You are on page 1of 9

qr

,~~~ ~

REQUM *'e-W11H, J. Duncan Glover ~ ~


4m<el
t

ASW

fADVANCED LOAD FREHQUENCY CONTROL


Fred C. Schweppe Massachu tts Institute of Technology Cam ridge, Massachusetts

Service Corporation J7D ectricPower New York, New York

Abstract-The problem of designing a load frequency ntrol law which reduces transient frequency oscillations (swings) and reduces the number of control signals sent to power houses is investigated. A linear is developed, and a model of an aren discrete time, linear-plus-deadband, feedback control law is designed. Feedback variables include cumulative inadvertent interchange, frequency deviation, integral of frequency deviation, real power absorbed by loads, and governor-turbine variables. This linear-plus-deadband control is an application of a special case of a more general "set-theoretic" class of control laws. A simulation of two areas with two hydro sources is presented. The dynamic response to a step load change is determined for the case of no load frequency control, load frequency control presently used by power companies, and load frequency control designed in this paper. I. INTRODUCTION flJR

ete time o-second control period) rather than continuous time LFC is designed here. Thus it is difficult to compare the performance of the two control laws. The deadband portion of the control law is designed (Section 111.3) to reduce the number of LFC signals sent to power plants. The deadband logic is related to the "Error Adaptive Control Computer" (EACC), which was developed by Ross [31 and is presently used in some LFC laws. The differences between EACC and deadband logic are: 1) EACC logic is based on a probabilistic analysis. Deadband logic is based on a magnitude bound analysis. 2) EACC analyzes the area control error (and its integral). Deadband logic analyzes all of the state variables (or an estimate of them) of the area. It is assumed here that all of the (state) variables of the model are monitored (perfectly) and are available for control implementation. If not then a dynamic state estimator could be used to estimate the state Present load frequency control (LFC) laws adequately perform the variables. The state estimator would precede the LFC (as in Ref. 2). tasks they were designed to handle. However it is conceivable to use a Simulations of a two area power system with two hydro power more sophisticated LFC law (denoted advanced LFC) which can do sources are presented (Section IV). The dynamic response (frequency, more. This possibility is investigated here by developing such a control tie-line power, and source mechanical power) to a step load change for law and comparing its performance with LFC presently used (denoted the cases of no LFC, conventional LFC, and advanced LFC is determined. conventional LFC). The linear-plus-deadband control is an application of a special case Recently there have been other investigations [], [2] using mod- of a more general class of control laws designed using set-theoretic techern control theory techniques to develop new LFC laws. Fosha and Elgerd niques. Such set-theoretic controls are especially designed to handle [1] use a state variable model and the state regulator problem of op- problems with unknown-but-bounded disturbances. Set theoretic contional control theory to develop new feedback control laws. Cavini, trol is discussed in Ref. 4. Budge and Rasmussen [21 develop a control law which is related to Fosha and Elgerd's, but also includes a state estimator. II. MODEL It is desired that advanced LFC achieves the same objectives as conventional LFC; i.e. the control law should perform the following A model of an-area of an interconnected power system is presented. functions: The area consists of a group of power sources, loads, transmission lines 1) Each area regulates its own load fluctuations (if possible). within the area, and tie-lines which interconnect the area to the system. 2) Each area contributes to the control of system frequency. First each of the components (sources, loads, lines, ties) is modelled 3) In steady state, frequency and net tie-line power are returned separately; then the component models are connected to form the area to schedule in all areas (if all areas can regulate their own load fluctua- model. tions). TURBINE GOVERNOR In addition to the objectives of conventional LFC, it is desired that advanced LFC: 4) Reduces transient frequency oscillations (swings), without a lairge increase in the magnitude and speed of control. 5) Reduces the number of LFC signals sent to power plants, without compromising other LFC objectives. In order to design advanced LFC laws, a model of an area of an interconnected power system is developed (Section II). The (state) variables of the model include mechanical variables of sources, electrical frequency and angle deviations of the sources, load variables, and tie-line variables. The model is chosen to be linear and time-invariant, and also contains uncertainty. The uncertainty is assumed to be "unknown but bounded"; i.e. the uncertainty is contained in a specified bounded set, but is otherwise unknown. Fig. 1. Mechanical Dynamics - Hydro Source Using the model, a discrete time, "linear-plus-deadband," feedback control law is designed (Section III). The linear portion of the control law is designed (Section 111.2) to reduce swings, and is related to Fosha 1. Power Source and Elgerd's design. However, a different model is used here, and a dis-

vAf<sto~--

$t3

PA'il;

Paper T72 085-4, recommenided anid approved by the Power System Engineering Committee of the IEEE Power Engineering Society for presentation at the IEEE Winter Meeting, New York, N.Y., January 30-February 4, 1972. Manuscript submitted September 16, 1971; made available for printing November 15, 1971.

The dynamics of a power source consist of three parts: mechanical, electrical, and electromechanical coupling. The mechanical dynamics are represented by a linear model. The inputs of the model are the LFC of the source, u(t), and the change

2095

Authorized licensed use limited to: Sven del Pino. Downloaded on September 15, 2009 at 16:44 from IEEE Xplore. Restrictions apply.

