You are on page 1of 6

Decentralized Load Frequency Control for

Load Following Services


Dulpichet Rerkpreedapong, Student Member, IEEE, and Ali Feliachi, Senior Member, IEEE

Abstract--This paper proposes a decentralized controller for according to their load following contracts [3-4]. Therefore,
the load frequency control operated as a load following service. the parameters of the proposed controllers are simultaneously
Decentralization is achieved by developing a model for the inter- designed to control the generated power of each generating
face variables, which consist of frequencies of other subsystems. unit to meet contractual requirement. The effectiveness of the
To account for the modeling uncertainties, a local Kalman filter proposed controllers is demonstrated using two test systems,
is designed to estimate each subsystem’s own and interface vari-
ables. The controller uses these estimates, optimizes a given per-
each consisting of a three-area power system. The first system
formance index, and allocates generating units’s outputs accord- has one generator in each area, and the second has one area
ing to a deregulation scenario. Two test systems are given to with two units operating under a deregulation scenario.
illustrate the proposed methodologies.
II. DYNAMIC MODEL
Index Terms--Load frequency control, Automatic generation
control, Decentralized control, Kalman filter, Estimation, Ancil- A large interconnected power system consists of a number
lary services, Deregulation. of subsystems or control areas. Each area can be modeled in
great details depending on the generators models and their
I. INTRODUCTION prime movers. But, to illustrate the proposed idea, a simple
dynamic model, shown in Fig. 1, is presented in this section.
T HE power system consists of several interconnected con-
trol areas where each one is traditionally responsible for
its native load and scheduled interchanges with neighboring
The test system has a more elaborate model.
1
Ri
areas. Load frequency control (LFC) or automatic generation ∆PDi
control (AGC) is the mechanism by which the energy balance 1 ∆PVi 1 ∆PTi 1 ∆f i
is maintained. Under deregulation, such a mechanism can be + 1 + sTHi 1 + sTTi
+ D i + sTPi
used as a load following service operated by the regulating Governor T urbine
units according to a given contract. The regulating unit in a Bi
control area changes may change its output to match the ∆PCi = ui
∑ ∆Ptieij +
power demands in other areas, as the contract desires. j

Conventionally, the area control error (ACE), which is a ACEi


combination of a frequency error (∆f) and a tie-line power
Fig. 1. Block diagram of the ith-generating unit.
error (∆Ptie), reflects the control area’s performance, and is
used as the input to the load frequency controllers. Such con- PTi: turbine power PVi: governor valve
trollers are PI (Proportional-Integral) controllers whose pa- PCi: governor set point PDi: power demand
rameters are tuned using lengthy simulations and trial-and-
fi: frequency ΑCEi: Area control error
error approaches. Several optimization techniques have been
Ptie,ij: tie-line power between area i and j
proposed to solve this problem, but they require information
∆: deviation from nominal values
about the entire system rather than local information [1-2].
This paper proposes a completely decentralized LFC
The state space model for this system is given by:
scheme for load following services. A model for the interface
variables, which consist of frequencies of other control areas x& i = Ai xi + Bi u i + Gi z i + Fi ∆PDi (1)
or subsystems, is developed. To account for the modeling where
uncertainties, a local Kalman filter is designed to estimate
each subsystem’s own and interface variables from only local  Di 1 1 
 −T 0 − 0
 0 
measurements, namely area frequency and tie-line power. The  Pi T Pi T Pi 
deregulation scenario considered here assumes that generat-  1 1 
 0 
0 − 0 0  1 
ing units in each area supply a portion of the regulated power  TTi TTi 
  , Bi =  
Ai =  − 1 0 −
1
0 0  T Hi 
This work was sponsored in part by National Science Foundation under  R i T Hi T Hi   0 
grant ECS-9870041 and in part by US DOE/EPSCoR WV State Implementa-
 N   
tion Award.
 2π ∑ Tij 0 0 0 0  0 
 j =1 
The authors are with the Lane Department of Computer Science & Elec-  j ≠i 
trical Engineering, West Virginia University, Morgantown, WV 26506.  B i 0 0 1 0 

0-7803-7322-7/02/$17.00 © 2002 IEEE


1252
0-7803-7322-7/02/$17.00 (c) 2002 IEEE
o
 1 
T
u xi = u io = ∆PCi
o
(7)
Gi = [0 0 0 − 2π 0]T , Fi = − 0 0 0 0
 TPi  Define xi′ = xi − xio (8)
 
