Professional Documents
Culture Documents
Furthermore, if the simple size is big enough, the last In some experiments, factors can have different levels. The
hypothesis can be relaxed (Central Limit Theorem). Hence, interaction between them can mask the principal effect if this
the observations represented by Eq. (1) can be described by interaction is considerable. Therefore, before reaching any
the following statistical lineal model [9]: conclusion, this interaction effect must be studied.
X ijk = µ + τ i + β j + γ ij + ε ijk (2) To determine the interaction between two factors, the
following hypothesis testing is performed:
where µ is the total media, τi denotes the effect of level i of the
first factor, βj the effect of level j of the second factor, γij the H 0 : γ 11 = γ 12 = ... = γ ab = 0
interaction effect between level i of the first factor, and level j (5)
H1 : at least one γ ij ≠ 0
of the second factor, and εijk the experimental error.
Moreover, the experimental observations related the
statistical model described by Eq. (2) can be formulated as considering the following relation [9]:
follow:
( ) (
x ijk = x + x i • − x + x • j − x + ) (3)
Λ* =
SSPres
SSPint + SSPres
(6)
( ) (
x ij − x i• − x • j + x + x ijk − x ij ) where Λ* is known as Wilks’s Lambda. The hypothesis H0
where x is the total sampled mean vector, x i• represents the is rejected if Λ* is less than a given threshold. Normally this
sampled mean vector that correspond to the first factor level i, test is done before testing the effect of the principal factors. In
x • j is the sampled mean vector of the second factor level j, the case that the interaction test fails, then no additional test is
performed.
and x ij is the sampled mean vector due to the first factor level
Once this test is passed, the effect of different factors is
i and the second factor level j. calculated using the following hypothesis testing are
The generalized variance estimation can be represented by: performed:
a b n a
H1 : at least one IJ i ≠ 0
SSPcorr SSP1
b
+ ¦ ®¯an(x
j =1
•j −x )(x • j − x )t ½¾¿ H 0 : β1 = β 2 = ... = β b = 0
H1 : at least one β j ≠ 0
(8)
SSP2
(4)
a b the statistical value used for the hypothesis testing
+¦¦ (
i =1 j =1
n x − x − x + x x − x − x + x t ½
® ij
¯
i• •j ij i• •j )(
¾
¿
) described by Eq. (7) and Eq. (8) is similar to the one described
by Eq. (6):
SSPint SSPres
a b n
Λ1* = (9)
SSP1 + SSPres
+¦¦¦ (
i =1 j=1 k =1
x −x x −x t½
® ijk
¯
ij ijk ij ¾
¿
)( )
SSPres
SSPres Λ 2* = (10)
SSP2 + SSPres
where SSPcorr is the total sum of squares, SSP1 is the sum of
squares for the main effect due to the first factor, SSP2 , is the
sum of squares for the main effect due to the second factor, * *
the hypothesis H0 is rejected if Λ 1 and Λ 2 are below a
SSPint is the sum of the interaction, and SSPres is the sum of threshold. Once the multivariate analysis indicates that an
squares of the errors. effect exists, comparisons must be drown to detect specific
The test procedure is arranged in an multivariate analysis of
variance (MANOVA) table, shown in Table I.
differences. There are several comparison methods, such as
the Bonferroni method or Hotelling method [9].
A. Data normalization There are residuals that are bigger than the rest of the
The first step is to verify if the data satisfy the statistical residuals; these residuals are identified as outliers. Since the
lineal model hypothesis of Eq. (2). The error, also referred as presence of one outlier may alter the analysis of the results,
residual, is defined as the difference between the observation before any analysis, the origin of these outliers should be
value X ijk and the estimated value X ij : determined. A simple technique to identify outliers is to
examine the accumulative distribution of the normalized
residuals. In this example a normalized residual that is outside
εˆijk = X ijk − X ij (11) the region (-3,3) is assumed as a potential outlier
The observation of Fig. 2 and the normal distribution test
proposed by Shapiro & Wilks [9], suggest that market prices
The accumulative distribution residuals graphic is used as
are not normal distributed. Normal test results from the WEM
the initial diagnostic tool [9]. The residuals are normally
market price at each hour are described in Table II
distributed with zero mean; therefore its representation should
(Appendix).
be a line. As an illustration, Fig. 1 shows the simulation results
Although the normality hypothesis testing is not satisfied,
for Monday, January, hour 24.
the number of outliers is not severe enough to stop the
analysis.
