640 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 27, NO.

2, APRIL 2012

Optimal Sizing of Hybrid Wind/PV/Diesel Generation

in a Stand-Alone Power System Using Markov-Based
Genetic Algorithm
Ying-Yi Hong, Senior Member, IEEE, and Ruo-Chen Lian

Abstract—Owing to the Kyoto Protocol and the growing deple- system and microgrid. A small stand-alone power system can
tion of natural resources, renewable energies have attracted much be found on offshore islands. Renewable energies will help
attention. This paper considers 25-kW wind-turbine generator, reduce COx, SOx, and NOx emissions. This paper addresses
5-kW PV and 30-kW diesel generator as unit sizes for generation
planning in a stand-alone power system. The investment cost the generation expansion planning for small stand-alone power
(installation and unit costs) and fuel cost are minimized while re- systems.
taining the reliability requirement and CO2 emission limit. First, There are some studies on generation expansion planning for
the fuzzy-c-means (FCM) is employed to cluster the operation small stand-alone power systems. These studies can be divided
states for system load, wind-turbine generations (WTG), and PV into three main categories, namely, reliability analysis [3]–[6];
in 8760 h. Then, the Markov models for the system load, WTG,
and photovoltaic (PV) are established. The Markov models are optimization [7], [8]; and enumeration [9]–[11], as discussed in
embedded into the genetic algorithm to determine the optimal the following text.
sizes for WTG, PV, and the diesel generator. The simulation Reliability analyses generally involve loss of load proba-
results reveal that computation time can be reduced greatly while bility (LOLP) or loss of power-supply probability (LPSP),
optimality can be still retained, compared with the traditional which implies the probability for imbalance between electricity
method using chronological data.
supply and load. Billinton utilized renewable energy in capacity
Index Terms—Genetic algorithm (GA), Markov model, optimal planning through cost and LOLP that recognized the highly
size, reliability, renewable energy.
erratic nature of renewable energy sources while maintaining
the chronology and interdependence of their inherent random
I. INTRODUCTION variables [3]. Karki presented an approach to capacity planning
of small isolated systems, in which the conventional proba-
ONCENTRATIONS of greenhouse gases (COx, SOx, bilistic and the well-being indices were used jointly [4]. Ai
C and NOx) in the atmosphere have been increasing. The
Kyoto Protocol [1] is an agreement under which industrialized
proposed a complete set of match calculation methods for the
optimum sizing of the wind/PV hybrid system and the annual
countries should reduce their collective emissions of green- LPSP values of the wind/PV hybrid systems when different
house gases in 2012 by 5.2% compared with that in 1990. In capacities of PV array and battery bank were considered [5].
the Copenhagen Accord, the convention agrees to reduce global Nelson used the number of wind turbines as a free input param-
emissions so as to maintain the increase in global temperature eter and then the number of PV panels is calculated using the
below 2 C [2]. LPSP technique to match generation with load in [6].
Renewable energies, such as solar, wind, geothermal, The optimization method employs a set of algorithmic steps
biomass, tidal, and hydropower, constitute a type of distributed to reduce the investment, operation, and fuel costs. Kabouris and
electricity resource and have recently received much attention Contaxis proposed an optimization method with planning and
as alternatives for electricity generation. The use of renewable operational constraints incorporated with the stochastic nature
energies can mitigate the greenhouse effects to meet the strict of the meteorological conditions, loads, and the availability of
the diesel units [7]. Kabouris and Contaxis also determined the
requirements stated in the Kyoto Protocol. In particular, wind
and solar photovoltaic (PV) power generations play essential optimal expansion policy over a period of years for a limited
roles in a grid, especially in a small isolated (stand-alone) power size generation system consisting of wind generators and PV
panels [8]. With the historical load data, the characteristics of
renewable sources, the cost of energy, and the associated esca-
Manuscript received January 22, 2011; revised March 18, 2011, September
17, 2011, and November 15, 2011; accepted November 15, 2011. Date of publi-
lation rate, the dynamic programming algorithm was employed
cation January 05, 2012; date of current version March 28, 2012. This work to find the optimal expansion policy for the period of the next
was supported by the National Science Council, Taiwan, under Grants NSC years.
99-2632-E-033-001-MY3 and NSC 100-3113-E-007-006. Paper no. TPWRD-
00060-2011.
The enumeration method considers various types and capaci-
The authors are with the Department of Electrical Engineering, Chung Yuan ties of distributed-generation resources for comparison of the in-
Christian University (CYCU), Chung Li 320, Taiwan (e-mail: [email protected]. vestment, operation, and fuel costs. Lund analyzed them solely
cycu.edu.tw; [email protected]). from a technical point of view; investments are not included [9].
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. The PV and wave power were excluded when considering an op-
Digital Object Identifier 10.1109/TPWRD.2011.2177102 timal solution. Different sources are analyzed in the range of an
0885-8977/$31.00 © 2012 IEEE
HONG AND LIAN: OPTIMAL SIZING OF HYBRID WIND/PV/DIESEL GENERATION 641