(from 60 hertz) in electrical frequency of the source, Afs(t). The output is the change (from nominal value Pmnom) in mechanical power per unit delivered by the source, APms(t). A model of the mechanical dynamics of a hydro source [ 5 ] is shown in Figure 1 and is used in simulations presented here (Section IV). The electrical dynamics of the source are modelled by the following voltage source;

3. Load
A load is modelled here by representing its voltage and power characteristics. It is assumed that the voltage across each load is:

V (t)

,TCOS

eS (t)

VL (t) = f2 V 0 (t) = Wot 4-

6L

CoSQL (t)
(t)
=

WOt 4. bs t) = 2 n 60 WO w (t) = 2 nr f$ 9 = t 2 rr & f (t) AW J(t=


=

Qs (t)

A6w 1 (t)
(t)
-

WL (t)

-w

-t

dt

L (t)

dt

dt

&L(t)

(3)

ot +

= Lo

(t)

w
=

5-d S(

A s (t) =

(t)

bsnom

dt

dt

(t)

The load voltage is assumed to be sinusoidal with constant magnitude (RMS value V). The phase angle 6L(t) is assumed to be slowly time varying. The power absorbed by each load is modelled as:

It is assumed that the three-phase power system is balanced (balanced sources, loads, lines), and only equivalent one-phase models are presented. The magnitude of the sinusoidal voltage source is assumed to be constant (RMS value V). The phase angle 6s(t) is assumed to be slowly time varying compared to wot (65s(t) is a narrow band process modulating the phase of the source voltage). A6s(t) is the change (from nominal value bsnom) in phase angle of the source. A source reactance is also modelled, but is assumed to be "lumped" into the transmission line system. The electromechanical coupling is modelled by Newton's Second Law (swing equation).

fAPL (t)

= 0

(4)

where APL(t) is the change (from nominal value) in electrical power per unit absorbed by the load. Only step load changes (i.e. Eq. 4) are considered here. Also the dependence of load power on frequency (and voltage) is neglected. Kirchoff's laws and a knowledge of the transmission line configuration of the area can be used to determine the source electrical power as a function of source phase angles and load powers. Assuming small angle differences, a linearized relation is:

dtAWt)

=_

(t)

APms tt)
M

AP(.t )

(2)

SP(t)

(t)
LARL (t)J

its S(t)1

(5)

APS(t) is the change (from nominal value) in electrical power per unit delivered by the source. M is the normalized moment of inertia of the source, and D is the normalized damping constant.
2. Transmission Line
A transmission line is modelled as a constant reactance. Real and reactive losses on the line are neglected.

EL (t)]

APS(t) (an Ns vector) is the vector of (changes in) electrical powers (per unit) delivered by the Ns sources in the area. A6s(t) (Ns vector) is the vector of (changes in) source angles. APL(t) (NL vector) is the vector of (changes in) electrical powers (per unit) absorbed by the NL loads in the area. D = [5s PL] is a constant (Ns by Ns + NL) matrix which is determined from the system configuration (of sources, loads, and lines), and Kirchoff's laws.

Fig. 2. Block Diagram of an Area of an Interconnected Power System


2096

Authorized licensed use limited to: Sven del Pino. Downloaded on September 15, 2009 at 16:44 from IEEE Xplore. Restrictions apply.

4. Isolated Area Model


A complete model of an isolated area, consisting of Ns power sources, NL loads and NT transmission lines, is obtained by marrying the component models previously discussed. The result is a linear system which is shown in Figure 2 (disregard tie-line dynamics and "w" terms). It is assumed that the area is not connected to other areas (no tie-lines). The inputs to the area model consist of the vector of electrical load changes APL(t), and the vector of LFC signals, u(t).
Jal W-

x (t)

x
I

.w

(t)

[ MI,

-1 r(-) .. 13

(t) -WS (t) 311.iI (t) )., 11 (t)

(9)

5. Interconnected Area Model

If the area is connected to a power system via tie-lines, then the area is affected by electrical power flows on the tie-lines. The power flow on a transmission line (tie-line or otherwise) is a function of the difference in phase angles at the two nodes to which the line is cona nected. These two phase angles can be written as a function of load (10) d t x (t) = A x (t) + 2 O (t) + 5 (t) powers and source phase angles of the system. It is assumed, however, that models of only those sources and loads within the area are available. For this reason, only an approximate model x(t) is the "state" vector of the system, which consists of the mechanical of a tie-line power flow can be obtained. It is assumed, then, that a tie- variables of the sources, frequency and angle deviations of the sources, line power flow is a function only of load powers within the area, and load variables, and tie-line variables. u(t) is the vector of LFC signals. possibly some loads outside the area but electrically "near" the tie-line. Iw(t) is the vector of uncertainties, which is constrained via Eq. 9. The matrices A, B, and G are determined from Figure 5 and previous discussions. X (t5} A YXTl (t)l A discrete time version of Eq. 10 is:

Eq. 9 defines an ellipsoid 2w, which contains the vector w(t). The positive definite matrix W determines the "size" of the ellipsoid, and is determined from the bounds on the uncertain variables. The ellipsoid set is chosen for mathematical convenience. A model of an area of an interconnected power system, with uncertainty added, is shown in Figure 2. It is a block diagram of a linear, time-invariant model. In equation form, the model can be represented via:

dt LXT ( t)
(t)
=

AT

XT2

(t)J

(t) (6)

X( i e

+ e

) =-ccx( i 6)

Au( i C-)
V

VW( j G)
.1

( 1)

L,

t2(t

T.

