IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 18, NO.
2, MAY 2003
Robust Load Frequency Control Using Genetic
Algorithms and Linear Matrix Inequalities
Dulpichet Rerkpreedapong, Member, IEEE, Amer Hasanović, Student Member, IEEE, and
Ali Feliachi, Senior Member, IEEE
Abstract—In this paper, two robust decentralized control design
methodologies for load frequency control (LFC) are proposed. The
first one is based on control design using linear matrix in-
equalities (LMI) technique in order to obtain robustness against
uncertainties. The second controller has a simpler structure, which
is more appealing from an implementation point of view, and it
is tuned by a proposed novel robust control design algorithm to
achieve the same robust performance as the first one. More specifi-
cally, genetic algorithms (GAs) optimization is used to tune the con-
trol parameters of the proportional-integral (PI) controller subject
to the constraints in terms of LMI. Hence, the second control
design is called GALMI. Both proposed controllers are tested on a
three-area power system with three scenarios of load disturbances
to demonstrate their robust performances.
Index Terms—Automatic generation control, decentralized con-
trol, genetic algorithms, , linear matrix inequalities, load fre-
quency control, robust control.
I. INTRODUCTION
I N a restructured system, power systems consist of several
interconnected control areas where each one is responsible
for its native load and scheduled interchanges with neighboring
areas. Load frequency control (LFC) is the mechanism by which
a balance between power generation and demand is satisfied.
In a regulation service, generation companies (Gencos) submit
their bids including prices (U.S. dollars per megawatt) and ramp
rates (megawatts per minute) to the market operator. After a
bidding evaluation, those Gencos selected to provide regulation
service must perform their functions according to the ramp rates
approved by the independent system operator (ISO).
Traditionally, the area control error (ACE), which is the com-
bination of area frequency bias times frequency deviation
and net power interchange error or net tie-line flow error
, is used as the input of the load frequency controller
whose objective is to control the ACE and the interconnection
frequency deviation in order to comply with the North Amer-
ican Electric Reliability Council (NERC)’s control performance
standards.
Usually, the load frequency controllers used in the industry
are proportional-integral (PI) type and are tuned online based
Manuscript received August 21, 2002. This research was supported in part by
the National Science Foundation under Grant ECS-9870041 and in part by the
U.S. DOE/EPSCoR WV State Implementation Award.
D. Rerkpreedapong is with the Department of Electrical Engineering,
Kasetsart University, Bangkok, Thailand.
A. Hasanović and A. Feliachi are with the Advanced Power Engineering Re-
search Center in the Lane Department of Computer Science and Electrical En-
gineering, West Virginia University, Morgantown, WV 26506-6109 USA.
Digital Object Identifier 10.1109/TPWRS.2003.811005
0885-8950/03$17.00 © 2003 IEEE
Authorized licensed use limited to: Iraq Virtual Science Library. Downloaded on October 27, 2009 at 05:13 from IEEE Xplore. Restrictions apply.
855
on trial-and-error approaches. Several optimization techniques
have been proposed to tune the control parameters using sim-
ulation of the entire system rather than just the control area
being studied. Some of them simply assume that all subsystems
are identical, which is not the case of actual power systems
[1]–[3]. Subsequently, a number of decentralized load frequency
controllers were developed to eliminate the above drawback
[4]–[8]. But most of them are complex state-feedback or high-
order dynamic controllers, which are not practical for industry
practices.
This paper proposes two robust decentralized LFC con-
trollers. The first one is based on theory, and results in a
high order controller. The second controller is a PI controller
tuned by a novel robust control design algorithm to achieve
the same robust performance as the first one, but it is more
appealing from an implementation point of view. In more
details, the second control design is first cast into the robust
control design in terms of linear matrix inequalities (LMI)
in order to obtain robustness against uncertainties [9], [10]. An
additional constraint is that the structure of the controller is
predefined as a PI type, which is ideally practical for industry.
In order to obtain the optimal controller parameters with
regards to the and controller structure constraints, genetic
algorithms (GAs), a powerful probabilistic search technique
[12], [13], is used to find the control parameters of the PI
load frequency controller at the upper level, whereas the linear
matrix inequalities resulting from constraints are solved
via optimization routines provided with MATLAB’s LMI
control toolbox [11]. The coordination between GAs and LMI,
proposed in this paper and called here the GALMI technique,
is needed because the formulated control design is in terms
of nonconvex optimization problem, which cannot be solved
by using LMI techniques alone. For both control designs, the
controlled variables include frequency deviation, area control
error, and governor load setpoint. The design objective is to
reduce unit wear and tear caused by equipment excursions, and
it addresses overshoot and number of reversals of the governor
load setpoint signal.
Technical background including robust control design
by LMI approach and genetic algorithms is given in Section II.
Next, a dynamic model of each control area for load frequency
control problem is presented in Section III. Afterwards, the
algorithm of the proposed robust control design using the
GALMI technique is illustrated in Section IV. Two types of
robust load frequency controllers, which are based on the
conventional control and GALMI tuned PI control, are
tested on a three-area power system with three scenarios of load
856 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 18, NO. 2, MAY 2003

where

Fig. 1. Close-loop system via robust H control.