electricity production from 0 to 100% of the electricity demand TABLE I

[9]. Diaf considered various types and capacities of system de- WEATHER STATISTICS FOR MT. JADE IN 2007
vices with configurations that can meet the desired system reli-
ability by changing the type and size of the device systems. The
configuration with the lowest “levelized cost of energy” gave the
optimal choice [10]. Kaldellis used the available experimental
probability density function distributions in the past. The min-
imum acceptable nominal power of the wind turbine was then
predicted using nominal power, capacity factor, and the mean TABLE II
power coefficient [11]. WEATHER STATISTICS FOR ORCHID ISLAND IN 2007
As discussed before, most of the previous studies are on re-
liability analyses [3]–[6] and enumeration methods [9]–[11].
The optimization-based approaches were rarely implemented
due to the difficulty in solving the integer variables [7], [8].
In this paper, the investment cost (installation and unit cost)
and fuel cost are minimized while the reliability requirement
and CO emission constraint are retained. First, the fuzzy-c-
TABLE III
means (FCM) is employed to cluster the operation states for VARIOUS COSTS FOR DIESEL, WTG, AND PV UNITS
system load, wind-turbine generation (WTG), and solar PV.
Then, the Markov models for the system load, WTG, and PV
are established. The genetic algorithm (GA) incorporated with
the Markov models is utilized to determine optimal sizes for
WTG, PV, and diesel generator.

II. MODELS
solar radiations/temperatures in 8760 h will be transferred into
The problem here is to determine the numbers of WTG, 8760 corresponding generations using manufacturer specifica-
PV, and diesel generators for a given system load. The IEEE tions of WTG and PV, respectively.
Reliability Test System (RTS) load profile [12] for the 365
24 8760) h in a year is utilized in the studied stand-alone B. Cost Components
power system.
The unit cost, installation cost, and even maintenance cost
of new diesel units and renewable energy units should be con-
A. Kilowatt Generations of WTG and PV Using Historical sidered in the planning stage. Compared with the cost of PV
Chronological Weather Data material, the installation cost for PV is rather inexpensive and
can be neglected. These considerations are the same as those
The weather statistics (wind speed, solar irradiation, and tem-
in [3] and [4]. Table III illustrates the costs of these distributed
perature) for the studied power systems should be taken into ac-
generations.
count when planning generation expansion. The effects of wind
The fuel cost of the diesel generator can be modeled as a
speed and solar irradiation in different areas have great impacts
quadratic function as
on the kilowatt generations. In particular, the efficiency of the
PV will increase in a high mountain area because of higher solar
irradiation and lower temperatures. In general, the typical height (1)
of an anemometer in a weather station is 15 m. The wind speed (2)
measured at this height should be transformed to that at 30 m,
which is the height of the hub of a typical 25-kW WTG. As- where (in U.S. dollars per hour), [in U.S.$/kilowatt-hour
sume that the friction coefficient for the ground is 0.14. Tables I (kWh)], and (U.S.$/kW h) are the cost coefficients. and
and II show the wind speeds (in meters per second) and irra- denote the price (U.S.$/Gcal transformed from U.S.$/liter)
diation (in watts per meter ), measured in 2007, for Mt. Jade of oil and kW generation, respectively. The values of , and
and Orchid Island in Taiwan in all four seasons, respectively. for the 30-kW diesel generator used in this paper are 1.07,
The symbol “SD” means standard deviation. In Taiwan, the 0.0657, and 0.00006, respectively.
spring, summer, autumn, and winter are March–May, June–Au- The total cost includes three terms: 1) capacity cost and instal-
gust, September–November, and December–February, respec- lation cost for new diesel and wind generators, 2) capacity cost
tively. As can be seen, the irradiation in the high mountain is for new PV, and 3) fuel cost for total durations that diesel gen-
very high and the average wind speed for Mt. Jade exceeds 5 erators supply kilowatts. If the system load at hour is greater
m/s, implying that Mt. Jade is suitable for developing wind and than wind power generation plus PV power generation at hour
PV power generations. However, the irradiation in the Orchid , then there is kilowatt generation from the diesel generator,
Island is not as high as that in Mt. Jade. The wind speeds and and the fuel cost at hour can be computed by (2).
642 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 27, NO. 2, APRIL 2012

