Electrical Power & Energy Systems, Vol. 20, No. 8, pp.

563–569, 1998
q 1998 Elsevier Science Ltd. All rights reserved
Printed in Great Britain
PII: S0142-0615(98)00016-7 0142-0615/98/$ - see front matter

Optimal reactive power dispatch

using an adaptive genetic
algorithm

Q H Wu*, Y J Cao and J Y Wen

Department of Electrical Engineering and Electronics,
The University of Liverpool, Liverpool L69 3GJ, UK
mathematical assumptions, such as analytic and differential
This paper presents an adaptive genetic algorithm (AGA) for
properties of the objective functions and unique minima
optimal reactive power dispatch and voltage control of
existing in problem domains, have to be given to simplify the
power systems. In the adaptive genetic algorithm, the prob-
problem. Otherwise it is very difficult to calculate the
abilities of crossover and mutation, p c and p m, are varied
gradient variables in the conventional methods. Further, in
depending on the fitness values of the solutions and the
practical power system operation, the data acquired by the
normalized fitness distances between the solutions in the
SCADA (supervisory control and data acquisition) system
evolution process to prevent premature convergence and
are contaminated by noise. Such data may cause difficulties
refine the convergence performance of genetic algorithms.
in computation of gradients. Consequently, the optimization
The AGA applied for optimal power system reactive power
could not be carried out in many occasions.
dispatch is evaluated on an IEEE 30-bus power system in
In the last decade, many new stochastic search methods
which the control of bus voltages, tap position of trans-
have been developed for the global optimization problems,
formers and reactive power sources are involved to minimize
such as simulated annealing, genetic algorithms and evolu-
the transmission loss of the power system. q 1998 Elsevier
tionary programming. Genetic algorithms (GAs) are a class
Science Ltd. All rights reserved
of stochastic search algorithms that start with a population of
randomly generated candidates and ‘‘evolve’’ towards better
Keywords: Reactive power dispatch, genetic algorithm
solutions by applying genetic operators (crossover, muta-
adaptive search, global optimization
tion, inversion, etc.), modelled on the genetic processes
occurring in nature. Recently, the evolutionary computation
techniques have found many applications in power systems,
I. Introduction especially in the economic operation area [5–16]. The first
The reactive power dispatch problem has a significant attempt of the application of genetic algorithms in power
influence on secure and economic operation of power systems is in the load flow problem [5]. It has been found that
systems. The reactive power generation, although itself the simple genetic algorithm (SGA) quickly finds the normal
having no production cost, does however affect the overall load flow solution for small-size networks by specifying an
generation cost by the way of the transmission loss. A additional term in the objective function. The economic
procedure which allocates the reactive power generation so dispatch problem has been investigated by both the binary
as to minimize the transmission loss will consequently result and float point based genetic algorithms in Refs [7,10–
in the lowest production cost for which the operation con- 13,15,16]. The evolutionary programming has also been
straints are satisfied. The operation constraints may include applied to the problem of reactive power dispatch [6]. A
reactive power source capabilities, nodal voltages, phase number of approaches to improving convergence and global
angles and transformer tap position. This forms the optimal performance of GAs has been investigated [8,11,12,17,18].
reactive power dispatch problem. The conventional The main difficulty of application of GAs in engineering
gradient-based optimization algorithms have been widely problems is their premature convergence, i.e. the GAs do
used to solve this problem for decades [1–4]. Obviously, this degenerate in many cases to a local optimum instead of the
problem is in nature a global optimization problem which global one. In order to avoid the premature convergence and
may have several local minima and the conventional opti- improve the convergence properties of GAs, the idea of
mization methods easily lead to local optimum. On the other adapting crossover and mutation operators to improve the
hand, in the conventional optimization algorithms, many performance of GAs has been employed [18–24].
This paper proposes an adaptive genetic algorithm (AGA)
for optimal reactive power dispatch and voltage control of
* Corresponding author. Tel: 0044 151 7944 535; Fax: 0044 151 power systems. The proposed algorithm is based on the
7944 540; e-mail: [email protected] recently developed adaptive strategy [24], but introduces a