N u i′ = u xi − u io (9)
 N ,
x iT =  ∆f i ∆PTi ∆PVi ∑ ∆Ptie ,ij ∫ ACE i
z i = ∑ Tij ⋅ ∆f j
 j =1  j =1
 j ≠i  j ≠i Substitute eq. (8) and (9) into (4), a new system is obtained:
xi: State variables of the ith area
ui: Input of the ith area, zi: Interface variables x& i′ = Ai xi′ + Bi u i′ (10)
Tij : Synchronizing power coefficient of tie-line i-j
TPi: Generator’s time constant Subsequently, the controller parameters (Ki) in (2) are
TTi: Turbine’s time constant, THi: Governor’s time constant designed using LQR with the following performance index.
Ri: Droop characteristic, Di: Damping coefficient
N: Number of interconnected areas ∞
(
J i = ∫ xi′T Qi xi′ + u i′T ri u i′ dt ) (11)
0
III. CONTROL DESIGN
where
The proposed controllers are designed for a given state Ji: Performance index of the ith subsystem
space model using an LQR (linear quadratic regulator) ap- Qi : System weighting matrix
proach. It is known that the LQR has good gain and phase ri: Input weighting matrix
stability margins, but an accurate model is needed and all of
its state variables are essential for its implementation. This is The optimal controller is:
not suitable for a decentralized control structure because of u i′ = − K i xi′ (12)
the interface variables. In this paper, local state feedback or u xi = − K i xi + u io + K i xio (13)
gains (Ki) are designed using LQR, but an additional feed-
forward gain (KDi) is separately designed to cancel the effects
In a deregulated power system, the generated power from
of the interfaces [5].
generating units will be allocated according to given load
following contracts shown as
A. Control Design with Available State and Interface
Variables
 ∆PG1  α11 α12 α1m   ∆Pdc1 
∆P  α α α  
In this section, it is assumed that both state and interface  G 2  =  21 22 2 m   ∆Pdc 2 
variables are available for feedback. Then, using the state (14)
 M   O  M 
space model (1), a controller (ui) is designed as:     
 ∆PGn  1α4
n1 α n 2
44424444
α nm  ∆Pdcm 
3
u i = − K i xi + K Di z i = u xi + K Di z i (2) α
where where
KDi: Interface cancellation gain ∆PGi: Required change in pu MW of the ith generating unit
Ki : Stabilizing feedback gain, u xi : Stabilizing input ∆Pdcj : Change in demand of the jth distribution company
αij: Contract factor that indicates ratio of required change
then the closed loop system is expressed by: in pu MW of the ith generating unit to change in de-
mand (∆Pdcj) of the jth distribution company
x& i = Ai xi + Bi u xi + (Bi K Di + Gi )z i + Fi ∆PDi (3)
 ∆PD1   β 1 0 L 0   ∆PG1 
 ∆P   0  ∆PG 2 
The gain (KDi) interfaces are designed later to cancel the in-  D2  =  0 β 2 (15)
terface variables. Hence, the closed-loop system now has the  M  M O M  M 
form:     
∆PDN  0 0 L β N   ∆PGn 
x& i = Ai xi + Bi u xi + Fi ∆PDi (4)
β i = [1 1 L 1]1×k (16)
i
For a load frequency control problem, a non-zero set point,
i.e., steady state condition, is present when there is a change ∆PDi : Total change in demand for which the ith area are
in power demand (∆PDi). The set point is given by: responsible
ki: Number of generation units of the ith area
x& io = 0 = Ai xio + Bi u io + Fi ∆PDi (5)
yielding: Consequently, designing the controller parameters (Ki)
must also satisfy contract-based constraints. At a desired set
xi = xio (6) point,

1253
0-7803-7322-7/02/$17.00 (c) 2002 IEEE
u io = ∆PCi
o o
= ∆PGi = − K i xio (17) ing the required state and interface variables using only avail-
able measurements at the expense of some performance deg-
xio = − ( Ai − Bi K i )−1 Fi ∆PDi (18) radation. The structure of the proposed controller using esti-
where mates obtained from a Kalman filter is illustrated in Fig. 3.
xio : Set point of state variables wi(t) vi(t) PDi

u io : Set point of inputs


zi(t) xi(t)
Subsystem
ui(t) = ∆P Ci yi(t)
The input weighting matrix (ri) used for designing Ki will (area #i)
be selected by “lsqnonlin”, a search routine embedded in the
optimization toolbox in MATLAB, to satisfy eq. (17) and
(18). ^
xi(t)
+ -Ki Kalman
∆P di β i (List of generating
^z (t) Filter
units of each area) KDi i

