Soft Computing Techniques for Short

Term Load Forecasting

Carolina Tranchita, Member IEE ,Alvaro Torres, Senor Member IEEE
Department of Electrical Engneering, Universidad de los Andes - Bogota, Colombia
b8N Short term load forecasting is a recurrent topic in
the operatve planning activities of companies dedicated to the
distrbuton and trade of energy around the world due to the
compettive electrcity market, in which an advantage in the
previous kowledge of demand could mean the diference
between obtaining big benefts or incur in economic losses.
In this paper a novel method for short term load forecasting is
proposed based on the similar day approach and the use of Sof
Compnting techniques. This approach is founded on the search
for the most similar day in history, to the forecasted day, based
on the exlanatory meteorologcal varables forecast for the load
of this day. Once the "similar" day is found, the load forecast
will be the same of that day with an adjustment for load growth.
LAA-fz-clusterng techniques, regression trees, CAT
classifcation and fz inference for the peak power, daily
energy and load cure forecast are used.
The validation of the proposed method is made with
meteorologcal and load data from a Colombian city.
KeordShort term load forecast, Peak Power and Daily
Energy Forecast, Similar Day Approach, LAA, CAT.
I. INTODUCTION
T
he art of load forecasting, particularly short tenn,
became one of the top gowing areas of study and
sigifcance since te emergence of competition in the
energy markets. Aer the privatization and deregulation
tendency of power systems in many countries, including
Colombia among them, the purchase and sale of energy in
this new maket constitute the base of business for the
distibution and electicity trade companies.
For these companies it is vital to have good kowledge, or to
obtain a good forecast of the short ter load with the purose
of fxing prices in advance, according to their service supply
costs. This contributes to these companies savings, for
exaple in Great Britain, the savings were in the order of US
$30 millions for a 4% improvement in load forecasting [ 1].
Diverse metods for short term load forecasting have been
tested with dfferent degees of success. There are two main
tendencies of models for short term forecast: the causal or
conventional models, which outline the relationships among
load and factors infuencing it, mainly social and weather
variables; and the time series models which outline the
relationships with values observed in the past.
The model presented in this paper is based in te simila day
approach and a forecast for the load curve, the peak power
and the daily energy consumed is achieved.
0-780J-8718-X/04/$20.002004.
The "similar day approach has been used in forecasting
applications where the dependent variable is higly correlated
with the meteorological variables and the calendar location.
The method proposes to seach in history a ver similar day
based in some kown explaining independent variables,
generally of climate and time (day of te week, month, etc.)
and to predict the dependent variable, in this case the electic
energy demand, in accordance to what had happened to this
variable in the similar day found. It is important to note that
in this method there are no mathematical fnctions found for
the explaining variables.
The hourly electric load in a Colombian coast located
company and the hourly temperature of the coresponding
city are used for the research development. The data used for
methodology validation go fom Janua 1998 to June 2000.
II. STAT OF TH AT
Conventional methods in short-ter load forecasting ae
among others: regession models, AM, AM or
SAM models [2], Box and Jenkins transfer fnctions
[3], optimization techniques [4], non-paametric regession,
structural models, curve adjustment procedures and tere are
some more recent methods whose results ae compaed with
the ones mentioned such as expert systems [5], artifcial
neural networks [6], fzy inference [7] and Neuro-fz y
models [8].
The load to forecast in the system is a random non-stationa
process composed by tousands of individual components.
Therefore, the range of possible approaches to the forecast is
wide. Usually te only possibility is to take a macroscopic
view of the problem, and tr to model the fture load as a
refection of previous behavior. This still leaves open space
to very different solutions. Due to the nature of load, the only
objective method to evaluate the approaches is experimental
evidence [9].
Expert systems have been also used in load forecasting.
These systems are heuristic models, which usually ca
include both quantitative ad qualitative factors. A typical
approach is tring to imitate the reasoning of a human
operator. The idea ten is to reduce the analogical thougt
behind the intuitive forecasting to formal logic steps. A
popular approach has been to develop rules based on fuz
logc [9].
Another important technique in short term load forecasting is
the neural network. In these models, the network inputs
generally are present and past load values ad the outputs ae
fture load values. The network is trained using real load
data fom te past [10].
There have also been suggested models with no-supesed
learing for load forecasting. The purpose of these models
can be the classifcation of days in different types of day, or
to choose the most appropriate days in histor to be used as
base for the real load forecasting.
. MODEL PROPOSED
In the method proposed in this paper peak power and daily
energy forecasts ae made in order to predict te load curve,
lookng for the most similar day in history based on the
meteorologcal forecasts for tomorow's day. U is achieved
using classifcation and identifcation with LAMDA ad
CART methodologies, obtaining and implementing a rule
based Fuzzy Inference System and te similar day approach.
A. STAGE .Data Analysis
The data used for methodology validation go fom Janua of
1998 to June of 2000. For a more complete proof, load data
of a longer period of time would be prefered.
Some cyclical efects for the load data, see Figure 1, can be
observed starting fom te workng day; the weekend rhym
is followed by most people. In labor days social activities are
in a higer level than in Saturdays and Sundays, therefore the
load is also higher. The series begins with fve quite similar
patters tat are the Monday to Friday load cuves. Next two
different patters continue for Satuday and Sunday. Then
this same weekly patter is repeated [ 10].
Hourly behavor of load September 2000
550000
,
I
500000
450000

