similar day types bear a greater resem-
blance. For example, a Tuesday's load pattern is more
similar to that of the previous Tuesday than to that of
the immediately preceding Monday.
The 24-hour average load forecasts over four weeks
Fig. 2. Load forecasting model using a fuzzy neural network (FNN).
in winter and summer are shown in Fig. 5(a) and (b)
respectively, and the 168-hours-ahead average load
4. Load forecasting results forecasts are given in Fig. 6(a) and (b). It is observed
that the average load forecast is better than the peak
The 24-hours-ahead peak load prediction over four load forecast. This phenomenon is due to the aver-
weeks in winter is shown in Fig. 3(a). The correspond- aging eect of the load pattern.
ing percentage of error (PE) is shown in Fig. 3(c). It is The comparison of the percentage of errors (PEs)
observed from these gures that the predicted output between the ANN, FNN and SFNN models for aver-
closely follows the actual load pattern, and the error is age load forecasting is shown in Fig. 7(a) and (b).
mostly within 23%. Similarly, the 24-hours-ahead Fig. 7(a) gives the PE for 24-hours-ahead forecasts in
peak load forecast and its corresponding error values the month of January (winter) and Fig. 7(b) gives the
for summer season are shown in Fig. 3(b) and (d). comparison for 168-hours-ahead forecasts in the
From Fig. 3(a)(b) it is clear that the prediction in month of June (summer). Similarly, Fig. 8 shows the
winter is better than in summer. This can be attributed comparison for 168-hours-ahead peak load forecasts in

Fig. 3. 24-hours-ahead actual vs predicted peak load, (a) Winter (b) Summer; corresponding percentage of error(c) Winter (d) Summer.
 P. Dash et al. / Engineering Applications of Articial Intelligence 11 (1998) 307316 313

Fig. 4. 168-hours-ahead actual vs predicted peak load, (a) Winter (b) Summer.

Fig. 5. 24-hours-ahead actual vs predicted average load, (a) Winter (b) Summer.

Fig. 6. 168-hours-ahead actual vs predicted average load, (a) Winter (b) Summer.

the month of January. The gure shows the improved days are within a PE value of 1.0, 80% within a PE of
performance achieved by the SFNN model in terms of 1.5 and 90% within a PE of 2.0, which is much better
faster convergence and improved overall accuracy, in than the results obtained by Mohammed et al. [8],
comparison with the ANN and FNN approaches. Khotanzad et al. [19] and Bakirtzis et al. [21].
The results in Fig. 9 show the percentage of the Similarly, the percentage of the days of each week-
number of days on which dierent week-days lie day type, having dierent levels of megawatt (MW)
within a certain percentage of error over the whole error (dierence between actual and predicted load), is
year. Here, in the results for 24-hours-ahead peak load shown in Fig. 10. The results in Fig. 10(a) show that
forecasts, it is seen thatout of all the days, cate- almost 50% of the days of any week-day type lie
gorised into individual week-day types70% of the within a 20 MW error range, 70% are within 40 MW
314 P. Dash et al. / Engineering Applications of Articial Intelligence 11 (1998) 307316

Fig. 7. Comparison of PEs between ANN, FNN and SFNN models in average load forecasts; (a) 24 hours ahead (b) 168 hours ahead.

Fig. 8. Comparison of PEs between ANN, FNN and SFNN models Fig. 9. Percentage of days on which a particular weekday had dier-
in peak load forecasts; 168 hours ahead. ent PE values over a year.

Fig. 10. Percentage of number of days on which a particular weekday had dierent Mega Watt errors over a year. (a) 24-hours-ahead peak load
forecasts, (b) 168-hours-ahead average load forecasts.

and 90% within a 60 MW range for 24-hours-ahead 24-hours-ahead as well as 168-hours-ahead forecasts
peak load forecasting. For 168-hours-ahead average for 50 weeks of a year. It can be observed from these
load forecasting almost 80% of the days for any given gures that the weekly MAPE values for average load
weekday type fall within a 20 MW range. This is forecasts mostly occur within MAPE values of 0.05 to
shown in Fig. 10(b). 0.15, and for peak load within 0.5 to 1.5, which can be
Fig. 11(a) and (b) show the mean absolute percen- compared with results obtained by Mohammed et al.
tage of error (MAPE) calculated over each week for [18] and Khotanzad et al. [19].
 P. Dash et al. / Engineering Applications of Articial Intelligence 11 (1998) 307316 315

Fig. 11. Occurrences of weekly MAPE values; (a) Peak load (b) Average load.

Fig. 12. Error distribution of peak load forecast over a year, (a) 24 hours ahead and (b) 168 hours ahead.

Fig. 13. MAPE and standard deviation for each month over a one-year period for (a) 24-hours-ahead peak load forecast, (b) 168-hours-ahead
average load forecast.