Load following
Contracts Fig. 3. Decentralized control structure for the ith subsystem.
(αij)
wi : Plant noise, vi: Measurement noise
∆P Gi , ∆PDi ui: Control input, yi: Output
System
parameters Lsqnonlin
Ai , B i , Fi To design a Kalman filter that will estimate both local and
interface variables a dynamical model for these variables is
desired. This model is obtained by introducing the dynamics
ri , Ki
of the interface variable (zi), a combination of the deviations
of frequencies from the other areas, in the following form:
Fig. 2. Flowchart of determination of control parameters (Ki).
z&i = w fi (23)
Next, the interface cancellation gain (KDi) is designed to
cancel the effects of the interface on the integral of the area
control error, which is one of the state variables. where wfi is a fictitious white noise. The reason for this as-
sumption comes from the nature of the area frequencies
~ ~ whose deviations keep oscillating around zero, and are
yi = ∫ ACEi = Ci xi (19)
bounded when NERC’s performance standards are met.
However, the variance of the fictitious noise must be properly
The effect of the interface on the considered output ( ~
yi ) is chosen to increase the accuracy of the above model. In this
paper, the value of this variance is chosen at 0.001 by trial-
~
y zi (∞ ) = − Ci ( Ai − Bi K i )−1 (Bi K Di + Gi )z i
~ (20) and-error.
Augmenting the model given by (1) by adding the interface
dynamics given by (23) gives the following model:
then the interface cancellation gain (KDi) is selected to make
eq. (20) equals to zero:
x& i = Ai xi + Bi u i + GiWi + Fi ∆PDi (24)
~ ~
Ci ( Ai − Bi K i )−1 Bi K Di = − Ci ( Ai − Bi K i )−1 Gi (21) y i = C i xi + vi (25)
where
~
[
K Di = − Ci ( Ai − Bi K i )−1 Bi ]−1 ~
Ci ( Ai − Bi K i )−1 Gi (22)
x   wi 
xi =  i  , Wi =   , Ai =  i
A Gi 

B  F 
, Bi =  i  , Fi =  i 
 zi   w fi  0 0 0 0
From eq. (21), interface cancellation gain (KDi), however, can T 1 0 0 0 0
0 0 1 0 0 0 
be obtained as far as the number of inputs are not less than Gi =   , Ci = 0 0 0 1 0 , Ci = [Ci 0]
that of the outputs. 0 0 0 0 0 1 0 0 0 0 1
B. Control Design with Estimation of State and Interface xi : Augmented state vector
Variables
A Kalman filter, based on (24) and (25), is designed to
In an interconnected power system, not all the state vari- estimate the augmented state vector. It is given by:
ables are measurable, and the interface variables cannot be  x&ˆ 
&
obtained from local measurements. In this paper, a local xˆ i =  i  = Aˆ i xˆ i + Bi u i + Li y i + Fi ∆PDi (26)
Kalman filter is used to overcome this limitation by estimat-  z&ˆ i 

1254
0-7803-7322-7/02/$17.00 (c) 2002 IEEE
Aˆ i = Ai − Li Ci (27) The system matrix (A) can be written as
~ ~
where Li is the Kalman gain that can be determined by a A = A+G (32)
MATLAB function called “KALMAN.” where
The full dynamical model of the ith subsystem including ~
A : Ideal system matrix without interface consideration, i.e.,
the dynamics of the actual system in (1) and the Kalman es- ~
A = A where any Gij = 0
timator in (26) is expressed in (28).
~
G : Interface matrix
 x& i   Ai 0   xi   Bi  Gi   Fi 
 &ˆ  =  ˆA   xˆ  +  B u i +  0  z i +  F  ∆PDi (28) The closed loop system is:
 xi   Li Ci i  i   i     i
x& = ( A − BK ) x + F∆PD = Acl x + F∆PD (33)
In Fig. 3, the input of the subsystem using the estimates of
state variables and interface is defined as ( ~
) ~ ~
Acl = A − BK + G = Acl + G
~
(34)

u i = − K i xˆ i + K Di zˆ i (29) The performance of the composite closed-loop system (Acl)