C. CO Emission
The renewable sources can mitigate CO emission efficiently. If
The stand-alone power system may be allocated with strict an-
then can be computed by (2);
nual CO emission and be supplied with a limited number of
new diesel generators. CO emission can be evaluated from the can be computed by (3). ;
heat in (1) as follows:
If
Heat then can be computed by (2);
(3)
can be computed by (3);
The values of , and for the 30-kW diesel generator used in
Next ;
this paper are 0.028144, 0.001728, and 0.0000017, respectively.
8760;
III. PROBLEM FORMULATION
Total fuel cost ;
The aforementioned problem can be formulated using the
concept of chronological data including hourly load, wind Total (in kilograms).
speed, and irradiation/temperature as follows:
This pseudocode is developed for calculating the dependent
variables in the problem. This is also the advantage of GA that
(4) can solve the independent and dependent variables separately;
however, most of the traditional optimization methods solve
all variables simultaneously. Hence, if is zero, then the
(5) emission and fuel cost will be zero at hour , as can be seen
(6) in the 4th line of the above pseudocode. Please note that if the
(7) number of diesel generators is zero, the third term of (4)
is zero too.
(8)
According to (4)–(9) and the pseudocode for calculating the
(9) fuel cost, LOLP and CO at hour , seeking the optimal sizing
of distributed generations in a stand-alone power system is a
where , and represent the (unknown) numbers of
mixed-integer programming (MIP) problem. Specifically, the
WTG, PV, and diesel generators, respectively. and
unknown and are integer (independent) variables
denote the costs per unit ($/kW) for the WTG, PV, and
and the fuel cost, LOLP, and CO at hour are continuous (de-
diesel generator, respectively. is the installation cost for the
pendent) variables. The reasons for adopting GA used to study
WTG. and are the installation and fuel costs for the
the optimal sizing problem are as follows.
diesel generator, respectively. The superscript “max” and sub-
1) Traditional MIP based on binary linear programming or
script “min” signify the maximum and minimum limits, respec-
branch and bound can only handle explicit variables, con-
tively.
straints, and objectives. The problem of implicit variables
The response of a diesel generation is very fast. The output
and objective cannot be solved by traditional MIP but can
power from 0 kW to the rate generation of a diesel generator
be solved by the GA. The system frequency response dis-
can be achieved in several minutes. The ramping constraints on
cussed in [13] is an example of an implicit objective which
diesel generation are not considered in this paper because the
cannot be expressed in a closed form. LOLP and CO at
time interval (step) in the chronological data is one hour. The
hour defined in the above pseudocode are examples of
ramping rate was not considered in [3]–[11] either.
implicit variables, which are defined by the cause-effect
Once the unknown and are obtained (by the GA
relation (e.g., “IF-THEN”).
in this paper), the chronological data, including hourly load,
2) Linear programming and binary linear programming can
wind speed, and irradiation/temperature can be employed to ex-
only study the problems of linearized objectives and con-
amine (8) and (9). Let be the index for loss of load. More-
straints. This leads to the traditional branch and bound not
over, let the chronological variables and be the
being suitable to solve the nonlinear integer problem, pri-
known kilowatt generations for the WTG and PV at hour ,
marily because the validity of the branching rules is tied
respectively. The chronological variable represents the
with the assumption of linearity (see [14, p. 171]). It is ob-
known total system load at hour . The pseudocode for calcu-
vious that (8) and (9) cannot be linearized due to the im-
lating LOLP and CO at hour is described as follows:
plicit variables (i.e., the fuel cost, LOLP and CO ). How-
ever, GA is able to solve the problems of nonlinear objec-
;
tives and constraints by dealing with the independent and
1, 2, 8760; dependent variables separately.
3) Traditional MIP can be solved by the master-level algo-
If
rithm (i.e., Benders decomposition) if the cost from con-
then 0 0 and 0; tinuous variables and that from the integer variables can
HONG AND LIAN: OPTIMAL SIZING OF HYBRID WIND/PV/DIESEL GENERATION 643

be added together in the objective function [14]. However, Step 3) Update for all
LOLP, CO , and the cost of distributed generations cannot
be added together. (13)