563
564 Optimal reactive power dispatch: Q. H. Wu et al

normalized fitness distance between the current individual X

and other individuals in the population to control the prob- 0 ¼ QGi ¹ QDi ¹ Vi Vj (Gij sinvij ¹ Bij cosvij ) i [ NPQ
j[Ni
ability of mutation. It is observed that as the GA converges,
the fitness distance between the chromosomes in each and (3)
population will become smaller and smaller. Therefore, it
would be sensible to incorporate the fitness distances Vimin # Vi # Vimax i [ NB
between individuals into the adaptive updating of the prob-
ability of mutation. Application of the AGA to the optimal Tkmin # Tk # Tkmax k [ NT
reactive power dispatch problem of large-scale power
Gi # QGi # QGi i [ NG
Qmin max
systems is discussed in the paper. The performance of the
proposed algorithm is evaluated on an IEEE 30-bus power
Ci # QCi # QCi i [ NC
Qmin max
system. The simulation results show that the AGA converges
to the global optimum in far fewer generations than the where power flow equations are used as equality constraints,
canonical genetic algorithm (CGA). reactive power source installation restrictions, reactive
power generation restrictions, transformer tap-setting
restrictions and bus voltage restrictions are used as
II. Mathematical formulation of optimal inequality constraints.
reactive power dispatch In the most of the nonlinear optimization problems, the
The objective of the reactive power dispatch is to minimize constraints are considered by generalizing the objective
the active power loss in the transmission network which can function using penalty terms. In the reactive power dispatch
be described as follows: problem, the generator bus voltages, V PV and V s, the tap
X X position of transformer, T, the amount of reactive power
fQ ¼ Pkloss ¼ gk (Vi2 þ Vj2 ¹ 2Vi Vj cosvij ) (1) source installations Q C, are control variables which are self-
k[NE k[NE constrained. Voltages of PQ-buses, V PQ, and injected
reactive power of PV-buses, Q G, are constrained by adding
where k ¼ (i,j); i [ N B; j [ N i. The minimization of the them as penalty terms to the objective function (equation
above function is subject to a number of constraints: (1)). The above problem is generalized as follows:
X X X
0 ¼ PGi ¹ PDi ¹ Vi Vj (Gij cosvij þ Bij sinvij ) i [ N0 F Q ¼ fQ þ lVi (Vi ¹ Vilim )2 þ lGi (QGi ¹ Qlim
Gi )
2

j[Ni i[NVlim i[NQlim

(2) (4)

Figure 1. Flow chart of AGA algorithm for optimal reactive power dispatch
 Optimal reactive power dispatch: Q. H. Wu et al 565

where l Vi and l Gi are the penalty factors, Vilim and Qlim

Gi are while some mutation is required to prevent the premature
defined as convergence of the GA to suboptimal solutions. Identifying
( max optimal settings for p c and p m is an important problem for
Vi ; Vi . Vimax
lim
Vi ¼ (5) improving the convergence performances of GAs and has
Vimin ; Vi , Vimin been studied by many researchers [19–24].
( max The key idea of the AGA is to adapt the probabilities of
lim
QGi ; QGi . Qmax Gi
crossover and mutation based on the fitness statistics of
QGi ¼ (6) population at each generation. In Ref. [24], it has been
QGi ; QGi , Qmin
min
Gi observed that the difference between the maximum fitness
value and average fitness value of the population, fmax ¹ f̄ ,
likely to be less for a population that has converged to
optimum solution than that for a population scattered in
III. Adaptive genetic algorithm the solution space. Therefore, the values of p c and p m should
The AGA is developed based on the canonical genetic be varied depending on the value of fmax ¹ f̄ . On the other
algorithm (CGA). The CGA consists of an n-tuple of hand, if p c and p m have the same values for all the solutions
binary strings b i of length l, where the bits of each string of the population, which means solutions with high fitness
are considered to be the genes of an individual chromosome values as well as the solutions with low fitness values are
and where the n-tuple of individual chromosomes is said to subjected to the same level of mutation and crossover, this
be a population. In a multiple variable optimization problem, will certainly deteriorate the performance of GAs. The
the individual variable coding is usually concatenated into a adaptive strategy for updating p c and p m developed in Ref.
complete string. To decode a string, bit strings with specified [24] takes the following forms
(
string length are extracted successively from the con- k1 (fmax ¹ fc )=(fmax ¹ f̄ ), fc . f̄
catenated string and the substrings are then decoded and pc ¼ (7)
mapped into the value in the corresponding search space. k3 , fc # f̄
There are three main GA operators: reproduction, crossover, and
and mutation. The reproduction operator allows highly (
productive chromosomes (strings) to live and produce off- k2 (fmax ¹ fi )=(fmax ¹ f̄ ), fi . f̄
springs in the next generation. The crossover operator, used pm ¼ (8)
with a specified probability, exchanges genetic information k4 , fi # f̄
by splitting two chromosomes at a random site and joining where k 1, k 2, k 3 and k 4 have to be less than 1.0 to constrain p c
the first part of one chromosome with second part of another and p m to the range 0.0–1.0, f c is the larger of fitness values
chromosome. Mutation introduces occasional changes of a of the individuals selected for crossover and f i is the fitness of
random string position with a specified mutation probability. the ith chromosome to which the mutation with probability
The significance of p c and p m in controlling GA per- p m is applied.
formance has long been acknowledged in GA research. The As the GA converges, the fitness distance between the
crossover probability p c controls the rate at which solutions chromosomes in each population will become smaller and
are subjected to crossover. The higher value of p c, the smaller. The sum of fitness distance between the ith
quicker are the new solutions introduced into the population. chromosome and the other chromosomes is denoted by
As p c increases, however, solutions can be disrupted faster
than selection can exploit them. Mutation is only a secondary X n
C(fi ) ¼ lfi ¹ fj l, i Þ j: (9)
operator to restore genetic material. Nevertheless, the choice j¼1
of p m is critical to GA performance. Large values of p m
transform the GA into a purely random search algorithm, The normalized fitness distance is obtained by dividing the

