Forecast For Domestic Electric Power Expenses
作者:
系級:
學號:
D9765656 D9765660 D9765554 D9765906 D9765851
開課老師:
課程名稱:
開課系所:
開課學年: 99 學年度 第 二 學期
AREMOS
1990
1
2010
12
252
12
ARIMA
MSE MAD MPE MAPE
Abstract
Abstract
Abstract
Abstract
In order to better understand electric power expenses in Taiwan, we study “the
domestic energy use - electric power” which was obtained from the AREMOS
economic statistics information database. The time period is from January 1990 to
December 2010, total 252 observations. We save the last 12 observations for
out-of-sample forecasts. We apply four kinds of statistical methods/models for
forecasting, which included the time series regression,
exponential smoothing
,
decomposition method
, and the ARIMA model to model this time series. In order to
evaluate the forecast performance, we employ MSE, MAD, MPE, and MAPE four
criteria to evaluate the performance of the four methods. Based on MAPE criterion,
we found that the
decomposition method with X11 and regression technique has the
best performance (about 2.3%) among the four methods which
is able to effectively
forecast domestic electric power to expend in Taiwan for the future .
Keywords
ARIMA; Electric power expense; out-of-sample forecast;
Decomposition Method; X11, Exponential Smoothing.
5
8
9
13
13
15
15
(Time Series Regression)
16
(Decomposition Method)
20
(Exponential Smoothing)
24
ARIMA
25
30
32
33
2-1.1
9
3-1.1
15
3-2.1
log
16
3-2.2
17
3-2.3
95%
(
) 19
3-3.1
20
3-3.2
20
3-3.3
21
3-3.4
21
3-3.5
21
3-3.6
22
3-3.7
95%
23
3-4.1
24
3-4.2
95%
95%
25
3-5.1
25
3-5.2
log
26
3-5.3
26
3-5.4
ACF
PACF
27
3-5.5
ACF
PACF
27
3-5.6
ACF
PACF
28
3-5.7
29
3-5.8ARIMA
95%
(2010
)
30
1-1.1
99
7
2-3.1 MAPE
14
3-4.1
(
)
24
3-5.1 ARIMA
29
3-6.1
31
235
238
...
(
)
1
(
)
0.62
2007
2
24
(2008)
28
C
9.6
60W
11W
36.8
13.5
POLO
(2008)
1-1(
99
)
AREMOS
1990
1
2010
12
252
12
ARIMA
MSE MAD
MPE MAPE
2-1.1
(
)
AREMOS
Excel
(
)
(KLOE)
1990
1
2010
12
252
252
240
12
2010
1
2010
12
ARIMA
2010
ARIMA
MSE MAD MPE
MAPE
2-1.1
/
ARIMA
2010
(
)
1.
(Trend
TR )
t(Season
SN )
t(
ε
t)
t t t tTR
SN
y
=
+
+
ε
2.
(Decomposition Method)
(Trend;
TR )
t(Season;
SN )
t(Cyclical;
CL )
t(Irregular;
IR )
t(
y )
t ty =f(
TR ,
tSN ,
tCL ,
tIR )
tX11
ty =
TR +
tSN +
tCL +
tIR
t ty =
TR
t×
SN
t×
CL
t×
IR
t3.
Exponential Smoothing
Winters Method‐Additive
Winters
Method‐Multiplicative
Damped Trend
Winters Method‐Additive
m s t tm t m t s t t t t t t t t t t s t t tS
b
L
F
S
L
Y
S
b
L
L
b
b
L
S
Y
L
+ − + − − − − − − −+
=
−
+
−
=
−
+
−
=
+
−
+
−
=
)
(
)
1
(
)
(
)
1
(
)
(
)
)(
1
(
)
(
1 1 1 1 1δ
δ
γ
γ
α
α
Winters Method‐Multiplicative
m s t tm t m t s t t t t t t t t t t s t t tS
b
L
F
S
L
Y
S
b
L
L
b
b
L
S
Y
L
+ − + − − − − − − −+
=
−
+
=
−
+
−
=
+
−
+
=
)
(
)
1
(
)
/
(
)
1
(
)
(
)
)(
1
(
)
/
(
1 1 1 1 1δ
δ
γ
γ
α
α
Damped Trend Exponential Smoothing
1 1 1 1
)
1
(
)
(
)
)(
1
(
− − − −Φ
−
+
−
=
Φ
+
−
+
=
t t t t t t t tb
L
L
b
b
L
Y
L
γ
γ
α
α
4.
ARIMA
ARIMA
Box-Jenkins
.
ARIMA
3
(1)
(
AR
) (2)
(
MA
) (3)
(
ARIMA
) ARIMA
ARIMA
(1)
ARIMA
(2)
(3)
(4)
AR(P)
Z
t=
φ
1Z
t−1+
φ
2Z
t−2+
...
+
φ
pZ
t−p+
a
tMA(q)
Z
t=
a
t−
θ
1a
t−1−
θ
2a
t−2−
...
