Biogeography-Based Optimization and The Solution of The Power Flow Problem
Biogeography-Based Optimization and The Solution of The Power Flow Problem
SMC 2009
entering the network from the bus, respectively. For the That is, the power injected into the network at Bus i is equal to
standard IEEE 30-bus system used in our simulations [16], the the power supplied by the generator minus the power absorbed
voltage magnitude and the phase angle vectors are denoted by by the load.
V = (V1 ,V2 , ,V30 ) , = (1 , 2 , , 30 ) (1) Every power system has a reference bus called a slack bus.
The slack bus has a generator connected to it and is usually
The explicit dependence of the active and reactive powers denoted as Bus 1. The voltage magnitude V1 and angle 1 are
on the voltages, angles, and admittances in the system is given specified with 1 = 0 being specified as the reference angle for
by the power flow equations
all voltages and currents in the system. The active and reactive
Nb
powers of the generator, PG1 and QG1 , are not specified, so P1
Pi (V , ) = Vi V j [Gik cos(ik ) + Bik sin(ik )]
k =1
and Q1 are not specified but are left as free variables in order
(2)
Nb to guarantee that the system of equations in (2) has a solution.
Qi (V , ) = Vi V j [Gik sin(ik ) Bik cos(ik )]
k =1 At every bus in the system to which a load is connected, the
powers, PDi and QDi , absorbed by the load are specified in the
where i) for i k , Gik + jBik is the negative of the admittance
design requirements. Some of the buses will have a generator
connected between Bus i and Bus j; ii) for i = k , Gii + jBii is connected to them, and the active power PGi supplied by the
the sum of all admittances connected to Bus i; and iii) N b is generator is specified except for the generator at the slack bus.
the number of buses in the system. At each bus there are four And some of these generators may also have the reactive power
unknown quantities: Vi , i , Pi , and Qi . The system of QGi specified by design because of the type of power plant or
nonlinear equations in (2) has 4 N b unknowns and only 2 N b other conditions requiring both the active and reactive powers
equations, and therefore it is underdetermined. to be maintained at a certain level. So in these cases, both PGi
and QGi are specified. Buses for which all four powers,
Gen 1 Gen 2
PGi , QGi , PDi , and QDi , are specified are called PQ buses.
Bus 1 V11 V2 2 Bus 2 At the remaining buses to which a generator is connected,
P1 , Q1 P2 , Q2
only the active power of the generator PGi is specified; the
reactive power QGi is left unspecified. The voltage at these
Load 1 buses is controlled to specified values by adjusting the
generators output. These are called PV buses.
Transmission
Line
We can combine the generator and load power
specifications into a single quantity called the specified power
Bus 3 at Bus i. It is defined as the power generated at the bus minus
V33
P3 , Q3 the load connected to the bus:
Load 2
Pi sp = PGi PDi
(4)
Figure 3: A 3-Bus Power System Qisp = QGi QDi
The general practice in the analysis and design of power The first equation in (4) is well defined for all buses except the
systems is to specify two of the four unknown quantities at slack bus, namely all PQ and PV buses. The second is well
each bus so that the system in (2) can be solved. Depending on defined for all buses for which QGi is specified, namely the PQ
which quantities are specified, the buses are divided into three
buses.
categories: slack buses, PQ buses, and PV buses. These
categories are described next. In order to solve the power flow equations in (2) iteratively,
an initial value for the voltage-angle vector (V , ) is generated
For the sake of the mathematical notation involved, it is
assumed that every bus in the system has both a generator and a to start the algorithm. The values of Pi and Qi can then be
load connected to it. Under this assumption, four additional calculated for all buses in the system using (2). These quantities
quantities are defined for Bus i: the active and reactive powers represent the net estimated power entering the network from
supplied by the generator, PGi and QGi , and the active and Bus i and are denoted by Pi est and Qiest . The power mismatch
reactive powers absorbed by the load, PDi and QDi . If a is then defined as
particular bus does not have a generator or a load, then we set
the corresponding quantities to zero. Using this notation, the Pi = Pi est Pi sp = Pi ( PGi PDi )
power balance equations are defined at Bus i as (5)
Qi = Qiest Qisp = Qi ( QGi QDi )
Pi = PGi PDi
(3) Mismatches are nonzero when the power estimated from a
Qi = QGi QDi
given voltage-angle vector (V , ) differs from the specified
SMC 2009
power at the bus. The goal of the power flow problem is to The GA and BBO algorithms require a fitness function for
find a vector (V , ) such that the power balance equation at roulette wheel calculations. Since the theoretical minimum of
each bus is satisfied: c(V , ) is zero, the fitness function was chosen to be
Pi (V , ) = 0 100
(6) f (V , ) = (10)
Qi (V , ) = 0 c (V, ) + 0.005
For PV buses, QGi is not specified, and for slack buses, to avoid extremely large fitness values or divide-by-zero errors.