II
I l( V lc
400000
350000
IJ II .t l| I( \1 f If \ f I! I
|| _' '|
1i
+ ''
I '
300000
250000
I
1 36 71 10 141 176211 246281 316 351 386421 456491 526
Hours
Fig. I. Monthy demnd of enegy during Septemer 2000.
On the other hand the daily rhym takes place because of the
synchronous behavior of people dung the day. Most of
people slee at night; consequently the demand is low during
night hours. Also during the day, many activities tend to be
simultaneous for most of people (labor schedule, lunch hour,
etc.). The daily rhythm changes throughout te year [10].
There are also differences between days of the sae month or
meteorological condition. Therefore, in load forecasting days
are ofen divided in numerous tes of da, each has its own
characteristic load patter. Satrdays and Sundays have
different load cuves tan other days. Ofen Mondays andor
Fridays also sepaate fom other labor days, because the
proximity of weekend can have a sligt effect in load [10].
In the proposed model the explanatory variables used for the
forecast were: te day of the week, the type of day (classifed
as labor, weekends and holidays) ad te meteorological
conditions of the day to which the load is to be predicted.
D. STAGE I: LAMDA to determine the similar day
The similar day technique is founded on te search for the
most similar day in histor to tomorow, or to the day to
predict, based on the explanatory meteorologcal variables for
the electic power consumption.
The objective is that with the meteorological forecasting for
tomorow, or for the day to predict, the most similar day in
history can be found. Some sofware progams, available in
the market, look for the most similar day using clustering
methods with neual networks.
In the work developed the LAMDA (Learing Algorithm
Multivariate Data Analysis) classifcation method is proposed
to be used with this aim. LAMDA was chosen because it
shows some advantages that oter classifcation systems don't
have. One of tese is deterning the number of clustes,
without being the user who decides te input space partition,
and this is not excessive. Another interesting feature is that it
can be chosen whete the supervised or the not supervsed
classifcation is used [11].
LAMDA is a fz classifcation method that is rooted in the
adequacy degee analysis of each object to a class, concet
related with the membership degee to a fzy set. The
contibution of each one of the attributes or individual's
describers, to te membership degee to a class is called
marginal adequacy (M, Marginal Adequacy Degee).
These marginal adequacies are combined with the use of
"Fuzzy Mixed" operators to finally fmd the adequacy degee
of the individual to each one of the classes (GAD, Global
Adequacy Degee) [11].
As no patter was kown to determine similar days it was
necessa to car out frst not supervised classifcation, so
that later supervised classifcation could be done in such a
way that when entering a new day (individual), the algorithm
was able to identif at least one (1) most similar day.
The "leaing" consists on fmding parameters defming each
one of the classes' describers. The type of parameters and the
way to fmd them deends on whether the work is done with
quantitative or qualitative describers, and the means to
determine them depends on whether a supervised learing is
being caried out or not [11].
A non information class exists (NIC=Non Informative Class),
which determines te minimum threshold of global adequacy
that an individual should have in order to belong to some
Z
class. In the case of supervised leag, the individuals that
don't overcome te GAD of the NIC class ae said not to
belong to any of the classes. For not supervised leaing,
starting fom the NIC paeters and the indvidual that was
not classifed, a new class is generated. So starting fom only
one goup, all te individuals are fmally classifed locating
them in goups tat are generated depending on the GAD's.
For the similar day identifcation, the IDEAM (Colombian
institution for meteorology) made possible the access to their
meteorological database. Basically data of years 1998, 1999
and 2000 were obtained. However there were problems wit
te database as it was not flly completed and the variables
measures are not hourly taken. Therefore te variables used
for obtaining the similar day are the following: average
temperature and average humidity in te moring, aferoon,
and nigt of te next and current day; solar shine in the
moring and aferoon of the current day; average speed of
the winds in the morg, aferoon, and night of the cur ent
day; maximum and minimum speeds of te winds in the
curent day and maximum and minimum temperatue in the
next and curent day.
Once the variables were chosen, tests on the not supervised
classification were made with the sofware elaborated by the
contol goup of the Universidad de Los Andes. Several
simulations with different demand values ad distributions
were made using 900 of te base of 1088 data for training,
because in this type of problem there is one seasonal patter
per year and it was important to t to embrace all of the
possible cases.
Aer the tests it was determined that the best dstribution for
classifng data was te normal when the demand parameter
is 0.39 since the number of classes is not excessive and the
classification is more homogeneous that in other cases. The
result of not supervsed classifcation is shown in Figure 2.
Fig.2. Classification of the days according to meteorological conditions.
L STAGE 1Forecast for peak power and daily energ
1) Part a: Search of relevant variables and obtaining ofif
then Rules uing CART
Once made the daily load curve analysis it was decided to
forecast te peak power and the daily energy, other important
data such as te load factor ad the minimum power were not
forecasted since tey are quite constant with te test data.