"LL(t)

(7)

where the state x(je) is described at discrete times je, j = 0, 1, 2,. . at intervals of e seconds. The matrices ox, 13, and -y are given by:

Eqs. 6 and 7 represent a second order linear system. The input of the system is APL(t), and the output is APT(t), the net change (from nominally scheduled value) in tie-line power flow out of the area. The constant matrices A B H and HT2 are determined from a knowledge of the system configuration and tie-line oscillations (period and time constant). A model of an area of an interconnected power system, including tie-lines, is shown in Figure 2 (disregard "w" terms).
6. Uncertainty
Many assumptions and approximations are made to obtain the linear model of Figure 2. In order to account for modelling errors caused by these assumptions and approximations, the following variables are augmented:
t

0e=

-J %dr yA(e(r -r)


A

7eA(d.1. G 0_

(12)

The bound on the discrete-time uncertainty vector, yzje), is represented as in Eq. 9, with "t" replaced by "je".
7. Discussion

P-L (W-, API, (t) + RitL (t) (t) A ES (t) + MS (t)8 (E t ST E T MT (t) (t)
-u (t)

Figure 2 contains Ns sources. A source could represent one generator, one plant, or a group of "closely coupled" plants. Also the NT transmission lines and NL loads could represent all the lines and loads in the area, or a reduced set (e.g. just the EHV System, with the lower voltage system "lumped" into the loads). Thus the model is flexible, allowing for as much (little) complexity as desired. The variables of the model are written as changes from nominal values. The nominal values can be determined from a "load flow"

(t)

u(t)

(8)

equation. The purpose of the model is to develop advanced LFC laws. First the model is used to design control laws. Then it is used in simulations to analyze control law performance.
III. CONTROL

A vector of "uncertain" variables, wL(t), is added to the vector of load powers. wL(t) can be viewed as a small perturbation of APL(t) in order to account for the uncertain nature (inexact modelling) of the loads. Similarly "uncertain" vectors ws(t) and wT(t) are added to the source and tie-line power vectors. An "uncertain" vector yu(t) is added to the vector of LFC signals in order to account for the inexact modelling of the mechanical dynamics of the sources. It is assumed that the uncertain variables are bounded or constrained in magnitude by the following equation:

1. Description

This section discusses advanced LFC. Only the basic LFC signals, which are assumed to be dispatched every two seconds, are considered. Economic dispatch and adjustments of frequency schedule and net tieline power flow schedule (to control system time deviation and cumulative inadvertent interchange) are not considered. These can be viewed as refinements of the basic LFC which operate more slowly (minutes versus

2097

Authorized licensed use limited to: Sven del Pino. Downloaded on September 15, 2009 at 16:44 from IEEE Xplore. Restrictions apply.

seconds). The following control law is used to achieve the control objectives.
ii(t) =~ ~ ~ E 31(t) = s(je--i)

xV(j j
11

-+E)
-

olxl( je)+ -l'L


+- .82K2

+ -Wl

j )

(18)

if x (je) is not contained in 92 otherwise

A,

oQ - "9I I

(19)
(20)

je Lt - (j+1) e

(13)

J31

= J1

e= 2.0 seconds Eq. 13 is a discrete time, "linear-plus-deadband," feedback control law. Every two seconds LFC is either updated by an amount proportional to the change in state variables or left unchanged. K is a feedback gain matrix which linearly relates the state vector to the LFC vector. Q2 is a "deadband set" which specifies a region in state space. The LFC is left unchanged if the state vector is contained in O.

Assuming Eq. 18 defines a stable system (eigenvalues of oil less than one in magnitude) Qss is given by the solution to the following "set equation" [4].

ss -~SS=C

f2 ss 0 SO,1

AL

31

<l

(21)

2. Feedback Gain Matrix The feedback gain matrix is based on the discrete-time model, Eq. 11, and is determined from the state regulator problem of optimal control theory [ 6].
A quadratic cost function, J, is specified.

J -O

', (j) Q

(ji)+ u'(ja)Ru(Je)

(14)

I denotes vector sum; where a&l Qss denotes the set Qss translated by &1. 31 APL denotes the set consisting of the poillt (vector) 1 1 APL; and 11l2W denotes the set Qw translated by yl. In order to reduce the control action a reasonable approach is to stop updating LFC when the state is contained in Qss The LFC will not "track" the noise w(je) in this case. Thus Eq. 13 is used for LFC with a Unfortunately Qss given by Eq. 21, is not an ellipsoid, even though Qw is. For computational simplicity, an ellipsoid Qss3 which is an inner bound to Q2ss' is determined.

9ss,

ni ss -{i:[z1 - sSA LS jjzi-'.s > }


=
2 M -1 [I -kI Eq. =13

lize the (squared) magnitude of source frequency deviations. R is a positive definite, symmetric matrix which is chosen to penalize the (squared) magnitude of control signals. The control law which minimizes the cost J, subject to the model, Eq. ll,is

Q is a positive semi-definite, symmetric matrix which is chosen to pena-

(22)

rSS =
A

+Ct rS5;l

1
A

2ss is an ellipsoid with center css and matrix rss 2ss is contained in
Ss,

1U

(QE)

Lz (ii )

(15)

The linear-plus-deadband control law is

where K is the steady state solution to a discrete time Ricatti equation [6]. The uncertainty vector, w(je), is neglected in the design of K. After K is determined from Eq. 16, it is modified to include a feedback of cumulative inadvertent interchange, in order to regulate inadvertent interchange (so that the area regulates its own load fluctuations).

z;( ) j =t( i) if
(i
Pss.

