Beruflich Dokumente
Kultur Dokumente
allocation,
genetic
I. INTRODUCTION
( Pi Pj ) =
i, j k
n n Rij cos(i j )
= (
( Pi P j + QiQ j ) +
ViV j
i =1 j =1
Rij sin(i j )
+
(Qi P j PiQ j )
ViV j
(1)
(2)
O2 = PDG .
where PDG is the DG installed active power, in kW and
represents the investment cost. The tests in this study were
performed with a value for of 950 UDS/kW [2].
Optimality is achieved when the solution represents a
compromise between network benefits and capital
investments, and therefore the objective function is constituted
by two contradictory objectives. In order be able to
mathematically aggregate the two objectives of different
natures, the first one is also transformed into an economical
factor:
TABLE I
COMPUTATIONAL RESULTS WITH THE NONLINEAR OPTIMIZATION ALGORITHM
Pi = PDG, i PDi =
n
= U i [U k [Gik cos(i k ) + Bik sin(i k )]]
k =1
Qi = QDi =
(4)
(5)
(6)
(7)
Losses
[kW]
Comp. time
[s]
83.4252
260.3594
61
1794
84.233
797.1875
62
1794
n
= U i [U k [Gik sin(i k ) Bik cos(i k )]]
k =1
Si Simax
U imin U inom U imax .
PDG PDGmax .
No. of
DG
units
B. Genetic Algorithm
Before using any of the GA models, the problem must be
represented in a suitable format that allows the application of
genetic operators. GAs work by optimizing a single entity, the
fitness function. Hence, the objective function and the
constraints of the problem at hand must be transformed into
some measure of fitness.
Encodings. The first feature that should be defined is the
type of representation to be used, so that an individual
represents one and only one of the candidate solutions. A
candidate solution (or chromosome) designed in this paper for
the problem of finding the optimal location and size of one
DG unit is a two-component vector (Fig. 2,a).
Position(node number)
Unit 1
Position (node
number)
Unit 1
Size (max. 2
MW)
Unit 2
Unit 2
Position (node
Size (max. 2
number)
MW)
b.
Fig. 2. Chromosome encoding for one DG unit (a) and two DG units (b) to be
allocated
(8)
a.
where the first two terms are the ones in the objective function
and the following are penalty functions. The last three sums in
this fitness function are a measure of unfeasibility for each
candidate solution x. The penalty factors used in this study
were set to 10000.
The constraint expressed in (7) is satisfied each run, as the
limits for each individual are set within the main GA routine.
Details on genetic algorithms implementations can be
found in [8].
The results presented in Table II were obtained for roulette
wheel selection, a crossover rate of 0.7 and the population size
set to 50.
TABLE II
COMPUTATIONAL RESULTS WITH THE GENETIC ALGORITHM APPROACH
No. of
DG
units
Losses
[kW]
Comp.
time [s]
88.21 262.12
61
1500
83.91 864.83
62
861
61
886
73.76 1543.36
62
736
18
519
61
809
C. Solution Analysis
As it results from Tables I and II, the nonlinear
optimization algorithm cannot face the high complexity
problem of allocating more than 2 DG units, in comparison
with genetic algorithms. Even though the losses in the case of
GAs for one DG unit are slightly higher than the ones resulted
with the nonlinear optimization algorithm, the superiority of
GAs is proven when the problem complexity increases and the
nonlinear optimization algorithm fails to provide a solution.
For a better insight of the solutions supplied by the two
methods, a voltage level analysis is made in Figure 3, showing
the voltage levels in the network for one (Fig. 3,a) and two
(Fig. 3,b) DG units the two cases that were solved by both
methods.
The solution provided by the nonlinear optimization
method for placing and sizing one DG unit results in a better
voltage profile than the one in the case of genetic algorithms,
both having all buses with voltages within the admissible strip
(Fig. 3,a).
For two DG units, the genetic algorithm leads to slightly
better voltage levels (Fig. 3, b), proving that genetic
algorithms perform better than nonlinear optimization when
the number of variables increases.
b.