IV. PROPOSED METHOD

The central-processing unit (CPU) times required by Step 4) If , stop; else go to Step
GA-based methods are generally very long because each it- 2).
eration includes many solutions of chromosomes. Hence, this The symbol is the convergence tolerance.
paper presents a Markov-based GA to reduce the computation In this paper, the values of clusters (i.e., ) for the WTG, PV,
burden but to retain the accuracy of the solution. More specif- and system load are 13, 4, and 10, respectively. More specif-
ically, FCM is used to identify the operation states (discrete ically, there is no wind power generation if the wind speed is
and , where and are the below 3 m/s, and the wind power generation will be fixed if the
state indices) in the Markov model from the chronological data wind speed exceeds 16 m/s. Therefore, the wind speed can be
, and . Since the number of states is divided into 13 clusters in the case that the wind speed data are
much smaller than that (8760) of the chronological data, the evenly decomposed. However, FCM will not cluster the wind
computation time can be greatly reduced. speed evenly but coherently. That is, the amounts of data in all
clusters are not the same. The center in the cluster is the
A. Fuzzy-c-means (FCM) representative in this cluster. The values of clusters for the PV
and system load are attained in a similar way. On the other hand,
Traditional -means clustering has at least the following dis- the dimensions of and are both 1 1 and 8760.
advantages that were not considered in this paper: it does not
yield the same result with each run and does not ensure that
the result has a global minimum of variance. Hence, FCM is B. Genetic Algorithm (GA)
adopted herein. In FCM, each data has a degree (membership
function) of belonging to clusters, rather than belonging com- The basis of employing GA to solve the optimization problem
pletely to just one cluster. Bezdek defined a function is encoding all of the search parameters into discrete or binary
as an objective in the FCM algorithm [15] strings named genes. Then, according to the solved problem, a
fitness (objective) function is defined. An individual with better
fitness function values will be chosen and included in the mating
pool for the process of reproduction. Crossover and mutation
operations yield a new generation. The GA is performed by re-
(10) peating each of the aforementioned processes to produce the
fittest individuals.
where is the clustering number, represents the data In this paper, GA is employed to solve (4)–(9). The variables
number, signifies the vector of the center in the th cluster, , and are considered as the gene strings (binary bits)
is the th data vector for clustering, and implies the in the chromosome (control variables) while LOLP and CO are
membership function value as a weighting factor between the state variables. The lengths of binary bits for , and
and . The clustering number is fixed for developing FCM. are 5, 8, and 3, respectively. and cor-
The value of depends on the user’s requirement and the respond to the values of 31, 63, and 7, respectively. The fitness
problem characteristics. When the value of is mini- function is (4). The individuals for , and will be
mized, the data vectors can be partitioned into clusters. crossed over and mutated within their limits without the problem
Bezdek developed four solution steps to achieve the minimum of enforcing inequality constraints. To deal with inequality con-
as follows. straints for the state variables (LOLP and CO ) in GA more ef-
Step 1) Estimate a matrix of membership functions ficiently, the penalty functions are employed in this paper.
A penalty term is augmented to (4) for further genetic selec-
(11) tion. Specifically, the penalty function for dealing with (8) is
defined as follows:
where is the iterative index and 0 initially.
Step 2) Let . Compute the center of the th cluster (14)

where PF is called the penalty factor in this paper. Let the

symbol be the iterative index. Equation (14) implies that the
penalty weight should be gradually increased with
less effect on the initial iterations. The sign of the afore-
mentioned penalty factor is positive. Hence, a chromosome
with a large fitness value augmented with the penalty term will
(12) result in few chances to be selected for further mating.
644 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 27, NO. 2, APRIL 2012

TABLE IV
(a) MARKOV MODEL OF A 25-kW WTG. (b) MARKOV
MODEL OF A 25-kW WTG

Fig. 1. Three-state Markov model.