disturbances, and their robust performance is demonstrated in The following lemma [11] relates control design to LMI.
Section V using nonlinear simulation. Finally, the conclusion Lemma1: the closed-loop RMS gain or norm of the
is given in Section VI. transfer function from to , , does not exceed ,
if and only if there exists a symmetric matrix such that
II. TECHNICAL BACKGROUND
This section gives a brief overview of control design (4)
when formulated in terms of LMI. Also, the description of the
GA optimization is given. (5)

A. Robust Control Design via LMI An optimal control design can be achieved by mini-
Over the past two decades, robust control theory has been mizing the guaranteed robust performance index subject to
useful and applied to control system designs that require robust- the constraints given by the matrix inequalities (4) and (5). The
ness against possible disturbances such as parameter uncertain- MATLAB’s LMI control toolbox provides the function “hin-
ties, system modeling errors, plant and measurement noises, and flmi” to solve an control problem directly. This function re-
external disturbances. turns the controller parameters as shown in (2) with the
One major objective of robust control is to synthesize a con- optimal robust performance index . The obtained controller
troller that would guarantee internal stability of the system when is a dynamic type, whose order is the size of the system, and
bounded perturbations are present. This subsection describes the hence, very large in general.
control design via LMI approach, which is less complex Some control applications require predefined controller struc-
than standard frequency-domain approaches that require sub- tures such as proportional-integral (PI), or output feedback con-
stantial mathematical and computational effort. trollers, which are different than the above full-order dynamic
Fig. 1 shows a classical block diagram of the robust con- controller shown in (2). This might modify the current
trol problem. The objective is to design a control law based on control design into a nonconvex optimization problem, which
the measured variables such that the effects of the disturbance cannot be directly solved by LMI techniques. As a result, GAs,
on the controlled variables as expressed by the infinity a powerful search technique, is utilized as an additional tool for
norm of its transfer function does not exceed a given solving such a hard optimization problem.
value defined as guaranteed robust performance. In order to
synthesize an controller via LMI approach, state space re- B. Genetic Algorithms
alizations of the system and controller are needed. GAs represent a heuristic search technique based on the evo-
They are given by (1) and (2), respectively. lutionary ideas of natural selection and genetics. GAs solve op-
State Space System Model timization problems by exploitation of a random search. When
searching a large state-space, or n-dimensional surface, a ge-
netic algorithm may offer significant benefits over the classical
optimization techniques such as linear programming or non-
linear constrained optimization. Although randomized, using
(1)
the historical information, they direct the search into the region
of better performance within the search space. GA based opti-
( , ) is stabilizable, and ( , ) is detectable.
mization techniques are designed to mimic processes in natural
State space controller model
systems necessary for evolution. Since in nature, competition
among individuals for resources results in the fittest individuals
dominating over the weaker ones.
(2) Individuals in GAs are in the form of character strings that
are analogous to the chromosome found in DNA. Each indi-
Combining the above equations results in the following closed- vidual represents a possible solution within a search space. A
loop system: number of individuals constitute a population. The individuals
in the population are then made to go through a process of evo-
lution, in order to produce a new generation of individuals that
(3) is closer to the optimal solution. In this paper routines provided

Authorized licensed use limited to: Iraq Virtual Science Library. Downloaded on October 27, 2009 at 05:13 from IEEE Xplore. Restrictions apply.
RERKPREEDAPONG et al.: ROBUST LOAD FREQUENCY CONTROL USING GENETIC ALGORITHMS 857

..
.

..
Fig. 2. Dynamic model of control area i. .

with GAOT [14] toolbox were used to implement GA optimiza-

tion as required by GALMI technique.

