Professional Documents
Culture Documents
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
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.