neither PGi nor QGi are specified. In these cases, a mismatch
IV. SIMULATION RESULTS
for the corresponding quantity is not defined, and there is no
requirement that the power balance equations associated with The GA and BBO algorithms were applied to the power
these buses be satisfied when solving the power flow problem flow problem using the MATLAB programming environment.
[15]. The standard IEEE 30-bus system, based on the data from a
real power system in Virginia in the 1960s, was used as a test
Since the voltage magnitude is specified at PV buses, the case for the study. A diagram and the electrical specifications
given voltage magnitude vector V can be used to define the of the system are available at the Power Systems Test Case
voltage mismatch [14] as Archives website maintained by the University of Washington
[16]. The data file containing the electrical specifications is
Vi = Vi Vi sp (7) formatted in the IEEE Common Format for the exchange of
solved load flow data [17].
where Vi is the i-th component of the vector V in (1). As in (6),
The GA population size was chosen as 100. Each
a voltage balance equation can be then defined:
chromosome in the population has 60 genes made up of the 30
Vi = 0 (8) voltage magnitudes and 30 voltage phase angles in the vector
Note that Vi is not estimated from the power flow equations X = (V1 , V2 , ,V30 , 1 , 2 , , 30 ) (11)
est est
(2) as were Pi and Q , but is obtained from the estimate for
i The slack bus voltage V1 was fixed at 1 p.u. (per unit, a
V in the algorithm when solving the power flow problem. normalized unit of voltage measurement) and its angle 1 was
Typically numerical methods such as the Newton-Raphson fixed at zero degrees. The weights assigned to the cost function
method are used to solve the nonlinear system in (2) were Q = 1 and V = 100.
[8][13][15]. As mentioned above, an initial guess for the
voltage-angle vector (V , ) is required to start the algorithm. For the reproduction operation, a one-point crossover
This initial solution vector is used in (5) and (7) to generate a scheme was used. Each pair of parent chromosomes can
mismatch or estimation error. In numerical approaches, a potentially generate two offspring depending on the crossover
gradient method is then applied to the mismatch to generate a probability. Fifty parents were randomly chosen by roulette
new estimate for (V , ), and the process is iterated. wheel selection based on the fitness expression in (10).
Crossover was set to occur with a probability of 0.70 at a
In GA or BBO algorithms, an initial population of randomly determined crossover point in the chromosome.
individuals (chromosomes for the GA and islands for the BBO)
is generated by assigning a random vector (V , ) to every After the crossover step, a non-uniform mutation operation
is applied to each chromosome in the population. The
individual in the population. The components of the vector are probability of selecting a given chromosome for mutation is
real-valued voltages and phase angles constrained to be within 0.05. If a chromosome is chosen for mutation, then one of its
some prescribed limits based on considerations from the power genes, say g X , in (11) is changed to a new value g
system specifications. The components of these vectors
comprise the genes in the GA and the features (SIVs) in the according to the formula
BBO algorithm. The mismatches can then be calculated for
g ' = a '+ r ( b ' a ' )
each individual and a cost assigned from a weighted sum of the
squares of the mismatches: where a ' = g f (G ) ( g a )
(12)
b ' = g + f (G )(b g )
c (V , ) = Pi 2 + Q Q i
2
+ V V i
2
(9)
f ( G ) = (1 G / G max )
iPV PQ iPQ iPV 2
and
Here PV and PQ are index sets for the PV and PQ buses,
respectively. This cost is used as the objective function in both and r is a uniformly distributed random number in [0, 1], G is
the GA and the BBO algorithms. The goal of each algorithm is the current generation number, Gmax is the maximum number
to find a global minimum for the cost c(V , ) in as few of generations, and a and b are the lower and upper limits
generations as possible. For the power flow problem, the global associated with the gene gi . Duplicate chromosomes were
minimum is known to be zero from (6) and (8). removed at each generation step and replaced with random
mutations. Lastly, an elitism operation was applied in order to
preserve the six fittest individuals from each generation.