is analyzed based on its eigenvalues. Let µ be any eigenvalue
IV. PERFORMANCE ANALYSIS of the composite closed-loop system, i.e. an eigenvalue of Acl,
~
and λi be the ith eigenvalue of Acl . The objective of this sec-
The stability of the entire interconnected system, called tion is to estimate the maximum departure of µ from the ei-
here the composite or actual system, when the proposed con- ~
genvalues of Acl . For this purpose let:
trollers are implemented, is of a major concern rather than
~
that of individual subsystems. The composite system has the T −1 Acl T = D = diag {λ1 , λ 2 , L , λ n~ } (35)
following state-space model and control input: then
~ ~
x& = Ax + Bu + F∆PD (30) T −1 Acl T = T −1 Acl T + T −1GT = D + Γ (36)
u = − Kx (31)
Gershgorin Circle Theorem [6] states:
where
 n~ 
& T T ℑ i = d ∈ C : d − λ i ≤ ∑ Γij  (37)
x =  x&1T
& &
x& 2T L x& TN xˆ1T xˆ 2T L xˆ N  
  j =1
ℑi : Gershgorin disk’s region
u = [u1 u2 L u N ]T , ∆PD = [∆PD1 ∆PD2 L ∆PDN ]T
C : Complex domain
 A1 G12 L G1N 0 0 L 0  Γij : (i-j) entity of the matrix Γ
G A2 G2 N 0 0 M 
 21 From Gershgorin’s theorem, a Gershgorin disk’s region
 M O M M O  gives an estimate of the maximum departure of the composite
 
G GN 2 L AN 0 L 0  closed-loop system eigenvalues (µ) from the eigenvalues (λi)
A =  N1 of the closed-loop control area model. Since the strength of
 L1C1 0 L 0 ˆA
1 0 L 0 
 ˆA  the interconnection is preset, one can guarantee closed-loop
 0 L2 C 2 0 0 2 0  stability of the composite system by properly designing the
 M O M M O M  area controllers.
 
 0 0 L LN C N 0 0 L Aˆ N 
V. CASE STUDY
T
 B1T 0 L 0 B1T 0 L 0 
  A power system, which consists of three control areas in-
0 B2T 0 B2T
B=
M M  terconnected through a number of tie-lines as shown in Fig. 4
 M O M O 
  is used to illustrate the proposed idea.
T T
 0 L BN 0 L B N 
Gij = Gi Tij S j , S j = [1 0 0 L 0] tie-line
1442443 Area 1 Area 2
number of rows of x j

 K1 0 0 L
 
0 K2 M 
K = 0 NxN , K i = [K i − K Di ] Area 3
 M O 
 
 0 L K N  Fig. 4. A three-area power system.

1255
0-7803-7322-7/02/$17.00 (c) 2002 IEEE
The generating units and their prime mover models within The eigenvalues of the closed-loop, ideal (A ~
cl ) and com-
subsystems are shown as posite (Acl), systems are obtained but only eigenvalues close
to the imaginary axis are plotted in Fig 6. These two sets of
1
R eigenvalues are very close, and thus justify the application of
Governor
the proposed controllers as a feasible and effective decentral-
T2
T3 ized control structure.
_ +
∆Pc
+
Σ
1 1
1−
T2 +
Σ VI. SIMULATION
1 + sT1 1 + sT2 T3

The performance of the proposed controllers is assessed


+
Σ
+
Σ
+
Σ through simulation of two test systems: 1) a three-area power
+ + + system with a single generator in each area, and 2) a test sys-
K1 K2 K3 K4
tem similar to the previous one except area 1 now has two
1 1 1 1 generating units operating under a deregulation scenario.
1 + sT4 1 + sT5 1 + sT6 1 + sT7

Turbine
A. Test system #1
+ 1
∆PT _
Σ
_ D + sTP ∆f In this test system, the contract matrix (α) is identity, and
∆PD
∑ ∆Ptie Generator
each area has a single generating unit whose data are given in
Table 1. The following unit-step changes in power demands
Fig. 5. Generating unit and prime mover models. are applied: ∆Pdc1 = 200 MW, ∆Pdc2 = -100 MW and ∆Pdc3 =
150 MW. The MVAbase is 2000. The area control error
Table 1. Data for a three-area power system ~
( )
(ACE) of the closed-loop ideal system A cl and the closed-
Data Area 1 Area 2 Area 3 loop composite system are shown in Fig. 7.
Rating (MW) 1000 750 2000
Droop characteristic: R (%) 5 4 5
Damping: D (pu MW/Hz) 20 15 18
Constant of inertia: H (sec) 5 5 5
T1 2.8 3 2.5
T2 1 0 0
T3 0.15 1 1
T4 0.2 0.4 0.5
T5 6 0 5
T6 7 0 0
T7 0.5 0 0
K1 0.2 1 0.4
K2 0.2 0 0.6
K3 0.4 0 0
K4 0.2 0 0

Synchronizing power coefficient of tie line i-j:


T12 = 60 MW/rad, T13 = 200 MW/rad, T23 = 100 MW/rad
Fig. 7. Area control error for ideal and actual systems.