C. Markov Model
The Markov model is an approach to relating a state proba-
bility with its corresponding event frequency in the stochastic
process [16], [17]. The operation states and
, where and are the state indices, identified by
FCM, will be treated using the Markov model in this paper.
First, a single wind farm with three states is considered. The
failure rate is the state transition probability from state 1 to
state 2 while the repair rate is the state transition probability
from state 2 to state 1. Fig. 1 illustrates a three-state Markov
model.
Let 1, 2, be the probabilities for states,
where is the number of clusters. Then

(15)
TABLE V
MARKOV MODEL OF LOAD PROFILE
(16)

The upper triangle matrix of the squared matrix [ ] includes

the repair rates among states while the lower triangle matrix
consists of the failure rates among states. The diagonal term in
the ( , ) location of [ ] is the negative sum of all offdiagonal
terms at the th column. The absolute value of the diagonal term
for a state in (16) is called the rate of departure for state .
Solving (15) and (16) yields the probabilities for the states.
Let , and be the rate of departure, frequency, and
duration for state , respectively. The frequency and duration
for state are evaluated, respectively, as follows:

(17) corresponding probabilities for two states in different individual

(18) WTG and PV. For the same reason, the aggregate frequency and
duration for a new operation state can be attained.
Table IV shows the Markov model for a 25-kW WTG. The In addition to the probability distribution for all states, the
number of total states is 13, . The advantage of the Markov model is that it provides information
second row (kilowatt generation) in Table IV is the center (i.e., about the frequency and duration of each state. The informa-
defined in Section IV-A (FCM), of the cluster ). The prob- tion of frequency and duration is essential because they can be
ability, rate of departure, frequency (occurrence/h), and dura- employed to compute the kilowatts generated by diesel gener-
tion (h) are attained from the Markov model. Similarly, Table V ators. Once aggregate frequency and duration are known, total
shows the Markov model (state 1, 2, 10) for the system CO emission and total fuel cost can be further evaluated too.
load. For example, assume that the method for computing the total
Practically, WTG and PV exist in a stand-alone power system. fuel cost in a year is , as described
Assume that one WTG (13 states) and one PV (4 states) are con- in Section III, by chronological data. However, the method for
sidered here. There are 52 possible states in total for considering computing the total fuel cost in a year can be simplified by
WTG and PV. Then, the aggregate rate of departure for a new , where and
operation state (scenario) is the sum of the corresponding rates represent the duration and frequency at an aggregate state ,
of departure for two states in individual WTG and PV. The ag- respectively. And if
gregate probability for a new operation state is the product of the .
HONG AND LIAN: OPTIMAL SIZING OF HYBRID WIND/PV/DIESEL GENERATION 645

Fig. 2. Iteration performance of three methods (Mt. Jade).

TABLE VI
COMPARISON BETWEEN CHRONOLOGY-BASED AND MARKOV-BASED
METHODS (MT. JADE)

Fig. 3. Iteration performance for three different LOLP ’s.

chromosome. It can be found that it is easier for the Markov-

V. SIMULATION RESULTS based GA to yield a better result and the chronology-based can
The IEEE reliability test system (loads for 8760 h) [12] and only gain a local optimum due to complicate loops. The main
the weather data in Mt. Jade (see Table I) are studied first in difference among obtained results in Table VI is the number of
Sections V-A-C for illustrating the performance of the proposed wind generators (10, 11 and 9). The wind generators can provide
method. The same reliability test system with the weather data in sufficient electricity at night in winter but most loss of loads
Orchid Island (see Table II) are further investigated for showing occur in the daytime with the peak load. Hence, although the
the robustness of the proposed method in Section V-D. Discus- numbers of wind generators are different, the LOLPs are almost
sions are given in Section V-E. the same and bound by its limit.