III. DYNAMIC MODEL ..

.
A large power system consists of a number of interconnected
control areas, each of which has several generating units. In this
section, a dynamic model of a generic control area including
generating units is shown in Fig. 2. The units in each area
are assumed to be coherent. To obtain the area frequency ,
generators are lumped. The state space model is given in (6) and
(7). Also, its controlled variables for control design where
is defined as (8) turbine power;
governor load setpoint;
(6) area frequency;
(7) area interface;
(8) area load governing characteristic;
deviation from nominal values;
or tie-line synchronizing coefficient between area & ;
turbine time constant;
(9) droop characteristic;
(10) ramp rate factor;
governor valve;
where
power demand;
net tie-line flow;
area aggregate inertia;
governor time constant;
frequency bias;
number of control areas.
In Section V, two types of robust LFC controllers are devel-
oped. The first one is based on a robust design using LMI
technique. The second controller is a GALMI tuned PI con-
troller to achieve similar robust performance as the first one.
The optimization objective for both is to minimize the effects of
disturbances given in (10) on the controlled variables as defined
in (9). This objective is

(11)

where is the infinity norm of the transfer function

from to .
Specifically, the design objectives are 1) to regulate frequency
deviation and area control error; and 2) to reduce excess maneu-
vering and unit wear and tear caused by equipment excursions.
The , , and in (9) are weighting coefficients chosen by
the designer. In this paper, they are 0.5, 1, and 500, respectively.
The large coefficient “500,” for example, is chosen to limit con-
trol effort associated with overshoot and number of reversals of
the governor load setpoint signal.

Authorized licensed use limited to: Iraq Virtual Science Library. Downloaded on October 27, 2009 at 05:13 from IEEE Xplore. Restrictions apply.
858 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 18, NO. 2, MAY 2003

IV. NOVEL ROBUST CONTROL DESIGN VIA GALMI

In practice, the LFC controller structure is traditionally a pro-
portional-integral (PI)-type controller using the ACE as its input
as shown in (12). In this section, the robust control design algo-
rithm for such a load frequency controller using GALMI tech-
nique is presented. The objective of the proposed design is to
tune the PI controller parameters to achieve the same robust per-
formance as the conventional design

(12)

(13)
(14)
From the above equations, the desired controller is only a
simple static output feedback controller, and it is much less com-
plex than the one obtained from the conventional control
design shown in (2). To determine the control parameter vector
, (7) is first substituted into (14), which results in (15).
Next, (15) is substituted into (6) and (8), and the closed-loop Fig. 3. Robust control design via GALMI technique.
system is finally obtained as (16)
(15)

(16)
where

Subsequently, is searched for in order to minimize the per-

formance index subject to robust control constraints given by Fig. 4. Three-area power system.
matrix inequalities (4) and (5). The successfully obtained con-
trol vector also guarantees the system stability, which is
TABLE I
the small-signal stability of the control areas in this case. ROBUST PERFORMANCE INDEX
However, a solution of the consequent nonconvex constrained
optimization problem cannot be achieved by using LMI tech-
niques alone. Therefore, the proposed GA optimization tech-
nique is utilized to search offline for the control parameters
of the PI load frequency controller at the upper level, whereas
the LMI control toolbox is used at the lower level to solve the
has three generating units that are owned by different generation
linear matrix inequalities given as constraints for robust
companies (Gencos). Two types of robust decentralized load fre-
control design. The offline hybrid optimization algorithm that
quency controllers, 1) robust control designed according
describes the GALMI technique is shown in Fig. 3.
to the procedure described in Section II; and 2) robust GALMI
The proposed algorithm gives the constant optimal robust
tuned PI control designed based on the proposed GALMI al-
control parameters , when the minimum robust perfor-
gorithm presented in Section IV, are implemented in each area.
mance index is achieved. According to Fig. 3, NUM is
The obtained robust performance indices of both designs
the number of generations that the GA procedure is required
are almost identical as shown in Table I. These results show
to evolve when searching for the best . This number is
no degradation on the GALMI tuned PI control design. But its
prespecified and should be large enough to ensure that the
structure is much simpler than the robust design, whose
global minimum of is found. The described control design is
order is the number of system states (or the size of ) that in-
applied in Section V to design load frequency controllers for a
creases with the modeling details and the number of units, and
three-area power system.
it can be very large. In this section, the performance of the ro-
bust PI controllers is compared with that of the dynamic
V. CASE STUDY controllers for three scenarios of load disturbances.
The test system, shown in Fig. 4, consists of three control For scenario 1, random load changes, shown in Fig. 5(a), rep-
areas, and its parameters are tabulated in Appendix. Each area resenting expected load fluctuations, are applied to the three