−
θ
qa
t−qARMA(p,q)
q t q t t t p t p t t tZ
Z
Z
a
a
a
a
Z
=
φ
1 −1+
φ
2 −2+
...
+
φ
−+
−
θ
1 −1−
θ
2 −2−
...
−
θ
−:
1.
(White noise test)
,
,
H0
H1
P‐value >0.05
H0
P‐value<0.05
H0
H0
P‐value >0.05
2.
(Unit Root Test)
H0
(
)
H1
(
)
P‐value>0.05
H0
P‐value < 0.05
H0
H0
P‐value <0.05
3.
(1)
(2)
(3)
DW
*
:
Pr<DW
=0.05
Pr>DW
=0.05
12
n
y
ti
yˆ
tt
(
)
2 1ˆ
1
∑
=−
n t t ty
y
n
%
100
ˆ
1
1×
−
∑
= n t t t ty
y
y
n
%
100
ˆ
1
1×
−
∑
= n t t t ty
y
y
n
1.
(Mean Absolute Deviation, MAD)
0
MAD =
∑
=−
n t t ty
y
n
1ˆ
1
2.
(Mean Squared Error, MSE)
MSE = RMSE= MSE
3.
(Mean Percentage Error, MPE)
MPE =
4.
(Mean Absolute Percentage Error, MAPE)
MAPE =
MAPE
:
MAPE(%)
評估
<5
高準確的預測
6-10
尚可的預測
>10
不準確的預測
2-3.1 MAPE
ARIMA
2010
MSE MAD MPE
MAPE
3-1.1
3-1.1
2008
2009
3-1.1
7
9
12
2
3-1.1
2008
9
2008
2009
1
3-2.1 log
3-1.1
log
3-2.1
log
DW
Pr<DW
0.05
t tM
M
M
M
M
M
M
M
M
M
M
t
y
ε
+
+
+
+
+
+
+
+
−
−
−
−
+
=
∗ 11 10 9 8 7 6 5 4 3 2 10658
.
0
1327
.
0
1853
.
0
2015
.
0
1533
.
0
0112
.
0
0699
.
0
00722
.
0
0146
.
0
01232
.
0
0305
.
0
004557
.
0
3744
.
14
)
log(
* t ty
y
=
t t t t t
=
0
.
358025
ε
−1+
0
.
291779
ε
−2+
0
.
275252
ε
−3+
a
ε
)
,
(
~
2 . . .σ
µ
N
a
d i i tσ
ˆ
=
0
.
003742
Mt
12
12
95%
:
3-2.23-2.2
95%
2009
1
:
1.2008
7
9
2.2008
7
3.2008
9
2009
1
2008
9
2008
9
(
227
)
=
0
1
if
if
t
t
<
≥
227
227
t
s
DW
Pr<DW
=0.05
12
12
95%
:
3-2.3 95%
(Decomposition Method)
3-3.1
3-3.1
3-3.3 3-3.4 3-3.5
3-3.2
3-3.3
3-3.4
3-3.5
DW
Pr<DW
0.05
Pr<DW
Pr>DW
0.05
12
12
95%
:
3-3.6
95%
2008
9
DW
Pr<DW
=0.05
12
12
95%
:
3-3.7 95%
3-4.1
3.4-1
Winters
Method‐Multiplicative
3-4.13-4.1
12
12
95%
3500000 4000000 4500000 5000000 5500000 6000000 6500000
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
公 秉 油 當 量 公 秉 油 當 量 公 秉 油 當 量 公 秉 油 當 量
月份
月份
月份
月份
ACTUALPredicted value for POWER Upper 95% Confidence Limit Lower 95% Confidence Limit
3-4.2 95% 95%
3-3.2
95%
ARIMA
3-5.13-5.1
3-5.2 log
3-5.3
3-5.2
3-5.4 ACF PACF
3-5.4
ACF
(
)
lag12
lag24
3-5.5
ACF
MA(q)
ACF
lag1
lag12
(q=(1)(12))
Airline model
3-5.6 ACF PACF
log
ARIMA(0,1,1)(0,1,1)s NOINT
3-5.6
ACF
PACF
3-5.7
3-5.7
(
)
lag
=0.05
H0
lag
=0.05
H0
ARIMA(0,1,1)(0,1,1)s NOINT
3-5.1 ARIMA3-5.1
P-value
=0.05
ARIMA(0,1,1)(0,1,1)s NOINT
(1‐B)(1‐B
12) log(
Yt
)= (1-0.68274B)(1-
0.80104B
12)a
t~
(
,
)
2 . . .σ
µ
N
a
d i i tσ
ˆ
=
0
.
00149
12
12
95%
:
3500000 4000000 4500000 5000000 5500000 6000000 6500000Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
公
秉
油
當
量
公
秉
油
當
量
公
秉
油
當
量
公
秉
油
當
量
月份
月份
月份
月份
ACTUALPredicted value for POWER Upper 95% Confidence Limit Lower 95% Confidence Limit
3-5.8 ARIMA 95% (2010 )