The FNN results in terms of megawatt error distri- obtained using BP (back-propagation) in an ANN by
bution for all the days over a year is shown in Fig. 12. Mohammed et al. [18].
Fig. 12(a) and (b) show the results of both 24-hours- The MAPE and standard deviation (SD) of each
ahead and 168-hours-ahead peak load forecasts, re- month over a year are shown pictorially in Fig. 13. As
spectively. Taking the total number of days as 100, it shown in the gure, the maximum value of the MAPE
was found that this result is much better than those is below 1.4% for 24-hours-ahead peak load forecasts
316 P. Dash et al. / Engineering Applications of Articial Intelligence 11 (1998) 307316

(Fig. 13(a)), and the MAPE is within 0.16% for 24- Peng, T., Hubele, N.F., Karady, G., 1993. An adaptive neural network
hours-ahead average load forecasts (Fig. 13(b)). This is approach to one-week ahead load forecasting. IEEE Trans. on
PWRS vol. 8 (no. 3), 11951202.
signicant in comparison to other approaches e.g.,
Widrow, B., Lehr, M.A., 1990. 30 years of adaptive neural networks :
those of Dash et al. and Paraman et al. [22, 23]. perceptron, madaline, and backpropagation. in Proc. IEEE vol. 78
(no. 9), 14151492.
Zadeh, L., 1973. Outline of a new approach to the analysis of complex
5. Conclusion systems and decision process. IEEE Transaction on Systems, Man
and Sybernatics vol. SMC-3 (no. 1), .
This paper has addressed the problem of load fore- Hsu, Y., Ho, K., 1992. Fuzzy expert system: An application to short-
term load forecasting. IEE Proceeding-C vol. 139, 471477.
casting using a self-organising fuzzy neural network
Rahman, S., Bhatnagar, R., 1988. An expert system based algorithm
(SFNN) structure, which gave a signicant improve-
for short-term Load forecast. IEEE Trans. on PWRS vol. 3 (no.
ment in performance in comparison to ANN and 2), 392399.
FNN models. The weights in the dierent layers of the Lin, Y., Cunningham, G.A., May. A new approach to fuzzy-neural
network are optimised using a novel updating algor- system modelling. IEEE Transactions on Fuzzy Systems vol. 3
ithm. The network performs satisfactorily, starting (no. 21995), 190198.
from an initial set of random weights. Thus the pro- Horikawa, S., Furuhashi, T., Uchikawa, Y., 1992. On fuzzy modeling
blem of a proper choice of the initial weights is using fuzzy neural networks with the back-propagation algorithm.
IEEE Trans. Neural Networks vol. 3 (no. 5), 801814.
avoided. The ecacy of the network is validated by
Hayashi, Y., Bukley, J., Czogala, E., 1993. Fuzzy neural network with
selecting a practical load pattern in summer and win- fuzzy signals and weights. Int. J. Intell. Syst. vol. 8, 527537.
ter. As the summer load pattern is extremely tempera- Ishibuchi, H., Fujioka, R., Tanaka, H., 1993. Neural networks that
ture-sensitive, the peak load forecast demonstrates the learn from fuzzy if-then rules. IEEE Trans. fuzzy Systems vol. 1
network's eciency. The weekly MAPE value is (no. 2), 8596.
mostly within 0.51.5% for 24-hours-ahead peak load Schneider, A.M., Takenawa, T., Schiman, D.A. 1985. 24-hour elec-
forecasting, and within 0.050.15% for average load tric utility load forecasting. Comparative models for electric load
forecasting, which is a very good result in comparison forecasting. chapter 7. John Wiley & Sons Ltd. pp. 87108.
Fan, J.Y., McDonald, J.D., 1994. A real-time implementation of short-
to those obtained by BP neural networks.
term load forecasting for distribution power system. IEEE PWRS
vol. 9 (no. 2), 988994.
Mohammed, O., Park, D., Marchant, R., Dinh, T., Tong, C. et al.,
References 1995. Practical experiences with an adaptive neural network short-
term load forecasting. IEEE Trans. PWRS vol. 10 (no. 1), 254265.
Hargan, M.T., Behr, S.M., 1987. The time series approach to short- Khotanzad, A., Hwang, R., Abaye, A., Maratukulam, D., 1995. An
term load forecasting. IEEE Trans. on PWRS vol. PWRS-2 (no. adaptive modular articial neural network hourly load forecaster
3), 785791. and its implementation at electric utilities. IEEE Trans. PWRS
Box, G.E., Jenkins, G.M., 1976. Time Series Analysis Forecasting and vol. 10 (no. 3), 17161722.
Control. Holden-Day, San Fransisco. Dash, P.K., Liew, A.C., Rahman, S., 1995. Peak load forecasting using
Papalexopoulos, A.D., Hesterberg, T.C., 1990. A regression based a fuzzy neural network. Electric Power System Research 32, 1923.
approach to short-term system load forecasting. IEEE Trans. on
Bakirtzis, A.G., Theocharis, J.B., 1995. Short term load forecasting
PWRS vol. PWRS-5 (no. 4), 15351544.
using fuzzy neural networks. IEEE Trans. PWRS vol. 10 (no. 3),
Chen, S.T., Yu, D.C., Moghaddamjo, A.R., 1992. Weather sensitive
15181524.
short-term load forecasting using nonfully connected articial neural
network. IEEE Trans. on Power System vol. 7 (no. 3), 10981105. Dash, P.K., Satpathy, H.P., Rahman, S., 1995. Short term daily aver-
Peng, T.M., Hubele, N.F., Karady, G.G., 1992. Advancement in the age and peak load predictions using a hybrid intelligent approach.
application of neural networks for short-term load forecasting. IEEE, EMPD-95, Singapore, Catalogue No. 95TH8130.IEEE vol.
IEEE Transactions on PWRS vol. 7 (no. 1), 250257. 2, 565570.
Park, D., El-Sharkawi, M., Marks, R., Atlas, L., Damborg, M., May. Paarmann, L.D., Najar, M.D. 1995. Adaptive online load forecasting
Electric load forecasting using an articial neural network. IEEE via time series modeling. Electric Power System Research. Elsevier
Trans. on Power Systems vol. 6 (no. 21991), 442449. Science S.A. 32, pp. 219225.