CART is a classifcation and regession tees algoritm. H
makes supervised classifcation ad it can fnd the decision
rules of the cluster and in the application for load forecasting
it was used for this purose only afer the classifcation with
LAMDA was completed.
P the clusters had already been determined, it was necessa
to deterine whethe all the variables were relevant, or if it
was better to add other non meteorological variables in order
to obtain better fnctions of Takagi Sugeno type for the
defzcation. This way, several models were suggested
depending on meteorological vaiables, te day of the week,
the type of day ad the variable to predict in previous times.
With CART it was found that the meteorological variables,
the type of day and the day of the week were enoug to make
the peak power and energy forecasts. When autoregessive
models were used, that is, the output in previous times, te
error obtained was similar to te model that only uses
meteorological data but the more autoregessive the outut
was the less necessary the meteorological vaiables were.
Once defmed the clusters and variables, if-then rules were
also found for the fzzy inference system using CART.
However CART only provides Takagi - Sugeno rules of zero
order, which was not the optimum for the problem. This was
solved using least minimum squared on each cluster as later
explained.
P example, in the decision tree obtained for peak power one
of the rules obtained is:
I the dy tpe uweknd or holiday anJthe average speed of winds of the
current da in the aferoon u more than 4.4J anJ the humidit of the
curent day ules than or equal to 4.JanJthe temperature of te net da
in the night uless than or equal to 20.J then the peak power of the net da
is of 4J0/J.
2)Part b: Takgi- Sugeno (T S) Rule A4ustmet using least
minimum squared regessions
With the CART algorithm the classifcation rules for each
cluster were found but of zero order (a constant value),
therefore it was necessa to make an adjustment to tat
regession. We proceeded to make the regessions using least
minimum squared.
There were found excellent adjustments in several clusters
that exceeded all te statistical tests, this means, te t test tat
indicates te adjustment tolerance of each variable, the F test
that indicates the global adjustment tolerance of the model
and the value of R squared that indicates how much can the
model forecast. However, there were also found clusters
where the value of R squaed did not exceed 0.6 (fairly good)
and the value of F was baely more than te critical F, but in
spite of this in the geat majority of cases the typical eror of
adjustent was less than in the case where te rue was lef
with a constant value. In the cases where the defz cation
d
of the rule is constant, it is because the quantity of cluster
data was too small as to use least minimum squared.
Also in some of te clusters it was found that a fairly good
model with the explanatory variables did not exist, so that a
division of the cluster had to be made depending on the day
type or the day of te week, for example, te cluster 23 had
the following rule:
I the day of the wek is Monda or Tesday or Wednesday or Thursda or
Frida anJthe average speed of the wind of the current da is more than to
J.ZJand the temperature in the night of the net day is les than or equal to
2Z.!Jand . rkenthe daily energ of the net day is 404/+000.
P ca be seen, there was no diference whether te day was
Monday, Tuesday, Wednesday, Thursday or Friday. When a
least minimum squared was intended, none of the statistical
tests was satisfactor; therefore more explanatory variables
were used to improve the model adjustent. For this purpose
binary variables or dunnnies were maipulated to represent
the day of the week because it was kown that this variable is
relevant for the load forecasting. In this case there was better
adjustment makng these distinctions althoug it incured in a
geater division of the input space.
For example node 23 changes fom 1 to 4 rules, as follows:
* ^0Je2Ja:if the da of the week is Monda and the average speed of
the winds of the curent da umore than J.ZJand ...... .
* ^0Je2Jb:i the da of the wek is Tuesday and .... .
* ^0Je2Jc: i the day of the week is Wednesda and ... .
* ^0Je2JJ:i the da of the wek is Friday and the average speed of
the winds of the curent day is more than J.ZJand ..... .
In total tere were about 350 tests made to fnd 87 fnctions
for the consequents of the TS rules.
3) Part c: Implementation of the Fuzz Inference System
Once the Takag - Sugeno rules ad the number of
membership fctions were obtained the following step was
the implementation of the FIS. To create the fzy inference
system the "fzy" fction of the Fuzzy Logic toolbox in
MATLAB 6.1 was used.
The system has in total 25 inputs, for the 7 days of the week,
the three day types (labour, holiday or weekend), and te 15
meteorologcal variables. Since the day of the week and the
type of day ae concrete values the membership fnction tat
better fts is the singleton, not yet implemented in the
toolbox, therefore a very narow triangular fnction was used.
Bell fctions were used to implement meteorological
variables and fnction parameters were estimated using least
minimum squaed estimators. With the two FIS, for peak
power and energy, the forecast of these variables was made.
D. STAGE II: Integration of the fuz clustering and FIS
In Stage I, with the use of LAMDA meteorologically similar
days are found and in Stage II forecasts for peak power and
daily energy peak are made. Now the user will have to
choose which the most similar day in histor is ad make the
gowth adjustments to be able to forecast te load curve of
the day to forecast following the fu diagam in Figure 3.
From 2 to 5
meteorologcally
similar days
-
i