[Z1 (

C)_ -csSst[v)si> 0 Di)


(23)

otherwise

3. Deadband Set The deadband set Q2 is based on the discrete time model, Eq. 11, and some set theory. Since the model is driven by the uncertainty vector, wvOe), a steady state is never achieved. However, a "steady state region" can be defined. For a given load change, APL, the state will, after a transient period, settle into a steady state set (if the area and its model are stable), which is denoted by Qss The model, Eq. 11, is rewritten.

The deadband region can be tuned to a particular area by adjusting Each adjustment provides a new degree of freedom, and a change in performance. The use of the ellipsoid set, Qss results in a more conservative control law (LFC signals are dispatched more often). It is assumed that the noise vector, w(je), is magnitude bounded for each time interval (Eq. 9, with "t" replaced by "je", describes a discrete time white noise process). Many other types of noise (periodic, exponential, frequency banded) can be handled by including noise dynamics in the model.

4. Role of Set Theory

l(je+e) = M1;X1(j ) *-&1 J- L +*2 ia(j)4+ iE

)(16)

Eq. 21 is a set equation involving mathematical concepts which are probably not familiar to most readers, and the derivation of Eq. 22 is a nontrivial exercise. Explanations are not given here because they would greatly expand the paper and are not essential to the basic con-

The load vector, APL, is written as an input in Eq. 16 (rather than being included in the state). xl(je) is a reduced state vector (loads not included) and the matrices cl 1, 131,12 and 71 are determined from Eq. 1. Assume linear LFC is used, i.e.

cept of the linear-plus-deadband control of Eq. 23. Even though the details underlying Eq. 21 and Eq. 23 are omitted, the application of control design using set-theoretic techniques is a major contribution of this paper. Details are presented in Ref. 4.

IV. SIMULATION
1. System Description
sources is investigated (Figure 3). There are four nodes (five including

u.(jc)= iX(jE)

1ZI(j ) + K2

AFLL

(17)

where K is determined from Eq. 15. Using linear control, the model, Eq. 16 becomes

A power system consisting of two interconnected areas with two

2098

Authorized licensed use limited to: Sven del Pino. Downloaded on September 15, 2009 at 16:44 from IEEE Xplore. Restrictions apply.

YA

SiS

iY D

YB

"area control error", and is dispatched every two seconds. The "frequency bias coefficient" is chosen to be equal to the "area frequency response characteristic" of the Northwest U.S., and an optimum gain is chosen to obtain "best" conventional LFC performance.

&

TABLE 11. Discrete Time Model of Area One


.
0 4

{PS2(t)

'
-

PL2(t)
Two Areas, Two Sources

Fig. 3. Simulation Configuration

ground). Power sources are connected to two nodes, and loads are connected to two nodes. The nodes are connected via four transmission lines. That portion above (below) the dotted line in Figure 3 is viewed as area one (two). The two areas are interconnected via two tie-lines.

.962 1.11 .4c) -.001 -005 - 09 .13 -.001 -*05 -, 53 -,15 -.003 81. 7 565. E83.2 .948 8-.2 614. i11.9 2.06 0 0 0 0 0 0 iI 0 0
-

0
A.
=

0, 0

0 0

0 0

0 .0005 0 .0001 o, ooo4 0 -9-07 1 -. 65h 0 0 - .4 0 0 0 0

-ig46

.005 .030 .030 .0001-030 .0 0 -.64 -59. -1.01-58. .104t.8. .o46 -.12 0018 -.35 -1.7 .243

.0005.019 .019

-52.
0

0 0

1.0
0

1.0

.-qq 1..99
A(jf-+ F-)
X

.627 55.6 -147.

.0
X91

0-1

TABLE 1. Simulation Constants

Governor Turbine Constants


motor gain Al=-A2= 1.20 servo KHI''=Kli2-- 1.50 lever gain TR1=TR2= 7.00 seconds dashpot timeconstant
-

a-x(je) +.A-U(ifi) xl x2 x3 x4 x 5 x6 x7 x8
governor area one
-

X1,X2,x3
APM si.
=

turbine variables of

D'Rl=DR2= .430 dashpot compensation DP1=DP2= .05 permanent speed droop PB1=PB2= 826 megawatts- maximum power T'Wl=TW2= .860 seconds- turbine time constant

x4

IpMsl PMslnoml =-71X1+4-3x2-4.25x3


Si sl variables of tie-line power model

X5
AP x8

Swiyig Equation Constants


Ml=M2= .0177 normalized moment of inertia Dl=DCO-= .000177 normali-zed damping constant
-

x6,x7
T

=,3.5
AP Ll '4P L2

x6 -.18 x8

-t-

.4 5 xg

Transmission Line Admittances YA=YC=

X9

YB=YD=3.82
=

7.65

per unit (80 miles) per unit (160 miles)

.Base .u-antities SB77: 700 megawatts system power base V 500 kilovolts system voltage base (line-line,RMS) y SBAV 2 ) z-.0028 rnhos
-

model of Atea one, including the tie-line power flow and load in Area two, is given in Table 2, along with a physical interpretation of the state variables.

