Beruflich Dokumente
Kultur Dokumente
Constantin Barbulescu, Stefan Kilyeni, Attila Simo, Oana Pop, Cosmin Oros, Raluca Schiopu, Antheia Deacu
Power Systems Department
Politehnica University of Timisoara
Timisoara, Romania
constantin.barbulescu@et.upt.ro
Abstract The authors are focusing on developing a software tool
designed for power consumption forecasting based on artificial
neural networks (ANN). The hourly and peak consumed power and
daily load curves are forecasted. The backpropagation algorithm
is presented within the paper. Also, some practical considerations
are highlighted, necessary to be known to develop a real application.
The base propagation algorithm is completed with conjugated
gradient and parabolic interpolation. A software tool has been
developed in Matlab environment. The results are compared with
the ones provided by classic forecasting methods. The analyses
and the corresponding conclusions highlight the fact that the ANN
based forecasting algorithm has a better behaviour.
Index Terms ANN, backpropagation, consumed power, daily load
curve, forecasting.
I.
INTRODUCTION
yh kj f ( xi w1k 1 ),
ij
i 1
nh
y kj f ( xij w2k 1 ),
ij
i 1
j 1, nh
(1)
j 1, no
(2)
ok
i
e si
(1 e
sik 1 2
Eeik , i 1, no
(4)
ss k , i 1, nh
(5)
k 1
kh
i
e shi
(1 e
shik 1 2
where: ss k ss k 1 ok w2k 1 , j 1, no .
j
ij
k 1
gwih
kh
ij
i
t j ; i 1, nh, j 1, n
ij
Steps 3-6 are repeated for each pattern within the ANN
training set.
7. Conjugated gradient method:
gwihk ij
nh n
k
ih
i 1 j 1
nh n
i 1 j 1
k
gwho
ij
kho
i 1 j 1
nh n
i 1 j 1
(9)
k 1 2
gwho
ij
(12)
(8)
k 1 2
gwih
ij
no nh
optim
(6)
k
k 1
gwho
gwho
ok yh kj ; i 1, no, j 1, nh (7)
ij
RESULTS. DISCUSSIONS
k
k
k
k 1
wdih
gwih
ih
wdih
; i 1, nh, j 1, n (10)
ij
ij
ij
k
k
k
k 1
wd ho
gwho
ho
wd ho
; i 1, no, j 1, nh (11)
ij
ij
ij
(10)
k
w2ijk w2ijk 1 wd ho
; i 1, no, j 1, nh
(11)
ij
ij
10
11
12
13
14
15
16
17
18
19
20
21
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
122.0
134.0
127.2
114.9
116.0
91.1
110.9
108.9
117.0
112.6
113.6
126.9
119.5
107.7
108.7
83.5
102.6
100.6
107.5
104.8
111.4
123.5
116.2
104.8
105.2
81.3
100.9
99.2
102.4
100.6
111.9
121.8
115.4
104.3
103.3
80.0
98.0
99.7
100.2
102.6
113.6
123.0
115.6
106.5
104.8
81.9
99.4
99.0
101.5
102.2
120.8
131.4
122.5
113.5
109.5
87.7
103.9
103.6
104.6
102.6
139.3
151.4
138.7
127.2
119.9
97.9
115.3
114.6
115.0
116.0
160.2
172.7
156.5
143.3
137.0
111.6
134.4
135.2
134.7
135.7
162.8
177.0
161.7
150.1
147.0
120.2
144.9
143.1
144.0
144.5
161.6
166.0
160.5
148.5
148.6
122.8
146.5
142.6
145.1
145.3
154.6
161.1
153.4
143.4
147.5
121.4
143.9
141.1
143.7
143.3
152.5
160.4
152.4
145.9
147.3
123.7
145.1
143.5
146.2
146.4
148.3
156.3
149.2
143.5
145.1
122.7
143.8
142.6
146.2
145.4
148.0
155.6
148.1
144.3
143.4
123.7
143.0
143.6
145.1
146.9
146.8
151.2
145.6