Are these days reasonabJe?
Poity orer to follow:
I. Type of day
2. Day of wWeek
1.Vacations?, Cris1as?
4. Moth of wYea
Growh adjustment
The crve is raised to the
peak power forecast a
maximum, without exceeding
m-1%or+ 1%t Energ
forcast.
1
-

_.
I ' "" "" . . ", .. """.
y
1
t
ad curve fo
morow
Fig.3. Deision diag for lod cue forecast.
TA
TST CASE: SIA DAYS
4
Table I shows te result of te meteorologically most similar
days to a test day, on this case Wednesday October 18 of
2000. It was found according to the GA of the data in the
class tat the most similar days are: Friday August 3 of 2000,
Friday May 5 of 2000, Tuesday November 23 of 1998 and
Thursday October 14 of 1998 . We observe that the standard
deviation in most of the variables for the similar days is low.
Figure 4 shows one variable of those with which the search of
the similar day was made. P it can be seen these variables
ae quantitatively very near. Figure 5 shows the load curves
of the similar days obtained based only on the meteorological
variables. These load curves are similar evidencing that the
meteorological variables have explanation power for the
electic power consumption.
35,0
Similar days in temperature to october Zof
Z
30.0 +
::::
15,0

10,0
5,0
0,0
Temasig Tetasig Tenoig
Morning- Ahernoonand Night
Fig.4. Temeratures of similar days to October 2 according to the GA.
Once te similar days were identifed the meteorological
data, the day type and the day of the week were entered to the
fzy inference system for te short term load forecast.
Table II shows the result when the Takagi - Sugeno rules
consequent of frst order were used.
However there is still error due to the small database, since it
has not been possible to characterize all the different types of
days. One can observe that the days December 25 are non
typical days that could not be well characterized as
consequence of few leag data.
Load cre of mterlogiclly
similar days to Noveber 8of Z000
55000
. ....... . ..
50000
A- .
45000