SMC 2009
The BBO algorithm is similar to the GA except that in generations, although these simulations were not subjected to
BBO, the reproduction scheme of the GA is replaced by the Monte Carlo trials because of prohibitive computation times.
migration scheme described in Figure 2. The same mutation, Nevertheless, in numerous non-Monte Carlo simulations with a
duplication removal, and elitism schemes are used in both large number of generations, the BBO always found a smaller
algorithms. In BBO nomenclature, chromosomes are referred minimum cost and had a smaller average cost, even when the
to as islands and genes as features or SIVs. crossover or mutation rates were changed.
Many of the decisions and numerical values within both Some further Monte Carlo results are shown in Figure 5
algorithms depend on stochastically generated quantities, so and Figure 6. Figure 5 shows the minimum cost achieved over
700 Monte Carlo simulations were run to improve the all individuals, over all generations, and over all Monte Carlo
reliability of the conclusions about performance. Only 100 simulations. Figure 6 shows the minimum average cost over all
generations were simulated in order to reduce simulation time, generations and over all trials. Both figures indicate that the
which was 462 minutes on a 2.3 GHz processor. This number BBO algorithm is able to find a better solution than the GA
of generations is extremely small for this type of problem. For with respect to these statistical measures.
example, a GA very similar to the one in this study was applied
to the IEEE 14-bus standard system in [1] and required 10,000
generations to achieve a minimum cost very close to zero. But
since the goal of this study is a comparison of the relative
performance of the GA and the BBO algorithms, such time-
consuming simulations were not required.
The overall performance of the two algorithms is shown in
Figure 4. The plots were generated as follows. A cost is
calculated for each individual in a population. Each population
is associated with a generation, and each generation is
associated with a Monte Carlo trial. Thus, a cost is generated
for each individual in each generation in each trial; or, put
another way, the cost is a function of the individual index, the
generation index, and the trial index:
Figure 5: Minimum Cost from Monte Carlo Trials
cost = cost ( ind, gen, trial ) (13)
(
AveCost ( gen ) = ave ave cost ( ind, gen, trial )
trials ind
) (14)
V. CONCLUSION
The objective of this study was to compare the performance
of a novel biogeography-based optimization algorithm to that
of a traditional genetic algorithm in a practical application.
Both algorithms were applied to a standard power flow
problem in power systems analysis. Initial Monte Carlo testing
indicates that the BBO consistently performs better than the
GA with respect to certain basic statistical measures. The
mathematical basis for this improved performance is discussed
in [18].
Figure 4: Average Cost of 700 Monte Carlo Trials
For this study, the algorithms were applied to the relatively
The plot shows that for about the first 16 generations, the simple IEEE 30-bus test case so that the code could be tested
GA performs better than the BBO, but thereafter the BBO for accuracy more efficiently and initial results generated in a
performs better. This improved performance is consistently reasonable amount time. Although traditional numerical
observed in simulations with 1000, 5000, and 10000 methods for solving such problems have been successful, the
SMC 2009
difficulty in solving such problems increases rapidly as the [5] R. MacArthur and E. Wilson, The Theory of Biogeography, Princeton
number of buses increases or the system is stressed under University Press, 1967.
heavy loads approaching the maximum operating point of the [6] D. Simon, Biogeography-based optimization, IEEE Transactions on
Evolutionary Computation Vol. 12, No. 6, pp. 702-713, December 2008.
system. In the former situation, the Newton-Raphson method
[7] J. Carpentier, Contribution to the study of economic dispatch, Bulletin
often converges to a local minimum which is very dependent of the French Society of Electricians, Vol. 3, No. 8, pp. 431-437, 1962.
on the starting solution. In the latter situation, the method is
[8] A. Wood and B. Wollenberg, Power Generation Operation and Control,
more likely not to converge at all because the Jacobian tends to Wiley-Interscience, NJ, 1996.
be singular [14]. It is in these situations that population-based [9] D. Goldberg, Genetic Algorithms in Search, Optimization, and Machine
search algorithms have shown the most potential, since they are Learning, Addison Wesley, MA, 1989.
designed to search for global optima, and no derivatives or [10] J. Yuryevich and K. Wong, Evolutionary programming based optimal
matrix inversions are involved. The contribution of this work, power flow algorithm, IEEE Transactions on Power Systems, Vol. 14,
then, is that we have demonstrated the improved performance No. 4, pp. 1245-1250, November 1999.