Fig. 3 Voltage levels for solutions given by nonlinear optimization and genetic
algorithms for one DG unit (a) and two DG units (b)
losses [kW]
250
200
150
100
50
0
1
7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67
bus
Fig. 5. Variation of losses with the bus where a DG unit of optimal size is
connected.
170
150
110
90
11
00
12
50
14
00
15
50
17
00
18
50
20
00
21
50
23
00
24
50
26
00
27
50
29
00
80
0
95
0
70
65
0
130
50
0
Fig. 6. Variation of losses with the size, for a DG unit connected to bus 61.
Fig. 7 Voltage deviations when DG unit is installed in buses that lead to losses
higher than the base case
voltage [p.u.]
1
0.99
0.98
to be allocated. The first test was run for one DG unit, and
both methods provided similar results in similar periods of
time, the nonlinear solver having a small advantage regarding
the power losses and computational time. Next, the number of
DG units was increased to two, and the GA proved to provide
better results than the nonlinear solver, in a slightly higher
computational time. After increasing the number of units to
three, the GA outperformed the nonlinear optimization
algorithm, which failed to land a solution.
The network used for tests is the IEEE 69-bus distribution
system, for which the GA results in a DG size of 1500 kW,
installed at bus 61. Connecting this amount of DG at the
optimal bus leads to a power loss reduction from approx. 225
kW to about 88 kW. The voltage profile is substantially
improved as well.
The tight connection between the optimal location and size
is proved by allocating the optimal size at different buses in
the network and by allocating different DG capacities at the
optimal bus resulted from the GA.
Both studies show that system losses increase drastically.
In some cases they become even larger than the ones in the
base case. Furthermore, the voltage profiles are also degraded
if the optimal solution is not implemented.
Even though GAs are very dependent on their parameters,
they provide results when other methods fail. Firstly, the
algorithm is a multi path that searches many peaks in
parallel, and hence reduces the possibility of local minimum
trapping. Secondly, GA works with a coding of parameters
instead of the parameters themselves. Thirdly, GA evaluates
the fitness of each string to guide its search instead of the
optimization function. Hence, there is no need for computation
of derivatives or other auxiliary functions.
In future studies, several improvements should be taken
into consideration: the use of load profiles and assessing the
benefits accordingly to the DG type.
0.97
VI. REFERENCES
0.96
[1]
0.95
0.94
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67
[2]
bus
1 DG unit
2 DG units
3 DG units
[3]
[4]
V. CONCLUSIONS
The paper addresses the problem of optimal DG location
and sizing in a distribution network. Two solution methods are
proposed, one based on a nonlinear optimization algorithm
and one based on genetic algorithms. The objective function
comprises of both power losses and investment costs. The
studies are performed on multiple levels: a comparison
between the proposed approaches and the importance of
installing the right amount of DG in the best suited location.
In order to compare the solution methods, tests were
performed successively for an increasing number of DG units
[5]
[6]
[7]
[8]
VII. BIOGRAPHIES
Ioana Pisic received the B.S. degree in electrical engineering from the
University Politehnica of Bucharest in 2007; she is currently a PhD student
and teaching assistant in the Electric Power Engineering Department in the
same University. Her research interests are related to power systems stability,
FACTS devices and artificial intelligence applications in power systems.
Constantin Bulac (M02) graduated at the Polytechnic Institute of Bucharest
in 1982 and has received the Ph.D. degree in electrical engineering from the
University Politehnica of Bucharest in 1998; currently he is Professor in
the Electric Power Engineering Department in the same University. His
research interests are related to power systems stability, FACTS devices and
artificial intelligence applications in power systems.
Mircea Eremia (M98, SM02) received the B.S. and Ph.D. degree in
electrical engineering from the Polytechnic Institute of Bucharest in 1968 and
1977 respectively. He is currently Professor at the Electric Power Engineering
Department from University Politehnica of Bucharest. His area of research
includes planning and operation of transmission and distribution networks,
power system stability and applications of FACTS devices and artificial
intelligence applications in power systems.