141.1
138.9
121.5
141.9
141.3
144.1
143.1
135.5
148.6
136.4
131.1
129.2
116.1
135.3
134.2
137.0
135.4
128.9
147.8
133.1
126.5
123.5
111.9
131.3
127.4
132.3
129.7
127.7
147.4
128.9
123.6
122.4
110.6
127.8
125.7
129.3
126.7
126.4
148.7
127.2
122.3
120.1
109.3
125.3
121.4
126.7
124.7
133.3
156.2
131.3
125.4
119.8
111.0
123.3
120.6
125.0
122.7
158.4
174.4
148.2
140.8
131.0
120.5
128.3
125.5
133.0
127.3
22
23
24
TABLE II. REAL DAILY LOAD CURVES [KW] CORRESPONDING TO THE FORECASTED PERIOD
Year / Hour
10
11
12
11
111.4
104.9
102.8
99.8
100.9
103.0
116.5
137.4
147.6
151.1
151.2
154.0
12
118.5
109.2
103.9
103.2
104.3
105.0
115.3
133.8
145.0
148.4
148.0
152.0
13
127.9
116.3
111.7
110.4
111.3
113.8
122.7
139.9
147.9
151.3
150.8
153.0
TABLE II. REAL DAILY LOAD CURVES [KW] CORRESPONDING TO THE FORECASTED PERIOD (CONTINUATION)
Year / Hour
13
14
15
16
17
18
19
20
21
22
23
24
11
153.0
152.8
149.8
143.2
136.2
133.3
131.7
133.7
136.6
145.1
145.2
128.3
12
150.7
150.3
149.7
143.6
138.9
134.9
131.5
132.3
133.8
142.0
154.2
140.6
13
151.6
151.8
149.9
143.5
139.6
136.9
133.7
133.1
136.5
145.7
155.3
140.5
Prel
Pi Pavg
Pi
100, i 1; 24
(1)
For the 03-04 and 22-24 hours a miscorrelation is highlighted in case of the 6th year. The same conclusion is obtained
in case of 10-16 hours, for the 2nd year. Thus, it is concluded
that these curves have an acceptable correlation; they are not
strong correlated.
The ANN based algorithm for daily load forecasting has
been developed in Matlab environment. The ANN toolbox has
10
11
12
11
112.9
105.5
104.4
99.0
102.1
106.2
124.1
134.1
139.3
143.9
141.8
151.4
12
114.6
106.5
106.2
100.4
103.3
107.8
126.7
137.1
140.0
145.0
143.3
154.9
13
116.5
107.7
108.1
101.9
104.6
109.5
129.7
140.4
140.9
146.2
145.0
158.5
13
14
15
16
17
18
19
20
21
22
23
24
11
144.8
143.6
145.7
143.1
135.4
133.8
128.1
126.3
134.3
145.0
150.7
127.8
12
148.0
146.5
147.9
135.4
136.8
137.2
129.8
128.4
136.2
146.3
153.4
128.7
13
151.3
149.6
150.2
136.8
138.3
140.8
131.6
130.8
138.2
147.9
156.4
129.7
1st hour
3rd hour
5th hour
7th hour
9th hour
11th hour
13th hour
15th hour
17th hour
19th hour
21st hour
23rd hour
122.0
134.0
127.2
114.9
116.0
91.1
110.9
108.9
117.0
112.6
111.4
123.5
116.2
104.8
105.2
81.3
100.9
99.2
102.4
100.6
113.6
123.0
115.6
106.5
104.8
81.9
99.4
99.0
101.5
102.2
139.3
151.4
138.7
127.2
119.9
97.9
115.3
114.6
115.0
116.0
162.8
177.0
161.7
150.1
147.0
120.2
144.9
143.1
144.0
144.5
154.6
161.1
153.4
143.4
147.5
121.4
143.9
141.1
143.7
143.3
148.3
156.3
149.2
143.5
145.1
122.7
143.8
142.6
146.2
145.4
146.8
151.2
145.6
141.1
138.9
121.5
141.9
141.3
144.1
143.1
128.9
147.8
133.1
126.5
123.5
111.9
131.3
127.4
132.3
129.7
126.4
148.7
127.2
122.3
120.1
109.3
125.3
121.4
126.7
124.7
158.4
174.4
148.2
140.8
131.0
120.5
128.3
125.5
133.0
127.3
157.7
169.1
156.9