Authorized licensed use limited to: Iraq Virtual Science Library. Downloaded on October 27, 2009 at 05:13 from IEEE Xplore. Restrictions apply.
RERKPREEDAPONG et al.: ROBUST LOAD FREQUENCY CONTROL USING GENETIC ALGORITHMS
Fig. 5. System response for scenario 1. Solid (GALMI), dash-dotted (
control areas. The area control error (ACE), frequency deviation
, and governor load setpoint closed-loop responses
are shown in Fig. 5(b)–(d). From the results, both controllers
ramp generated power to match the load fluctuations effectively.
The performance of the GALMI tuned PI controllers is almost
identical to that of full order controllers. In addition, Fig. 6
shows the raise/lower signals allocated to all generating units in
area 1 conforming to their offered ramp rates.
For scenario 2, a large disturbance, a step increase in demand,
is applied to each area: MW, MW,
and MW. The purpose of this scenario is to test
the robustness of the proposed controllers against large distur-
bances. In fact, these large step changes in demand rarely occur
since a party, which causes a serious mismatch between actual
Authorized licensed use limited to: Iraq Virtual Science Library. Downloaded on October 27, 2009 at 05:13 from IEEE Xplore. Restrictions apply.
859
Fig. 6. Raise/lower signals of units in area1. Solid (GALMI), dash-dotted
(H ).
Fig. 7. System response for scenario 2. Solid (GALMI), dash-dotted (H ).
and forecast load, is likely to be penalized. Figs. 7(a) and (b)
show the responses of areas 1 and 2. The ACE and frequency de-
viation are effectively damped to zero with very small os-
H ). cillations by the GALMI tuned PI controllers. The control input
is also smoothly increased to the expected value without
overshoot and oscillations. These controllers perform as well as
the robust controllers. Above step changes in power de-
mand are considered significant because of the large overshoot
(0.1 Hz) in the frequency deviations.
For scenario 3, large step increases in demand are applied
to areas 2 and 3: MW, and MW.
To make the scenario drastic, it is assumed that LFC reset con-
trollers of areas 2 and 3 are out of service. As a result, the fre-
quency deviation of all areas cannot be driven to zero, which
causes the area interface , treated as a system disturbance, to
remain nonzero all the time. The purpose of the scenario is to
investigate the response of the system to such a severe condi-
tion, when only area 1 proposed LFC controller is kept active.
The responses of area 1, given in Fig. 8(a), show that the ACE is
860 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 18, NO. 2, MAY 2003

TABLE II
GENERATING UNIT PARAMETERS

TABLE III
GAOT TOOLBOX PROPERTIES

Fig. 8. System response for scenario 3. Solid (GALMI), dash-dotted ( H ). TABLE IV

PI CONTROL PARAMETERS FROM GALMI DESIGN

driven to zero successfully, and the governor load setpoint has

very small oscillations during transient and goes back to zero in
a very short time. The response of the governor load setpoint of
GALMI tuned PI controller is slightly degraded compared with
that of the controller. This is because the controller is
a high order dynamic controller where its order is as large as the
number of states of the system. In the paper, the controller is
of 9th order, but the GALMI controller is a simple PI controller.
Incidentally, the responses of ACE and frequency deviation of
the and GALMI tuned PI controllers are almost the same. In
addition, the responses of area 3 are shown in Fig. 8(b). Without
the LFC controller, the ACE cannot be driven back to zero.
VI. CONCLUSIONS
This paper proposes two robust decentralized control design
methodologies for load frequency control (LFC). The first
one is based on control design using LMI technique. The
second controller is a PI type, and is tuned to mimic the robust
performance of the first one. More specifically, instead of
solving the control problem to obtain a robust high-order
dynamic controller, the proposed GALMI technique, coordi-
nates GAs with the LMI control toolbox optimization in order
to obtain the control parameters and of a traditional PI
controller that satisfy the robust constraints. A three-area
power system is used as the test system with three scenarios
of load disturbances. The simulation results show that the
responses of GALMI tuned PI load frequency controllers are
almost the same as those of the robust controllers, which
have effective control performance and robustness against
possible disturbances.
APPENDIX
See Table II
Regulation requirement for each area 100 MW
Ramprate
regulation requirement
p.u./rad, p.u./rad, and
p.u./rad. See Tables III and IV.
REFERENCES
[1] M. L. Kothari, N. Sinha, and M. Rafi, “Automatic generation control
of an interconnected power system under deregulated environment,” in
Power Quality, 1998, pp. 95–102.
[2] V. Donde, M. A. Pai, and I. A. Hiskens, “Simulation and optimization
in an AGC system after deregulation,” IEEE Trans. Power Syst., vol. 16,
pp. 481–489, Aug. 2001.
[3] Y. L. Abdel-Magid and M. M. Dawoud, “Genetic algorithms applica-
tions in load frequency control,” in Proc. First Int. Conf. Genetic Algo-
rithms Eng. Syst.: Innovations and Applicat., 1995, pp. 207–213.
[4] H. Trinh and M. Aldeen, “Decentralized load-frequency control of inter-
connected power systems,” Proc. Inst. Elect. Eng. Int. Conf. Advances
in Power Syst. Contr., Oper. Manage., vol. 2, pp. 815–820, 1991.
[5] X. B. Chen and S. Stankovic, “Overlapping decomposition and decen-
tralized LQG control for interconnected power systems,” Proc. IEEE
Int. Conf. Syst., Man, Cybern., vol. 3, pp. 1904–1909, Oct. 1996.
[6] K. Y. Lim, Y. Wang, and R. Zhou, “Robust decentralised load-frequency
control of multi-area power systems,” Proc. Inst. Elect. Eng.—Gen.,
Transm. Dist., vol. 143, no. 5, pp. 377–386, Sept. 1996.
[7] K. Y. Lim, Y. Wang, G. Gua, and R. Zhou, “A new decentralized robust
controller design for multi-area load-frequency control via in complete
state feedback,” Optimal Control Applications & Methods, vol. 19, pp.
345–361, 1998.
[8] M. H. Rahi and A. Feliachi, “ H
robust decentralized controller for
nonlinear power systems,” in Proc. 30th Southeastern Symp. Syst.
Theory, Mar. 1998, pp. 268–270.
[9] P. Gahinet, “Explicit controller formulas for LMI-based H synthesis,”
in Amer. Cont. Conf., vol. 3, Jun. 1994, pp. 2396–2400.