O : Ideal system
Later, the frequency deviation (∆f) of each area regulated
X : Actual system
by the proposed controllers is shown in Fig. 8.

Fig. 6. Eigenvalues of the closed-loop ideal system and actual system. Fig. 8. Frequency deviation of actual system.

1256
0-7803-7322-7/02/$17.00 (c) 2002 IEEE
B. Test system #2

This test system is more realistic than the previous one.


Now area 1 has an additional generating unit identical to the
one in area 2. It is used to demonstrate the ability of the pro-
posed controllers to allocate generating units’s outputs ac-
cording to a given load following contract described as

 ∆PG1  0.6 0.2 0   0.05 


∆P     ∆Pdc1   
 G 2  = 0.4 0 0.1 ∆P  =  0.0475  pu
 ∆PG 3   0 0.8 0.1  dc 2  − 0.0325 Fig. 11. Frequency deviation for test system #2.
     ∆Pdc3   
∆PG 4  140 0 0.8   0.6  VII. CONCLUSION
42443
α
Later the identical changes in demands in the previous test This paper proposes a decentralized controller for the load
system are applied. Then the total change in demand for frequency control operated as a load following service. Decen-
which each area has to be responsible can be obtained as tralization is achieved by developing a model for the interface
 ∆PG1  variables, which is a combination of frequencies of other sub-
 ∆PD1  1 1 0 0    0.0975  systems. To account for the modeling uncertainties, a local
∆P  = 0 0 1 0 ∆PG 2  = - 0.0325 pu Kalman filter is designed to estimate each subsystem’s own
 D2     ∆P    and interface variables. The controller uses these estimates,
 ∆PD3  0 0 0 1  G 3   0.06 
optimizes a given performance index, and allocates generat-
∆PG 4 
ing units’s outputs according to a deregulation scenario. The
After the simulation is completed, the changes in turbine performance of the proposed controllers is assessed through
power (∆PT) of each unit are shown in Fig. 9. eigenanalysis and simulation of two test systems, each con-
sisting of a three-area power system. The first system has one
generator in each area, and the second has one area with two
units operating under a deregulation scenario. It is shown that
the proposed technique gives good results.

VIII. REFERENCES
[1] C. E. Fosha, Jr. and O. I. Elgerd, “The Megawatt-Frequency Control
Problem: A New Approach Via Optimal Control Theory,” IEEE Trans-
actions on Power Systems, vol. 89, no. 4, pp. 563-577, April 1970.
[2] M. L. Kothari, N. Sinha and M. Rafi, “Automatic Generation Control of
an Interconnected Power System Under Deregulated Environment,”
Power Quality’ 98, pp. 95-102, 1998.
[3] R. D. Christie and A. Bose, “Load Frequency Control Issues in Power
System Operations after Deregulation”, IEEE Transaction on Power
Systems, Vol. 11, No. 3, pp. 1191-1200, August 1996.
Fig. 9. Changes in turbine power for test system #2.
[4] R. D. Christie and A. Bose, “Load Frequency Control In Hybrid Elec-
tric Power Markets,” Proceedings of the 1996 IEEE International Con-
The area control error and frequency deviation of each area ference on Control Applications, pp. 432-436, September, 1996.
are shown as Fig. 10 and 11 respectively. [5] J. B. Burl, Linear Optimal Control, Addison Wesley Longman, Inc.,
1999.
[6] G. H. Golub and A.F. Van Loan, “Matrix Computations,” 2nd edition,
Baltimore: The John Hopkins university Press, 1989.

IV. BIOGRAPHIES
Dulpichet Rerkpreedapong received his MSEE from the CSEE Depart-
ment, West Virginia University in 1999. He is currently working towards a
Ph.D. in Electrical Engineering at WVU. His research interests are in power
systems control and operation, and power systems restructuring.

Ali Feliachi received the MS and PhD degree in electrical engineering


from Georgia Tech in 1979 and 1983 respectively. He joined the faculty of
Electrical and Computer Engineering at West Virginia University in January
1984 where he is now a Full Professor and the holder of the endowed Electric
Power Systems Chair position. His research interests are in modeling,
simulation, control and estimation of large-scale systems with emphasis on
Fig. 10. Area control error for test system #2
electric power systems.

1257
0-7803-7322-7/02/$17.00 (c) 2002 IEEE