.7 ., .' `
40000
- .
"""'<

'""
35000
`

30000
25000
20000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 2 21 22 23 24
Hours

Ma 211/99 -W e 30/699 -Tu 28/999 Tu 6/6/00 Tu 59/00 '

Fig.5. Load cues of the meteorologically similar days to Noy/ 8/2000.
Later, to fnd te demand curve, te most similar day is
chosen between the similar days found by fz clustering,
for this it should be kept in mind the day of te week, the day
type and the calendar date, for example, if it is Christmas day,
it has priority a similar day of one earlier Christmas that other
days of the year.
Makng the forecast of load curve for example for Thursday
November 24 of 2000, the most similar days foud were:
Tuesday June 2 of 1998, Friday June 5 of 1998, Thursday
July 9 of 1998, Tuesday July 27 of 1999 ad Wednesday
November 10 of 1999. Having as a priority the similarity of
the day of the week, it is chosen as similar day the day
Thursday July 9 of 1998, load data can be seen in Table .
Keeping in mind the criteria of moving up the cuve to the
peak power, without passing above 3% of the daily energy
forecasted, the curve goes up in a delta calculated as follows:
_ Ppeak forecasted-Ppeak of simi/ar day
Z
The result of the adjusted curve is shown in Figure 6. For
this day it can be observed that the error in the peak power
forecast was 0.55% and in the energy demand was 2.74%.
When the similar day approach is used for the demand curve
forecast it has an error of 2.96% on average, and the for of
the curve is similar.
TAII
FORECAST FOR POWER PEA USIG T T AGJ - SUGENO RULES CONSEQUENT FNCTIONS OF FIT ORER
Real Value Forecast Value
Error Error
Pleak Enerey PPeakDay Enerey PDeak Enerey
Sunday I 10 2000 489001 8652600 477568 8584621 2338% 0.786%
Wednesday 18 10 2000 509341 9711360 506233 9685412 0,610% 0.267%
Monday 30 10 2000 523222 9494236 525333 9547410 0.403% 0,560%
Sunday 5 11 2000 422995 7583608 433889 7935265 2575% 4,637%
Wednesday 8 11 2000 502626 9367717 501353 9404157 0.253% 0.389%
Wednesday 23 11 2000 501449 8910666 497966 9023493 0.694% 1.266%
Thursday 24 11 2000 520910 9331410 523782 9587424 0,551% 2.744%
Saturday 2 12 2000 462336 8559706 434077 8262351 6112% 3.447%
Sunday 3 12 2000 475611 8288801 487116 8447687 2419% 1.917%
Thursday 7 12 2000 533879 9564478 504708 9395214 5464% 1.77%
Tuesday 19 12 2000 536531 9330476 528959 9295416 1411% 0.376%
Monday 25 12 2000 447482 8143619 469725 8375258 4971% 2.844%
b
W
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

21
22
23

'
-
TALE III
R8U18
5 Em

3.70 3 3.10
331U.3 31.76
3374.71 31189.11
32703.03 337547.43
3352.55 345534.95
34702.28 357576.68
33376.72 3 .12
3.3 353920.77
376.44 3.b
377155.95 37670.35
3574.66 4 .0
3b.71 393211.11
. 32549.64
37618.63 386703.03
40149.84 41264.24
406.15 418176.55
3740.51 34611.91
30 .43 4079.83
4328.41 443319.81
4 .&
4. 474459.04
45572.40 466.80
408.86 418347.26
340376.38 3.78
WO
39131.0
331Z1.4
32490.11
31075.0
38.48
3594.10
353.55
37.&
30 .0
35215.5
374.5
376.64
376.40
3
.5
32172.6
319.
33.14
452.15
453581.47