150.0
147.9
139.8
147.8
143.6
139.9
149.3
Backpropagation ANN has been also used. Its configuration has been established as follows:
TABLE V. HOURLY LOAD FORECASTED VALUES [KW] 2ND DEGREE POLYNOMIAL FUNCTION
Hour
Year
1st hour
3rd hour
5th hour
7th hour
9th hour
11th hour
13th hour
15th hour
17th hour
19th hour
21st hour
23rd hour
150.3
156.5
163.9
130.3
134.0
138.3
130.9
136.4
143.0
130.4
135.4
142.1
147.7
150.8
154.7
150.84
157.87
166.63
136.83
142.91
150.49
131.22
137.29
144.94
133.87
140.07
148.42
147.59
150.56
154.47
118.4
124.0
130.8
104.5
108.9
114.3
105.8
111.0
117.4
119.8
125.6
133.0
150.8
157.9
166.7
148.8
154.4
161.3
151.1
156.7
163.3
118.94
125.37
133.68
104.96
109.92
116.41
106.25
112.07
119.70
119.87
125.81
133.80
151.11
158.85
168.97
149.16
155.37
163.29
151.63
157.89
165.70
All these figures are highlighting that the ANN based forecasting provides the best results compared to the real values. Greater
deviations are recorded in case of the classic forecasting methods.
C. Peak consumed power forecasting
The peak consumed power forecasting has been performed
for 9 (morning peak power) and 21 hours (evening peak power).
3 data sets have been used corresponding to the 2001-2006
period. Within the paper only 1 set is presented (Table VI).
TABLE VI. PEAK CONSUMED POWER [KW] KNOWN DATA
Data 26
26
26
26
26
26
Hour
Year
July August September October November December
10
9.5
16.5
18
21.5
12
09:00
2001
12
12.5
16
17
21
15
21:00
15
12.5
13
15
15.5
11
09:00
2002
15
15
14
15
17
13.5
21:00
14.5
11
13
10.5
15
7
09:00
2003
12.5 11.5
12.5
15
16
10.5
21:00
14
11.5
6
12.5
14.5
5
09:00
2004
12
13
10.5
14
14.5
9
21:00
12
10.5
11
10.5
14.5
5.5
09:00
2005
12.5 10.5
12
12
12
8.5
21:00
2006
10
10
8.5
9.5
12
10
11.5
12.5
8.5
13
7.5
10
09:00
21:00
Evening peak power data correlation is continuously improving. In case of morning peak power 90 % maximum value
is recorded; for September month it decreases.
It significantly decreases for the 3rd month.
ANN used for peak consumed power forecasting has the
same type and configuration as the one presented in case of
point B. The only difference is represented by the ANN
training set. 6 patterns have been used in this case.
The forecasted peak consumed power values using classic
methods are presented in Table VII.
TABLE VII. PEAK CONSUMED POWER [KW] CLASSIC FORECASTING
Hour
9:00
21:00
9:00
21:00
26
26
26
26
26
26
July August September October November December
2nd degree polynomial function
4.80 5.70
15.45
13.60
8.40
9.40
7.95 7.05
10.70
11.75
13.10
11.10
2nd degree exponential function
6.54 6.47
16.83
13.21
7.61
9.45
8.44 7.75
10.36
11.68
12.58
10.80
9:00
21:00
26
26
26
26
26
26
July August September October November December
10.76 7.80
13.32
12.64
12.44
6.43
10.97 9.10
13.96
13.93
12.40
8.96
IV.
CONCLUSION