Figure 2. IEEE 30-bus power system

566 Optimal reactive power dispatch: Q. H. Wu et al

Table 1. Variable limits (p.u.)

Bus 1 2 5 8 11 13
Qmax
G 2.0 1.0 0.8 0.6 0.5 0.6
Qmin
G ¹0.2 ¹0.2 ¹0.15 ¹0.15 ¹0.1 ¹0.15
max
VG VGmin max
VPQ min
VPQ T max T min Qmax
C Qmin
C
1.1 0.9 1.05 0.95 1.1 0.9 0.1 ¹0.1

sum of fitness distance by (n ¹ 1)maxj lfi ¹ fj l: evaluated on the IEEE 30-bus power system which is shown
X in Figure 2. The system data and operating conditions are
n
j ¼ 1 lfi ¹ fj l given in Appendix A. The network consists of 48 branches,
C̃(fi ) ¼ , i Þ j: (10)
(n ¹ 1)maxj lfi ¹ fj l six generator-buses and 20 load-buses. Four branches, (6,9),
(6,10), (4,12) and (27,28), are under load tap setting trans-
so that C̃(f i) is limited to the interval (0,1). It has been found former branches. The possible reactive power source instal-
effective to employ C̃(f i) to regulate the mutation operator. lation buses are 3, 10 and 24. Six buses are selected as PV-
When C̃(f i) is large, the chromosome is largely deviated from buses and Vv bus as follows: PV-buses: bus 2, 5, 8, 11, 13. Vv
the population, so we should mutate the chromosome with a bus: bus 1. The others are PQ-buses. The variable limits are
higher probability. This can be done by setting the given in Table 1. The transformer taps and the reactive
probability of mutation as: power source installation are discrete variables with the
(
k2 (fmax ¹ fi )C̃(fi )=(fmax ¹ f̄ ), fi . f̄ change step of 0.02 and 0.01 p.u., respectively.
pm ¼ (11) The system loads are given as follows:
k4 , fi # f̄
Pload ¼ 2:834 p:u: Qload ¼ 1:0445 p:u:
In equations (7) and (11), high-fitness solutions are pro-
tected, while solutions with subaverage fitnesses are totally The initial generator bus voltages and transformer taps are
disrupted. This adaptive strategy can prevent premature set to 1.0. The total generations and power losses are
convergence and speed up the convergence rate of GAs. obtained as follows:
PGS ¼ 2:893857 p:u: QGS ¼ 0:980199 p:u:

IV. Simulation results Ploss ¼ 0:059879 p:u: Qloss ¼ ¹ 0:064327 p:u:

The CGA and AGA-based optimal reactive power dispatch There are three bus voltages outside the limits in the net-
algorithms have been implemented based on the flowchart work:
shown in Figure 1 using the PASCAL programming
language and the software are run on the SUNSparc work- V26 ¼ 0:932; V29 ¼ 0:940; V30 ¼ 0:928
station under a UNIX 4.1 operating system. They have been The CGA and AGA have been evaluated on the above power
system. The adopted parameters in the algorithms are given
Table 2. Parameter values for AGA and CGA in Table 2. The objective function with the voltage and
reactive power penalty, equation (4), is used.
AGA CGA Owing to the randomness in the GA approach, the algo-
No. of variables 13 13 rithms are executed 30 times when applied to the test system.
Length of chrom. 65 65 The best and worst reactive power dispatch solutions
Population size 50 50 together with the associated power loss found by the AGA
No. of iteration 370 370 are tabulated in Table 3. For comparison purposes, the
pc 0.95 dispatch solutions obtained by the CGA, are summarized
pm 0.05 in Table 4. Table 5 lists the best control variables found by
k1 0.85 AGA and CGA in the 30 run times.
k2 0.5 The results in Tables 3 and 4 show that the dispatch
k3 1.0 solutions determined by the AGA lead to lower active power
k4 0.05 loss than that found by the CGA, which confirms that the AGA
is well capable of determining the global or near-global

Table 3. Simulation results obtained using AGA

P GS Q GS P loss Q loss P save P save%
Best 2.88326 0.66049 0.04926 ¹0.60151 0.01062 17.74
Worst 2.88391 0.90811 0.04991 ¹0.35389 0.00997 16.65
Average 2.88353 0.91784 0.04952 ¹0.34416 0.01035 17.28

Table 4. Simulation results obtained using CGA

P GS Q GS P loss Q loss P save P save%
Best 2.88380 1.02774 0.04980 ¹0.23426 0.01008 16.84
Worst 2.88614 1.16647 0.05214 ¹0.09553 0.00774 12.93
Average 2.88481 1.02788 0.05081 ¹0.23412 0.00906 15.14
 Optimal reactive power dispatch: Q. H. Wu et al 567

Table 5. Best control variables obtained by AGA and CGA (p.u.)

Method V1 V2 V5 V8 V 11 V 13 T1 T2 T3 T4 Q3 Q 10 Q 24
AGA 1.0751 1.0646 1.0422 1.0454 1.0337 1.0548 0.94 1.04 1.04 1.02 ¹0.00 0.37 0.06
CGA 1.0744 1.0625 1.0358 1.0408 1.0825 1.0610 0.96 1.00 1.00 1.04 ¹0.07 ¹0.01 0.06

optimum dispatch solution. In addition, the results summar- feature of robust computation. From the simulation study, it
ized in Table 6 show that the proposed AGA is about three has been found that the AGA converges to the global
times faster than CGA in speed. The optimization search optimum.
procedures by the AGA and CGA is shown in Figure 3. It can
be seen that, by using the adaptive probabilities of crossover
and mutation, the iterations for convergence can be reduced VI. Nomenclature
greatly. In both of the optimization procedures, all the bus
voltages are inside the limits. v ij voltage angle difference between buses i and j
(rad)
B ij transfer susceptance between bus i and j (p.u.)
V. Conclusion fQ active power loss in network (p.u.)
An adaptive genetic algorithm has been developed for G ij transfer conductance between bus i and j (p.u.)
determination of the global or near-global optimum solution gk conductance of branch k (p.u.)
for optimal reactive power dispatch and voltage control of N0 set of numbers of total buses excluding slack bus
power systems. In the adaptive genetic algorithm, the prob- NB set of numbers of total buses
abilities of crossover and mutation, p c and p m, are varied NC set of numbers of possible reactive power source
depending on the fitness values of the solutions and the installation buses
normalized fitness distances between the solutions. By this ND set of numbers of power demand buses
approach, the simulated-annealling-like strategy used in Ref. NE set of numbers of network branches
[7] has been replaced, which is relative to the reproduction NG set of numbers of generator buses
numbers in the evolutionary process. The performance of the Ni set of numbers of buses adjacent to bus i,
proposed algorithm demonstrated through its evaluation on including bus i
the IEEE 30-bus power system shows that the AGA is able to N PQ set of numbers of PQ buses
undertake global search with a fast convergence rate and a N PV set of numbers of PV buses
NQlim set of numbers of buses on which injected reactive
power outside limits
Table 6. Shortest and longest execution time by AGA and NT set of numbers of transformer branches
CGA NVlim set of numbers of buses on which voltages outside
Method Shortest Longest Average limits
execution execution execution P Di demanded active power at bus i (p.u.)
time (s) time (s) time (s) P Gi injected active power at bus i (p.u.)
Pkloss active power loss in branch k (p.u.)
AGA 50.72 57.98 54.28 Ps injected active power at slack bus (p.u.)
CGA 148.37 162.49 156.34 Q Ci reactive power source installation at bus i (p.u.)