of the BBO algorithm over the genetic algorithm in a special [11] S. Agrawal, B. Panigrahi, and M. Tiwari, Multiobjective particle swarm
case. We are now in the position to apply the algorithms to the algorithm with fuzzy clustering for electrical power dispatch, IEEE
Transactions on Evolutionary Computation, Vol. 12, No. 5, pp. 529-541,
much more difficult and important optimal power flow problem October 2008.
for a large number of buses where traditional approaches are [12] K. Lenin and M. Mohan, Ant colony search algorithm for optimal
less likely to succeed. This will also allow us to determine reactive power optimization, Serbian Journal of Electrical Engineering
whether the BBO algorithm continues to perform better than Vol. 3, No. 1, pp. 77 88, June 2006.
the GA in these more difficult problems. [13] J. Glover and S. Sarma, Power System Analysis and Design, Fourth
Edition, CL-Engineering, 2007.
Since evolutionary algorithms are probabilistic, the first
[14] K. Wong, A. Li, and M. Law, Development of constrained genetic
undertaking in future work would be to optimize the code for algorithm load flow method, IEE Proceedings Generation,
speed so that more Monte Carlo simulations could be applied to Transmission, and Distribution, Vol. 144, No. 2, pp. 91 99, Mar 1997.
both of the algorithms that we tested. This would allow the [15] J. Grainger and W. Stevenson, Power System Analysis, McGraw-Hill
student-t test [19] to be applied to the results from the two Science, NY, 1994.
algorithms to determine the statistical significance of the [16] R. Christie, Power Systems Test Case Archive, College of Engineering,
differences between the results. University of Washington, WA (Dec. 2008). [Online]. Available:
www.ee.washington.edu/research/pstca
One of the strengths of population-based search algorithms [17] H. Pierce, Jr., Common format for exchange of solved load flow data,
is their ability to determine global optima with probabilities IEEE Transactions on Power Apparatus and Systems, Vol. PAS-92, No.
that can be determined in many situations. It would be 6, pp. 1916 1925, Nov/Dec 1973.
beneficial to design experiments to measure the relative [18] D. Simon, M. Ergezer, and D.Du, Markov analysis of biogeography-
efficacy of the GA and BBO algorithms in finding a global based optimization algorithms, unpublished (Mar. 2009). [Online].
Available: http://academic.csuohio.edu/simond/bbo/markov
minimum for the optimal power flow problem.
[19] J. Kennedy and R. Eberhart, Swarm Intelligence, Academic Press, CA,
Many variations and enhancements are possible with both 2001.
the GA and the BBO algorithms, and some of these should be [20] S. Rahnamayan, H. Tizhoosh, M. Salama, Opposition-based differential
implemented so that comparisons can be made to determine evolution, IEEE Transactions on Evolutionary Computation Vol. 12,
No. 1, pp. 64-79, Feb. 2008.
whether the BBO algorithm maintains its superior performance
with these adaptations. One such variation fuses opposition- [21] M. Ergezer, D. Simon, and D. Du, Opposition biogeography-based
optimization, unpublished.
based learning [20] with biogeography-based optimization
[22] D. Simon, A probabilistic analysis of a simplified biogeography-based
(OBBO) [21] and has shown promising initial results for the optimization algorithm, unpublished (Mar. 2009). [Online]. Available:
power flow problem and for many benchmark problems that http://academic.csuohio.edu/simond/bbo/simplified
are used to test evolutionary algorithms.
Finally, different approaches to the migration operation
could be tested. Instead of the linear approximations that we
used for the migration rate curves and , more complicated
curves can be constructed. Also, the migration strategy that we
used can be described as an immigration-based strategy, but
others are possible, such as the emigration-based, single
immigration, and single emigration strategies discussed in [22].
REFERENCES
[1] L. Lai and N. Sinha, Genetic algorithms for solving optimal power flow
problems, Chapter 17, Modern Heuristic Optimization Techniques, pp.
471 500, IEEE Press, 2008.
[2] A. Wallace, The Geographical Distribution of Animals, first published in
1876, Adamant Media Corporation, 2006.
[3] C. Darwin, The Origin of Species, first published in 1859, Gramercy,
1995.
[4] E. Munroe, The geographical distribution of butterflies in the West
Indies, PhD Dissertation, Cornell University, Ithaca, New York, 1948.
SMC 2009