A. Comparison Between Chronology-Based and B. Impacts of Different LOLP Constraints (Mt. Jade)
Markov-Based GAs (Mt. Jade) In this section, impacts of different ’s on the cost
In this section, comparative results obtained from the and iteration performance are examined. Let CO be 200
chronology-based and Markov-based GAs are shown. Different 000 kg/yr. As shown in Table VII, when high reliability (small
numbers of states for WTG, PV, and load are also ) is requested, the investment cost becomes high. As
considered, as shown in Table VI. Suppose that and can be seen, when the LOLP constraint becomes stricter for high
CO are 0.03 and 200 000 kg/yr, respectively. Table VI illus- reliability, the CPU time and number of iterations do not make
trates the comparative studies. As can be seen, the CPU times any difference. Fig. 3 illustrates the iteration performance for
(1.19 and 1.60 s) required by the Markov-based GA are much these three different scenarios. Please note that the wind power
smaller than that (95.3 s) required by the chronology-based is sufficient at night in the winter, and the solar power is suffi-
method. Both LOLP and CO constraints for all methods are cient in the daytime in the summer. The loss of load (electricity
fulfilled. It can also be found that the cost (U.S.$225 595) for deficit) occurs in different scenarios in the second and third rows
the case with 14 load states is lower than that (U.S.$229 483) of Table VII.
with 10 load states and that (U.S.$227 402) with chronological
data. Fig. 2 illustrates the iteration performance of these three C. Impacts of Different CO Constraints (Mt. Jade)
methods. In this section, impacts of different CO ’s on the cost and
GA is able to gain a global optimum but does not always iteration performance are examined. Let be 0.03. As
guarantee to yield one. Each result shown in Table VI is the shown in Table VIII, when a small CO is requested, the in-
best solution among 50 runs. Chronology-based and Markov- vestment cost will become high but the fuel cost will become
based GA require 8760 and 520 (or 728) iterative loops for a low. As can be seen, when the CO constraint becomes stricter
646 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 27, NO. 2, APRIL 2012

TABLE VII
COMPARISON BETWEEN DIFFERENT LOLP CONSTRAINTS

Fig. 5. Iteration performance for two methods (Orchid Island).

TABLE IX
COMPARISON BETWEEN CHRONOLOGY-BASED AND MARKOV-BASED
METHODS (ORCHID ISLAND)

Fig. 4. Iteration performance for three different CO ’s.

TABLE VIII
COMPARISON BETWEEN DIFFERENT CO CONSTRAINTS

2) The case solved by the chronology-based GA evaluates

CO and the adequacy of power generation for 8760 h.
More CPU time may be required by using the chronology-
based GA in case more detailed data (e.g., 17 520 data of a
30-min interval) are studied. Furthermore, in the planning
stage, there are generally many plans (e.g., more than 200
scenarios) to be studied. These results will be discussed
for low emission, CPU time and the number of iterations do not step by step to achieve several applicable plans for making
make any difference. Fig. 4 illustrates the iteration performance policies at the final stage. Reducing the CPU time is there-
for these four different scenarios. fore essential.
3) In this paper, the premise that the optimal number of units
D. Studies on the Optimal Sizing of Wind/PV/Diesel is obtained ahead of time is made. In other words, the fore-
Generation for Orchid Island casted 8760-h wind speed, irradiation, and load profile are
used to determine the optimal numbers of different dis-
In this section, the weather statistics, as shown in Table II, for tributed generations. Therefore, accuracy of the models for
Orchid Island will be used to study the optimal sizes of hybrid the load profile, wind speed, and irradiation and the time
wind/PV/diesel units. The IEEE reliability test system (loads for step is very crucial. This issue was discussed in [3]–[11].
8760 h) [12] is also utilized. Table IX shows the comparative The load profile used in this paper is the same as that in
studies between different numbers of Markov states. As shown [3] and [4], and the weather model (8760 h) used in this
in Tables I and II, the irradiation in Orchid Island is not as high paper is the same as [5], [9], and [10]. References [3]–[6]
as that in Mt. Jade. The number of installed WTGs is therefore and [8]–[10] also considered the same time step (1 h) that
much larger than that of installed PVs for reducing the cost and is used in this paper.
satisfying LOLP and CO constraints. Fig. 5 illustrates the iter-
ation performance of the two methods. Table IX illustrates the
results indicating the efficiency of the proposed method. VI. CONCLUSION
A new method using the Markov-based GA for determining
E. Discussions optimal sizes of hybrid wind/PV/diesel units in a stand-alone
From the simulation results for both Mt. Jade and Orchid Is- power system is presented in this paper. The presented method
land in Taiwan, three comments can be made as follows. considers cost minimization, and both reliability and CO emis-
1) The CPU times required should be similar for both small sion constraints. The Markov-based GA can help reduce CPU
and large-scale systems. The reason is that the problem time greatly and provides competitive cost. The CPU times re-
formulation is independent of the numbers of buses and quired for the proposed method do not make any difference for
lines. different LOLP and CO emission constraints. It can also be
HONG AND LIAN: OPTIMAL SIZING OF HYBRID WIND/PV/DIESEL GENERATION 647