Figure 3. Optimization procedure with the genetic algorithms

568 Optimal reactive power dispatch: Q. H. Wu et al

Q Di demanded reactive power at bus i (p.u.) ing approaches to economic dispatch. IEE Proceedings Part C, Journal
of Generation, Transmission and Distribution, 1994, 141(5), 507–513.
Q Gi injected reactive power at bus i (p.u.) 12. Wong, K. P. and Wong, Y. W., Thermal generator scheduling using
Ti tap position of transformer i hybrid genetic/simulated-annealing approach. IEE Proceedings Part C,
Vi voltage magnitude of bus i (p.u.) Journal of Generation, Transmission and Distribution, 1995, 142(4),
V PQ voltage vectors of PQ buses (p.u.) 372–380.
V PV voltage vectors of PV buses (p.u.) 13. Walter, D. C. and Sheble, G. P., Genetic algorithm solution of economic
dispatch with valve point loading. IEEE Transactions on Power
Systems, 1993, 8(3), 1325–1332.
14. Lee, K. Y., Bai, X. M. and Park, Y. M., Optimization method for
reactive power planning by using a modified genetic algorithm. IEEE
VII. References Transactions on Power Systems, 1995, 10(4), 1843–1850.
1. Dommel, H. W. and Tinney, W. F., Optimal power flow solutions. IEEE 15. Dasgupta, D. and Mcgregor, D. R., Thermal unit commitment using
Transactions on Power Apparatus and Systems, 1968, 87, 1866–1876. genetic algorithms. IEE Proceedings Part C, Journal of Generation,
2. Sachdeva, S. S. and Billinton, R., Optimum network VAR planning by Transmission and Distribution, 1994, 141(5), 459–465.
nonlinear programming. IEEE Transactions on Power Apparatus and 16. Sheble, G. B. and Brittig, K., Refined genetic algorithm—economic
Systems, 1973, 92, 1217–1225. dispatch example. IEEE Transactions on Power Systems, 1995, 10(1),
3. Lee, K. Y., Park, Y. M. and Ortiz, J. L., A united approach to optimal 117–123.
real and reactive power dispatch. IEEE Transactions on Power 17. Cao, Y. J., Wu, Q. H. and Shimmin, D. W., Study of initial population in
Apparatus and Systems, 1985, 104, 1147–1153. evolutionary programming. In Proceedings of European Control
4. Hong, Y. Y., Sun, D. I., Lin, S. Y. and Lin, C. J., Multi-year multi-case Conference. Brussels, 1–4 July, 1997, paper no: 368, pp. 1–4.
optimal AVR planning. IEEE Transactions on Power Systems, 1990, 18. Cao, Y. J. and Wu, Q. H., Convergence analysis of adaptive genetic
5(4), 1294–1301. algorithms. In Proceedings of the 2nd International Conference on
5. Yin, X. and Germay, N., Investigations on solving the load flow Genetic Algorithms in Engineering Systems: Innovations and Applica-
problem by genetic algorithms. Electric Power System Research tions. University of Strathclyde, Glasgow, UK, 2–4 September, 1997,
Journal, 1992, 22, 151–163. pp. 85–89.
6. Wu, Q. H. and Ma, J. T., Power system optimal reactive power dispatch 19. Davis, L., Adapting operator probabilities in genetic algorithms. In
using evolutionary programming. IEEE Transactions on Power Proceedings of the 3rd International Conference on Genetic
Systems, 1995, 10(3), 1243–1249. Algorithms, Fairfax, VA, June 1989, pp. 61–69.
7. Wu, Q. H. and Ma, J. T., Genetic search for optimal reactive power 20. Fogarty, T. C., Varying the probability of mutation in genetic
dispatch of power systems. In Proceedings of the IEE International algorithms. In Proceedings of the 3rd International Genetic Algorithms,
Control Conference, Control 94, Vol. 1. Coventry, UK, April, 1994, pp. Fairfax, VA, June 1989, pp. 104–109.
717–722. 21. Goldberg, D. E., Genetic Algorithms in Search, Optimization, and
8. Wu, Q. H. and Cao, Y. J., Stochastic optimization of control parameters Machine Learning. Addison-Wesley, Reading, MA, 1989.
in genetic algorithms. In Proceedings of the IEEE International 22. Grefenstette, J. J., Optimization of control parameters for genetic
Conference on Evolutionary Computation. Purdue University, USA, algorithms. IEEE Transactions on System, Man and Cybernetics,
April, 1997, pp. 77–80. 1986, 16(1), 122–128.
9. Iba, K., Reactive power optimization by genetic algorithm. IEEE 23. Schaffer, D. and Morishma, A., An adaptive crossover mechanism for
Transactions on Power Systems, 1994, 9(2), 685–692. genetic algorithms. In Proceedings of the 2nd International Conference
10. Chen, P. H. and Chang, H. C., Large scale economic dispatch by genetic on Genetic Algorithms, Cambridge, MA, July 1987, pp. 36–40.
algorithm. IEEE Transactions on Power Systems, 1995, 10(4), 1919– 24. Srinivas, M. and Patnaik, L. M., Adaptive probabilities of crossover and
1927. mutation in genetic algorithms. IEEE Transactions on System, Man and
11. Wong, K. P. and Wong, Y. W., Genetic and genetic/simulated-anneal- Cybernetics, 1994, 24(4), 656–667.