3) Advanced LFC (Eqs. 13, 23) The, gain matrix, Ki, is designed as in Section 111.2. A discrete time

Machine constarits ai-e based ori measurements taken at the Dalles Dam ( Columbia river) and on erigineering judgment. Constants are chosen to represent ten 70-megawatt geDerators. Coristants labelled .1, (2) refer .to area 1 (2)"O
Each power source is considered to be a hyc'tro source, with mechanical dynamics as in Figure 1. The loads, transmission lines, and source electrical dynamics are modelled as in Section IL Constants are listed in Table 1. A "complete" model of the two sources, two loads, and four lines is used for simulations. Only an area model (plus a model of the load outside the area) is used to design the area LFC. The following control laws are simulated. 1) No LFC
u
U

TABLE III. Cost Function and Feedback Matrices

Q,

lqj.jl
7xlO 6,

;q44' =1.0
R
.=

q55

.10.1 qii
k;

otherwise

El-

-.00004- 0 0 =[-.068 --.310 -.090 .065 o4o -.ool463


cumulative inadvertant int.erchange feedback gain

(24)

2) Conventional LFC Conventional LFC in each area is proportional to the integral of the

The. cost function matrices g and R, and Ki (modified) are listed in Table 3. Q and R were adjusted in a "trial and error" technique by observing the system response. It was desired to minimize swings without a large incre'ase in the magnitude of control (the mechanical response of the, sources should stay in linear region).

2099

Authorized licensed use limited to: Sven del Pino. Downloaded on September 15, 2009 at 16:44 from IEEE Xplore. Restrictions apply.

TABLE IV

Deadband Terms

o101 41

0]

10
0 0
0 -487- 487-

-6~F
l.xlO
ap
Ll

78

1.41 o
CSS
-28.6

6.7x10-29
8x10

sources one and two is -.0640 radians. At time zero a step load change of .05 per unit power occurs in area one (APLl(t) = .05, t > 0). It is assumed that the discrete time control laws are zero for the first two seconds after the disturbance. The model (Eq. 10) is simulated on an IBM 360/65 digital computer, using the "Continuous System Modelling Program" (CSMP) subsystem [9]. A random number generator is used to model the load perturbations, wLl(je) and wL2(je). Two cases are considered, IwLi(je) < 10-4 and IwLiOe) < 10-3. Plots of frequency deviation, the envelope of tie-line power flow, and mechanical power are shown in Figures 4-8.
1.04
102 /

28.6 ILP

F(di agr) -s S

0.50 l.xlO0 .6j


l.

PMI.=PM.
2.0
40

TABLE V.

Discrete Time Model of Area One

1.01i
1.00 o
0.99

.962 1.11 .398 -.05 -.01 .125 -.05 -. 53 -.15 83.2 81.7
83.2
I

.1

0 0

.001 0 0 -.00061 _.001 0 0 - .0009 -.003 0 O -.0009 .948 0 .008 1.92 565. 614. 11.9 2.06 1 .oo8 1.93 0 0 0 .9998 1.967 0 0 0 0 -.0001 .967 0
1.991 1.99
-147.

6.0

8-0
T- SEC

ENVELOPE APTIE I

.00
-.Ol
-.-C3

20O

40

60

so
T-SEC

g2a =.627
10
-xI.(

=59.61

.049 .027 .0o49 .026 -95.5 -51.9 -95.7 -52.41 -.356 .898 .89a -. A15
)
=

.030

.0171

.6 -AFI-CPS

.4-

-2
-.4

e i- E

+II \(iF)

otl

xl(i) +.gsul(Je);LlJ

6L2F

-6

xll,x12, x13 - governor turbine vari ab:les -frequency deviation xi4 =AUt1 = 21Tf1&
x15

Fig. 4. Dynamic Response -No LFC

20
I

40
1

60

80
T-SEC

I^ -

=A s1
-

angle deviati-on

= xl 6 =

x17 =

(AP T +24o.xl 7 +. 18L1 -. L2,45P 5PL2 /120 OlPTcAt - cumulative inadverltant


; nt.rhnnA II 1 U Z.L- IL;Lltt 1 Ir, IZ

The deadband logic of Section 111.3 is used to determine css and A different model of the tie-line power flow, which includes the cumulative inadvertent interchange, is used, resulting in an area model shown in Table 5. Uncertainty is included only in the loads for this design and simulation. Also, rather than use the full deadband matrix, Pss only the diagonal elements of the matrix are used. This reduces the computational effort (the matrix inversion, I7l in Eq. 23, becomes trivial). The effect of this approximation is not analyzed. 4) Continuous Time Advanced LFC

'ss (Table 4).

FCPS I\

.4
02.0A _~
-.