concluded that strict LOLP and CO emission constraints will [12] Reliability Test System Task Force of the Application of Probability
create high cost. Methods Subcommittee, “IEEE reliability test system,” IEEE Trans.
Power App. Syst., vol. PAS-98, no. 6, pp. 273–282, Nov. 1979.
[13] Y. Y. Hong and S. F. Wei, “Multi-objective under-frequency load shed-
REFERENCES ding in an autonomous system using hierarchical genetic algorithms,”
[1] Kyotol Protocol to the United Nations—Framework Convention on Cli- IEEE Trans. Power Del., vol. 25, no. 3, pp. 1355–1362, Jul. 2010.
mate Change, Dec. 1997. [14] H. A. Taha, Integer Programming: Theory, Application and Computa-
[2] The Heads of State, Heads of Government, Ministers, and other heads tion. New York: Academic Press, 1975.
of delegation present, “Copenhagen Accord,” presented at the United [15] J. C. Bezdek, R. Ehrlich, and W. Full, “FCM: The Fuzzy c-means
Nations Climate Change Conf., Copenhagen, Denmark, Dec. 18, 2009. clustering algorithm,” Comput. Geosci., vol. 10, no. 2–3, pp. 191–203,
[3] R. Billinton and R. Karki, “Capacity expansion of small isolated power 1984.
systems using PV and wind energy,” IEEE Trans. Power Syst., vol. 16, [16] T. Manco and A. Testa, “A Markovian approach to model power avail-
no. 4, pp. 892–897, Nov. 2001. ability of a wind turbine,” in Proc. IEEE PowerTech Conf., Lausanne,
[4] R. Karki and R. Billinton, “Considering renewable energy in small Switzerland, Jul. 2007, pp. 1256–1261.
isolated power system expansion,” in Proc. Canadian Conf. Elect. [17] A. Shamshad, M. A. Bawadi, W. M. A. Wan Hussin, T. A. Majid, and
Comput. Eng., May 4–7, 2004, pp. 367–370. S. A. M. Sanusi, “First and second order Markov chain models for
[5] B. Ai, H. Yang, H. Shen, and X. Liao, “Computer-aided design of synthetic generation of wind speed time series,” Energy, vol. 30, pp.
PV/wind hybrid system,” Renew. Energy, vol. 28, pp. 1491–1512, 693–708, 2005.
2003.
[6] D. B. Nelson, M. H. Nehrir, and C. Wang, “Unit sizing and cost analysis
of stand-alone hybrid wind/PV/fuel cell power generation systems,” Ying-Yi Hong (SM’00) received the Ph.D. degree in electrical engineering from
Renew. Energy, vol. 31, pp. 1641–1656, 2006. the Institute of Electrical Engineering, National Tsing-Hua University, Taiwan,
[7] J. Kabouris and G. C. Contaxis, “Autonomous system expansion plan- in 1990.
ning considering renewable energy sources—A computer package,” From 1991 to 1995, he was an Associate Professor in the Department of Elec-
IEEE Trans. Energy Convers., vol. 7, no. 3, pp. 374–381, Sep. 1992. trical Engineering, Chung Yuan Christian University, Chung Li, Taiwan. He was
[8] J. Kabouris and G. C. Contaxis, “Optimum expansion planning of promoted to Full Professor in 1995. His areas of interest are power system anal-
an unconventional generation system operating in parallel with a ysis, power-quality analysis, and AI applications.
large scale network,” IEEE Trans. Energy Convers., vol. 6, no. 3, pp. Prof. Hong was the Chair of the IEEE Power and Engineering Society Taipei
394–400, Sep. 1991. Chapter in 2001. He is an Institute of Engineering and Technology Fellow.
[9] H. Lund, “Large-scale integration of optimal combinations of PV, wind
and wave power into the electricity supply,” Renew. Energy, vol. 31, pp.
503–515, 2006.
[10] S. Diaf, D. Diaf, M. Belhamel, M. Haddadi, and A. Louche, “A method- Ruo-Chen Lian received the B.Sc. degree in electrical engineering from the
ology for optimal sizing of autonomous hybrid PV/wind system,” En- Ching Yun University of Technology and Science, Chung Li, Taiwan, in 2008
ergy Policy, vol. 35, pp. 5708–5718, 2007. and the M.Sc. degree in electrical engineering from Chung Yuan Christian Uni-
[11] J. K. Kaldellis, “Optimum autonomous wind-power system sizing for versity, Chung Li, in 2010.
remote consumers using long-term wind speed data,” Appl. Energy, vol. His research interests include the applications of artificial intelligence and
71, pp. 215–233, 2002. power system analysis.