Appendix A Data for the 30-bus power system

(100MVA base)
Tables 7 and 8

Table 7. Load data

Bus no. Load Bus no. Load
P (p.u.) Q (p.u.) P (p.u.) Q (p.u.)
1 0.000 0.000 16 0.035 0.018
2 0.217 0.127 17 0.090 0.058
3 0.024 0.012 18 0.032 0.009
4 0.076 0.016 19 0.095 0.034
5 0.942 0.190 20 0.022 0.007
6 0.000 0.000 21 0.175 0.112
7 0.228 0.109 22 0.000 0.000
8 0.300 0.300 23 0.032 0.016
9 0.000 0.000 24 0.087 0.067
10 0.058 0.020 25 0.000 0.000
11 0.000 0.000 26 0.035 0.023
12 0.112 0.075 27 0.000 0.000
13 0.000 0.000 28 0.000 0.000
14 0.062 0.016 29 0.024 0.009
15 0.082 0.025 30 0.106 0.019
 Optimal reactive power dispatch: Q. H. Wu et al 569

Table 8. Transmission line data

Line no. From bus To bus Line impedance Line From bus To bus Line impedance
no. no. no. no. no.
R (p.u.) X (p.u.) R (p.u.) X (p.u.)
1 1 2 0.0192 0.0575 22 15 18 0.1070 0.2185
2 1 3 0.0452 0.1852 23 18 19 0.0639 0.1292
3 2 4 0.0570 0.1737 24 19 20 0.0340 0.0680
4 3 4 0.0132 0.0379 25 10 20 0.0936 0.2090
5 2 5 0.0472 0.1983 26 10 17 0.0324 0.0845
6 2 6 0.0581 0.1763 27 10 21 0.0348 0.0749
7 4 6 0.0119 0.0414 28 10 22 0.0727 0.1499
8 5 7 0.0460 0.1160 29 21 22 0.0116 0.0236
9 6 7 0.0267 0.0820 30 15 23 0.1000 0.2020
10 6 8 0.0120 0.0420 31 22 24 0.1150 0.1790
11 6 9 0.0000 0.2080 32 23 24 0.1320 0.2700
12 6 10 0.0000 0.5560 33 24 25 0.1885 0.3292
13 9 11 0.0000 0.2080 34 25 26 0.2544 0.3800
14 9 10 0.0000 0.1100 35 25 27 0.1093 0.2087
15 4 12 0.0000 0.2560 36 28 27 0.0000 0.3960
16 12 13 0.000 0.1400 37 27 29 0.2198 0.4153
17 12 14 0.1231 0.2559 38 27 30 0.3202 0.6027
18 12 15 0.0662 0.1304 39 29 30 0.2399 0.4533
19 12 16 0.0945 0.1987 40 8 28 0.6360 0.2000
20 14 15 0.2210 0.1997 41 6 28 0.0169 0.0599
21 16 17 0.0824 0.1932