Authorized licensed use limited to: Iraq Virtual Science Library. Downloaded on October 27, 2009 at 05:13 from IEEE Xplore. Restrictions apply.
RERKPREEDAPONG et al.: ROBUST LOAD FREQUENCY CONTROL USING GENETIC ALGORITHMS
[10] M. Chilali and P. Gahinet, “ H
design with pole placement constraints:
An LMI approach,” IEEE Trans. Automat. Control, vol. 41, pp. 358–367,
Mar. 1996.
[11] P. Gahinet, A. Nemirovski, A. J. Laub, and M. Chilali, LMI Control
Toolbox: The MathWorks, Inc., 1995.
[12] J. Holland, Adaptation in Natural and Artificial Systems. Ann Arbor,
MI: Univ Michigan Press, 1975.
[13] D. Goldberg, Genetic Algorithms in Search, Optimization, and Machine
Learning. Reading, MA: Addison-Wesley, 1989.
[14] C. Houck, J. Joines, and M. Kay, “A Genetic Algorithm for Function
Optimization: A Matlab Implementation,”, NCSU-IE TR 95-05, 1995.
Dulpichet Rerkpreedapong (S’00-M’03) received the B.Eng. degree from
Kasetsart University, Bangkok, Thailand, in 1995, and the M.S.E.E. and
Ph.D. degrees from the Lane Department of Computer Science and Electrical
Engineering at West Virginia University, Morgantown, in 1999 and 2003,
respectively.
He worked as a Lecturer at Kasetsart University until 1997 when he obtained
a Royal Thai scholarship to pursue his graduate studies. He was a Graduate Re-
search Assistant in the Advanced Power Engineering Research Center (APERC)
at West Virginia University, Morgantown, from 1999 until May 2003. Currently,
he is a faculty member in the Department of Electrical Engineering at Kasetsart
University. His research interests include power systems operation and control,
and power systems restructuring.
Authorized licensed use limited to: Iraq Virtual Science Library. Downloaded on October 27, 2009 at 05:13 from IEEE Xplore. Restrictions apply.
861
Amer Hasanović (S’02) was born in Tuzla, Bosnia and Herzegovina, in
1976. He received the B.S. degree from the University of Tuzla, Bosnia and
Herzegovina, in 1999, and the M.S. degree in electrical engineering from West
Virginia University, Morgantown, in 2001. He is currently pursuing the Ph.D.
degree at West Virginia University.
Currently, he is a Graduate Research Assistant in the Advanced Power Engi-
neering Research Center (APERC) at West Virginia University.
Ali Feliachi (SM’86) received the Diplôgme d’Ingénieur en Electrotechnique
degree from Ecole Nationale Polytechnique of Algiers, Algeria, in 1976, and
the M.S. and Ph.D. degrees in electrical engineering from Georgia Institute of
Technology, Atlanta, in 1979 and 1983, respectively.
Currently, he is a Full Professor and the holder of the Electric Power Systems
Chair endowed position at West Virginia University, Morgantown. He is also
the Director of the Advanced Power Engineering Research Center (APERC).
He has been a faculty member in the Lane Department of Computer Science
and Electrical Engineering at West Virginia University since 1984.