561.0
48128.61
4b.3
399.9

W
3.S,
2.S"
4.53,
4.63,
4.70,
0.495
0.281%
0.89 ,
0.28
3.0Z'
2.235
1.01%
3.935
0.01%
4.717
5.S,
2.27,
1.10,
2.19
3.431%
5.8
3.0,
1.81
7.827
2.0
eaIandadjustedcurvesto november24 ot2000
550000 T
500000
450000
400000
350000
300000 {
250000
1 2 3 4 5 6 7 8 9 1011121314151617 18 19202122 23 24
Adjusted cure -Real cure
Fig.6. Real and adjusted (forecast) load curves.
IV. CONCLUSIONS
A novel method is show to make short tenn demand forecast
using the bases of fzy logic and statistical methods. This
method allows making the forecast of the whole demad
curve conty to others in which the curve is forecasted by
means of an hourly forecast.
CAT classifcation ad regession trees allow identifng
the relevant variables for STLF without using autoregessive
models in the load. In the test case, the variables used were
temperature, relative humidity and speed of the winds.
Based on data analysis support rules can be generated for the
expert that show the reasoning line followed by the model. Is
a tool for peak power and daily energy forecast integated to
the similar day approach throug the fzy inference system.
The model developed can be applied to ay dynamic mime
system in which forecast is necessary to be made. However,
a limitation of the proposed method is te propagation of
error because a high degee of meteorological forecast is
needed.
V. BmLIOGRAPHY
[I] Ku Anne, "The A of Foecasting Demand", Global Energy Business
March April 2002, pp. 20-23
[2] Karanta I. y Ruusuen J., 1991, "Short term load forecastng in comunal
electic ulities", Research report A40, Systems Analysis Laboratory,
Helsink University of Technology
[3] Hagan M. y Klein K, "Identfcation Techniques of Box and Jenks
Applied to the Problem of Short Term Load Foecasting, IEE PES SM,
Papr A 77 168-2, Mexico City, Mexico, Jul. 1977
[4] Toyoda, J., M.-C. Chen, y Y. Inoue, 1970, "An application of state
estimtion to short-term lod forecastng, Pa I: Forcstng modeling, Part
II: imlementton", IEE Transactions on Powe Apparatus and Systems,
VOL. PAS-89, No. 7, September/October 1970, pp. 1678-1688
[5] Raman, S., Bhatagr, K, 1988, "An expert system based algorithm for
short term load forecast", IEEE Transactions on Power Systems, Vol.3, No.2,
May 1988, pp.392-399
[6] Park, D. C., M. A EI-Sharkawi, K J. Marks II, L. E. Atlas, M. J.
Damborg, 1991a, "Electic load forecasting using an artfcial neual
network", IEE Transactions on Power System, Vol. 6, No. 2, May 1991,
pp. 442-449
[7] H. Mori y H. Kobayashi, "Optmal Fuzy Infernce for Short Term Load
Forecastng", IEE Trans. on Power System, Vol. II, No. I, pp. 390-396,
Feb. 1996
[8] Ajit, A. yBalkunth, N., 2001, "A Neuo Fuzzy Approach fo modelling
electicity demand in Victoria". Monash University. Churchill, Austalia
3842
[9] Murto P, 1998, "Neual netwok models fo short-term load forecastng"
Helsinki Univesity of Technology, Department of Engneering Physics and
Mathematics, Finland, January, 1998
[10] Khotanzad, A, Abaye, A. y Maratukulam, D., 1997, "A Neual
Network Based Electic Load Forecastng Systems", IEEE Transactons on
Neual Networks, July 1997
[II] Gauthie, A., Isaza, C. V., Martinez, H. "Desarrollo de una her ent
comutacional pclasifcaci6n e identficaci6n de sistemas complejos con
tecnicas difsas". CIFI, Universidad d los Andes, Bogota, Junio 2003
N. BIOGRAPHY
Avaro Torres Macias: Eletical Engneer, UlS. Master in Electical
Power Eng. Master in Systems and Computers Eng. Ph.D in Electical
Engneering, Rensselaer Univesity, U.S. Specialist in Energy Transport,
ENSEM, France. Professo and Investigator of the Electic and Electonic
Departnt at the Univesidad de los Andes. Tehnical Manager of
Consultoria Colombiana. Work a: planning, modeling simulatng,
optmization, line and distibuton systems desig, sofware and informaton
systems.
Carolina Tranchita Rativa: Electical Engineer, Master in Electical
Engineering and Docorate stndent in Eletical Engineering at Universidad
de lo Andes. Professo at the Universidad de los Andes. Wok Areas:
planning, reliability and expert systems application in power systems.