B. Hypothesis testing
The effect of the interaction among different factors is
performed using the following hypothesis testing:
H 0 : γ 11 = γ 12 = ... = γ 12 7 = 0
(12)
H1 : at least one γ ij ≠ 0
SSPres
Λ* = = 0.005746 (13)
SSPint + SSPres
H 0 : IJ1 = IJ 2 = ... = IJ12 = 0
(14)
H1 : at least one IJ i ≠ 0
H 0 : β1 = β 2 = ... = β 7 = 0
(16)
H1 : at least one β j ≠ 0
then:
SSPres
Λ* = = 0.400154 (17) Fig. 3: Confidence intervals for months. Market Prices ($/MWh)
SSP2 + SSPres
C. Treatment effect
On condition that the hypothesis described by Eq. (14) and
Eq. (16) are rejected, then, it is important to identify the
factors that influence the rejection. Comparisons are
performed in order to identify specific differences among days
and months. The effect of months and days on the market
price is estimated by Eq. (18) and Eq. (19) respectively:
τˆ i = x i • − x ; i = 1,...,12 (18)
βˆ j = x • j − x ; j = 1,...,7 (19)
H 0 : µ L (k ) − µ D (k ) = 0 k = 1,...,24 (22)
SEPTEMBER 44,12393 0,844094 42,46238 45,78548
Table VII illustrates the test results. from which it can be OCTOBER 37,41957 0,831207 35,78339 39,05575
inferred that market prices during Sundays are significantly NOVEMBER 43,69457 0,844094 42,03302 45,35612
different compared to market prices during Mondays. Table DICEMBER 47,25071 0,831207 45,61453 48,88690
VIII shows the estimations (95%) related to the TABLE IV: PRICE MEAN , HOUR 12, DAILY
differences µ (k ) − µ (k ); k = 1,...,24 . Similar tests
M S
µ (12) Error LI LS
DAY
can be done for all other days, for example Table IX and X ($/MWh) ($/MWh) ($/MWh) ($/MWh)
show results from a comparison between Wednesday and
MONDAY 43,15354 0,641423 41,89094 44,41615
Thursday, concluding that results for both days are
TUESDAY 43,05650 0,641423 41,79389 44,31911
comparable. WEDNESDAY 43,00608 0,641423 41,74348 44,26869
Analogous tests can be produced between different month THURSDAY 43,04496 0,641423 41,78235 44,30756
types. For example a test to compare December and July can FRIDAY 42,43746 0,641423 41,17485 43,70006
be done as follows:
SATURDAY 41,60292 0,635671 40,35164 42,85420
SUNDAY 40,52708 0,641423 39,26448 41,78969
H 0 : µ Dic (k ) − µ Jul (k ) = 0 k = 1,...,24 (23) TABLE V: MONTHLY PRICES, H OUR 12
µ (12)
In this case, from the results shown in Tables XI and XII, it MONTH 1 2 3 4 5
($/MWh)
can be deduced that market prices for these months are
different. The same analysis for April and May is illustrated in OCTOBER 37,41957 x
Tables XIII and XIV, in this case the fact that prices are JUNE 39,14564 x x
related can be inferred. JULY 40,58321 x x x
MAY 41,56300 x x
V. CONCLUSION
MARCH 42,03057 x x x
The application of the methodology of analysis of variance APRIL 42,01564 x x x
of multivariate data (MANOVA) for competitive electric JANUARY 42,25836 x x x
market has been explored. This technique has been applied to FEBRUARY 43,33500 x x x
the Argentinean electric market. NOVEMBER 43,69457 x x x
Results suggest that statistical multivariate methods may SEPTEMBER 44,12393 x x x
become very useful for Market Regulators to evaluate the
AUGUST 45,42871 x x
impact of different factors on the determination of economic
DECEMBER 47,25071 x
parameters such as the market price.
TABLE VI: D AILY PRICES, HOUR 12
TABLE XIV: CONTRAST ESTIMATION- MONTHLY PRICES Diego Moitre (A’97, SM’03) obtained his M. Sc from
Pontificia Universidad Católica, Santiago de Chile. Currently
Hour
k ( )
µ Apr k − µ May k ( ) Hour
k ( )
µ Apr k − µ May (k ) he is associate professor and Dean of the College of
($/MWh) ($/MWh) Engineering, National University of Rio Cuarto. He is also
1 0,169429 13 0,652214 director of the Analysis of Power Systems Group (GASEP)
2 0,231429 14 0,608929
from the same university. His main research activities are
3 0,697286 15 0,563643
4 0,838143 16 0,665143
mathematical modeling and electric market design.
5 0,769000 17 0,360000 Fernando Magnago (SM IEEE) obtained his M. Sc and Ph.D.
6 0,691786 18 0,517500
from Texas A&M University, joined Nexant in 2000, he was
7 -0,213786 19 -1,20157
8 0,730857 20 -0,875357
previously a research engineer at the Institute of Electrical
9 0,904429 21 -0,943857 Power System Protection, Rio Cuarto, Argentina. Currently he
10 1,125429 22 -1,05721 is associate professor at the National University of Rio Cuarto,
11 0,720714 23 -0,997643 and works as software developer for Nexant.
12 0,452643 24 0,774786
VII. REFERENCES