.4

80

100

120

140

(t)]
u2

(t)

(t)

(25)

40 *4

-60-' T-SEC

The feedback gain matrix is designed via the continuous time state regulator problem [8]. It is assumed that before time zero the system is in steady state. Each source is generating 1.0 per unit power. The load in area one (two) is absorbing .864 (1.136) per unit power. The net tie-line power flowing from area one into area two is on schedule at .136 per unit. Also the frequency is sixty hertz (in both areas), and the angle difference between

Fig. 5. Dynamic Response - Conventional LFC


2. Discussion

For no LFC (Figure 4), frequency deviation in area one oscillates and, in steady state, is -.063 hertz. The frequency deviation in area

2100

Authorized licensed use limited to: Sven del Pino. Downloaded on September 15, 2009 at 16:44 from IEEE Xplore. Restrictions apply.

two is approximately the same as in area one. In steady state, the net tieline power flow is off schedule by -.025 per unit. The mechanical powers follow frequency and, in steady state, each source picks up onehalf of the unscheduled load change. For conventional LFC (Figure 5) the frequency oscillations are centered about sixty hertz and in steady state, frequency is returned to schedule. The oscillations are not reduced, compared to conventional LFC. In steady state, the net tie-line power flow is returned to schedule, and source one picks up all of the unscheduled load chalge.
1.05

schedule. Frequency oscillations are reduced by about ten to thirty percent after the first swing. After four and one half minutes, the state enters the deadband regions, and control action is reduced. As shown in Figure 7, control action is eliminated for wliw(je) < 10-4 and reduced by about fifty percent for IwLi(je) < 10-3.
1.05
PM2-*

PM1

-x

1.04
PM2PM1

*
x

1.04
1.03 1.02

1.03

1.02
1.01
20

1.01
20

-x 40

60

40

6o

8so

1.99 -

T- SEC

T-SEC

ENVELOPE &PTIEI

ENVELOPE &PTIEI
20
-

.00
40 60 80

20

40

60

80
i

-.01 -.02-

T-SEC

.01
T-SEC

- .02 - .03

.03

.6

FI CPS

.4o { \ ,:>> -20

40

60

so

-.2

T-SEC

-.4

T-SEC

-.6

1 -.8
-.81~

Fig. 8. Dynamic Response - Advanced LFC, Continuous Time


-

Fig. 6. Dynamic Response

Advanced LFC

For advanced LFC (Figure 6), the state is outside the deadband region during the first four and one half minutes, and linear control is applied. Control action in response to the disturbance is not delayed. In steady state, both frequency and net tie-line power flow are returned to

For continuous time advance LFC (Figure 8), frequency oscillations are reduced by about forty percent during the first swing and an order of magnitude thereafter. Also the period of the frequency oscillations is about one-half that of conventional LFC, and the sources respond about twice as fast as conventional LFC. Further reduction in frequency oscillations could be obtained (with a "larger" feedback gain matrix), but the sources would respond at a correspondingly faster rate. V. CONCLUSIONS

41

Log 10 Icli -cSSO2 rss

* - Iwi (je)l = X Iwi Cje)l=I

10-4
T- SEC

-3*50
-

Fig. 7. Deadband Logic

Based on the simulations presented in Section IV, the author makes the following conclusions. 1) A major reduction in swings can not be obtained with advanced LFC signals which are sent with a period of two (or more) seconds. Swings are reduced (by ten to thirty per cent after the first swing), but not substantially. 2) A major reduction in swings can be obtained by decreasing the advanced LFC period. This reduction is limited by the rate at which power sources can change their generation levels. 3) Advanced LFC is a promising approach to reducing the number of LFC signals which are sent to power sources. The control law is flexible, versatile, and easily implementable, provided the appropriate variables are monitored and are available for control use. Also, control action in response to large disturbances is not delayed by the deadband logic.
2101

Authorized licensed use limited to: Sven del Pino. Downloaded on September 15, 2009 at 16:44 from IEEE Xplore. Restrictions apply.

Advanced LFC laws are based on a dynamic model of an area of an interconnected power system. Consequently, advanced LFC performance is critically dependent on the validity and accuracy of the model. Not until adequate models of power system dynamics and an on-line dynamic state estimator are available, will advanced LFC be feasible. The linear-plus-deadband control presented here is only a special case of a more general theory. In Ref. 4 different but related designs that could be applied to LFC are discussed. The incorporation of a dynamic state estimator based on unknown but bounded uncertainty [10] is also discussed. The general concept of set theoretic control can also be applied to other power system problems, such as voltage regulation, load shedding, etc.

purposely selecting a control area on the basis of the electrical parameters of the system. Separate control areas which are very closely "coupled" by strong interconnections are becoming more frequent as individual companies undertake joint ventures in constructing new generating facilities. In the control model presented by the authors, is the '4area" considered to be comprised of closely coupled generation and loads and is the "area" assumed to be controlled by a single area controller? How would the author's model be affected if the separate control areas were closely coupled? Consistent with preceding discussion, it would appear that as interconnections between neighboring utilities are strengthened, the concept of a control area defined by company boundaries would invalidate the assumption that the power transferred to an area is consumed within the area. Have the authors investigated the effect of the approximate model for APT on the tie-line power flow and transient frequency oscillations? With regard to the conclusions of this paper, the discusser has two questions: ACKNOWLEDGEMENT 1. Would not a reduction of advanced LFC period increase the number of control signals sent to the power house? 2. Do the results expressed in this paper (Items 1 and 2) apply This work was supported by the Bonneville Power Administration, only to a system with hydro generation? Have the authors Portland, Oregon. The authors wish to thank Messrs. Arden Benson, conducted any studies with non-hydro sources? Arthur Brooks, and William Hauf for their criticism and support. The authors are again to be commended for their work and encouraged to pursue further the application of modern control theory to the control and operation of a power system. REFERENCES
C. E. Fosha, Jr. and 0. I. Elgerd, "The Megawatt-Frequency Control Problem: A New Approach via Optimal Control Theory", IEEE Trans. Pas 1, Vol. PAS 89, pp. 563-578, April, 1970. [2] R. K. Cavin, III, M. C. Budge, Jr. and P. Rasmussen, "An Optimal Linear Systems Approach to Load-Frequency Control" IEEE Trans. Winter Power Meeting, January, 1971. [3] C. W. Ross, "Error Adaptive Control Computer for Interconnected Power Systems", IEEE Trans. Pas, Vol. PAS 85, No. 7, pp. 742-749, July, 1966. [4] J. D. Glover and F. C. Schweppe, "Control of Linear Dynamic Systems with Set Constrained Disturbances", IEEE Trans. AC, Vol. AC-16, No. 5, October, 1971. [5] D. G. Ramey and J. W. Skoogland, "Detailed Hydro-Governor Representation for System Stability Studies", IEEE Trans. Pas, Vol. PAS 89, pp. 106-112, January, 1970. [6] R. E. Larson, "Survey of Dynamic Programming Computational Procedures", IEEE Trans. AC, Vol. AC-12, No. 6, pp. 767-774, December, 1967. [7] A. R. Benson, Control of Generation in the U. S. Columbia River System, Bonneville Power Administration, Portland, Oregon, 1966. [8] M. Athans and P. L. Falb, Optimal Control, McGraw Hill Book Company, New York, 1966. [91 IBM, Subsystem/360 Continuous System Modelling Program (360 A-CX-I6X) User's Manual, IBM Rept. H20-0367-2, New York, 1966. [10] F. C. Schweppe, "Recursive State Estimation: Unknown But Bounded Errors and System Inputs", IEEE Trans. AC, Vol. AC-1 3, pp. 22-28, February, 1968. [11] N. Cohn, Control of Generation and Power on Interconnected Systems, John Wiley and Sons, Inc., New York, 1966.
[1

Discussion
J. A. Hetrick (Ohio Edison Company, Akron, Ohio 44308): The authors are to be commended on their efforts to develop a more sophisticated control law for reducing power system transient frequency oscillations. The addition of a deadband control plus the investigation of "discrete time" load frequency control are contributions to previous studies on this subject. Unfortunately, anyone responsible for the operation of a power system must look into the future and ask: "Is this concept applicable to the real world?" "Will it achieve better control?'" and "How much will it cost?" Of course, the answer to these and similar questions must await the arrival of an actual, field tested, control system embodying the control theory presented in this article. The authors do, fortunately, indicate an awareness of the problems associated with adapting such a control scheme to the real world. Referring to the work reported in this paper, the discusser would like to comment on the term "area" as defined by the authors. Most present day control areas are defined by corporate system boundaries or by power pool agreements. For this reason, little thought is given to
Manuscript received February 9, 1972.

Charles W. Ross (Leeds & Northrup Company, North Wales, Pa. 19454): This discussion deals primarily with the itemized difference between the EACC and deadband logic given in the Introduction of the paper. The function of the EACC is to analyze disturbances (deterministic as well as probabilistic) in the system and to modify the controller and economic considerations based on the urgency with which control action is- required. The net effect of the disturbances in the system, including the dynamic response of the system to the disturbances, is reflected in the area control error ACE. The EACC takes advantage of the fact that even random (white noise) disturbances are colored (correlated) by the system. The effectiveness of the EACC results from utilization of the joint probability characteristics of the observed disturbances, i.e., analyze the steady -state value, the expected magnitudes, the expected times above the expected magnitudes, and the direction of the change in ACE. This is basicatly a pattern recognition approach which is applicable to .deterministic as well as probabilistic disturbances. The principle of joint probabilities is believed to give the fastest known method for analyzing disturbances and assigning a probability weighting to their importance from the control point of view. The EACC may also be made adaptive to deal with the various classes of disturbances which may be encountered in an electric power system. This is accomplished simply by retuning the EACC from fast responding disturbance state detectors as discussed in Reference (1). Perhaps the basic difference between EACC control (Reference (1)) and the "advanced control" with deadband discussed in this paper is that-the EACC deals basically with "load disturbance states"-the "advanced control" deals primarily with process states. If pure deadbands are assigned to prevent control of random process noise the steady state error may be substantial. If deadbands are assigned based on uncertainties in the model for estimates states, it is doubtful that these states would be active during control, i.e., inside of the generation rate limiting band for the dispatch units. Perhaps the authors could comment on the effects of these deadbands and generation rate limiting as well as how their optimum control law deals with these nonlinearities. It is not clear how the authors deal with the problem of steady state errors. Does the control scheme have reset action? If so, which state or states are reset? In conventional control only the area control error is driven to zero. If the authors are counting on the inadvertent interchange in the equation for K 1 (Table III) for reset action-this will result in unilateral interchange correction. Finally, can the reset and dead zones in the control scheme lead to reset-dead zone cycling? As the authors discuss, the process states are ill-defined and variable and the degree of controllability is limited by equipment constraints. Therefore, the question is raised-why attempt to use linear-optimal control laws over the limited and uncertain range?
REFERENCE
[1] C. W. Ross and T. A. Green, "Dynamic Performance Evaluation of Computer-Controlled Electric Power System," presented at IEEE Summer Power Meeting, Portland, Oregon, July, 1971. (71TP593)
Manuscript received February 17, 1972.

2102

Authorized licensed use limited to: Sven del Pino. Downloaded on September 15, 2009 at 16:44 from IEEE Xplore. Restrictions apply.

J. Duncan Glover and Fred C. Schweppe: We thank Messrs. Ross and Hetrick for their valuable discussions. Dr. Ross's comments are addressed first. The function of the deadband portion of advanced LFC is the same as that of EACC; i.e. to reduce control action without compromising other LFC objectives. It is uneconomical to continually change generation levels, and it may be futile to attempt to track rapidly varying loads and random disturbances. EACC [3] deals basically with the area control error (a weighted sum of frequency and net tie-line power fl6w deviations), which Dr. Ross refers to as "load disturbance states." Deadband logic deals with all available information, including frequency deviation, net tie-line power flow deviation, plant variables, and load variables. The basic concept of deadband logic is to determine a (multidimensional) region. If the control information (above variables) lies within the region, control action is eliminated. The construction of the region is based on a knowledge of the dynamics of the area, which includes generation response (rates, delays, limits), load, tie-line, and noise (randomness) characteristics. EACC uses a probabilistic approach towards the random disturbances. Deadband logic uses a magnitude bound approach; i.e. instead of considering averages or expected values, deadband logic considers bounds. Both approaches can take advantage of time correlation (coloring) by using "noise dynamics" (e.g. exponential filters; periodic filters, etc.). Since both cumulative inadvertent interchange and time deviation are considered in the deadband logic (they are two dimensions of the multidimensional region), steady state errors will not occur. In conventional LFC, the area control error is reset. Similarly in the linear portion of advanced LFC, frequency and net tie-line errors are reset. Proper adjustment of the reset gains results in mutual, rather than unilateral, interchange correction. Reset-deadband zones could lead to reset-deadband cycling. However, the deadband region (and the feedback gains) can be tuned or adjusted to prevent cycling. Tuning can also be used to adjust for nonlinearities (e.g. source limits and deadbands). Also advanced LFC could be made adaptive (this is not done here). The reason for considering advanced LFC, which uses more information and a knowledge of the dynanmics of the area, is simply to utilize a better control law which improves system performance and efficiency (see Introduction). Mr. Hetrick suggests the possibility of choosing a control area on the basis of the electrical parameters of the system. We feel that LFC should occur within the areas defined by corporate boundaries (or power pools) where it can be coordinated with other controls (e.g., economic dispatch) and planning decisions (which affect future LFC and the integrity of the area).

Manuscript received April 24, 1972.

The area model presented (Figure 2) represents Ns sources and NL loads, which may (or may not) be closely coupled. The degree of coupling depends on the specific characteristics of the area (transmission line strength, size and dynamics of sources and loads) and is specified in the area model by the values of the D matrix, and the source and load parameters (e.g. inertia constants, time constants). The area is assumed to be controlled by a single area controller. However, the area is affected by other areas (neighbors) of the system through its tie-lines (Figure 2). The degree of coupling of the area to its neighbors is represented in the model via (part of) the D matrix. It is not assumed that power transferred to an area is consumed within that area. On the other hand, LFC is (partly) designed to accomplish this. LFC causes each area to regulate its own load changes by maintaining net tie-line power flow on schedule (assuming that enough sources in each area are on automatic control and can respond fast enough to meet area load fluctuations). Scheduled changes in power transferred from one area. to another are accomplished via (mutual) changing of tie-line schedules. However, during transient periods, all areas in the system respond to disturbances. Following a disturbance (e.g. a load change) all areas first act to reduce frequency changes (generator shaft accelerations) via source governing action. Second, LFC causes all areas to aid in restoring system frequency to schedule. Third, as system frequency returns to schedule, LFC causes the area in which the disturbance occurred to correct its own load change (other areas revert to pre-disturbance conditions). As interconnection strength between neighboring utilities is strengthened, all areas contribute more to system frequency regulation. The approximate model of tie-line power flow is used only to design advanced LFC laws. For example, the two areas simulated here (Figure 3) are closely coupled (each area consists of one source and one load). The LFC law in each area is based on a model of the area source, load, and approximate tie-line model. The performance of the LFC laws (Figs. 4-8), however, is determined from "exact" simulations of both areas (no approximate tie-line model used). A reduction of the advanced LFC period will increase the number of signals sent to power houses. Although the deadband logic is designed to reduce control action, during a sustained load change more signals are sent with a smaller period. We are opposed to continuous time LFC for this (and other) reasons. The ideas of the paper (modelling and control) apply to a general power system. The conclusions are based only on simulations of a hydro system. We feel, however, that similar results would be obtained for typical existing systems. In a mixed-type system (e.g. AEP) often large fossil fuel units are base loaded and pumped storage hydro units are used for a major portion of regulation duties. Also the longer delays (time constants) inherent in larger fossil fuel units reduces their regulating effectiveness (especially with regard to "fringe" control). Thus the deadband logic is a promising approach to reducing control action and increasing the efficiency of (hydro, fossil fuel, nuclear, etc.) power plants and the system.

2103

Authorized licensed use limited to: Sven del Pino. Downloaded on September 15, 2009 at 16:44 from IEEE Xplore. Restrictions apply.

You might also like