Sie sind auf Seite 1von 11

1698 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, VOL. PAS-89, NO.

7, SEPTEMBER/OCTOBER 1970

STATE CALCULATION OF POWER SYSTEMS


FROM LINE FLOW MEASUREMENTS

J. F. Dopazo, 0. A. Klitin, G. W. Stagg, L. S. Van Slyck


American Electric Power Service Corporation
New York, N.Y.
ABSTRACT

This paper describes a new and efficient method for com- STATEMENT OF THE PROBLEM
puting a real-time load flow solution of a power system net-
work. The use of line flow measurements is proposed and it is Power System Variables
shown that this measurement scheme shows superiority in
speed, accuracy and numerical behavior together with advan- The quantities of interest for transmission network moni-
tages in error detection and identification. toring are:

INTRODUCTION I. At each network bus:


Voltage magnitude and angle.
With higher voltages and more interconnections of electric * Real and Reactive input/output power.
transmission systems, it has become necessary to develop Il. In each transmission line:
methods for real-time monitoring of the system variables. For- Real and Reactive power flow.
tunately, this requirement is simultaneous with the availability
of more powerful computers and computer systems and signifi- Some of these quantities must be measured; the remaining
cant advances in analytical techniques which are the tools ones can be calculated. Regardless of the number of real time
necessary to do the job. Various analytical techniques have been measurements, calculations are required for the following rea-
presented in recent papers.(l)(7) sons:

In general, these papers proposed using bus input/output . To determine bus voltage phase angles (which are con-
powers together with line power flows as basic data. The ana- sidered economically unfeasible to measure.)
lytical methods described to process the power system data * To reduce the cost of metering and communications
were: systems.
* To safeguard against a possible loss of metered data.
a. weighted least squares * To detect and identify bad data.
b. independent equations . To utilize the monitoring system in future applications
such as analysis of system contingencies.
Some disadvantages of these methods are:
Selection of Measured Variables
* Weighted least squares requires the proper and a limited
range of values for the weighting factors to obtain a A possible selection of measured quantities is the real and
solution. If considerable amounts of redundancy are not reactive power flows in all transmission lines. These measure-
incorporated in the algorithm, numerical difficulties are ments include the necessary redundancy for the solution of the
encountered in the processing of input/output powers and problem. Advantages of this measurement scheme over those
line flows. proposed previously(l)4(7) are:
* Independent equations requires a high level of accuracy in
the input data to insure detection of bad metered data.(1) a. All input data is compatible. Input/output powers mainly
determine voltages to ground while line flows specify bus
This paper proposes an analytical technique to pro- voltage differences. The concurrent processing of both of
cess power system data without any of the disadvantages these data types requires proper relative weighting of these
previously mentioned. These problems have been circumvented quantities to obtain the solution.
by: b. A set of compatible independent equations is almost
always available.
* Defining a redunant set of compatible data. c. The results of calculations are less sensitive to measure-
. Processing of the data using a very simple and stable ment errors.
algorithm. d. Measurements at both ends of lines are very effective in
detecting and identifying incorrect data.
e. The numerical computation is fast and stable.
Paper 70 TP 146-PWR, recommended by the Power System
Engineering Committee of the IEEE Power Group for presenta- Figure 1 shows an EHV transmission network with 42
tion at the IEEE Winter Power Meeting, New York, N. Y.,
January 25-30, 1970. Manuscript submitted September 17, busses, 49 lines and 6 transformers. This network is based on
1969; made available for printing November 13, 1969. the AEP 765/345 KV transmission system. A line measuring
DOPAZO et al.: STATE CALCULATION OF POWER SYSTEMS 1699

scheme would total 208 measurements, consisting of watt and * Analog to digital converters
var meters at each end of every line and on the low voltage side * Transmission (of analog data)
of each transformer bank. Since at most busses the injections
are composites of line flows and transformer flows thru one or Analog transmission errors of 1% of full scale range (2% of the
more banks, the number of measurements required by any absolute full scale) are likely to be expected. With identical full
other scheme would be almost the same. scales of say 2000 MVA, meters for a particular metering point,
employing analog transmission, are likely to have error bounds
For instance, assumming that the input at each node is com- more than 40MW or Mvar.
posed, on the average, of 1 1/2 elements to be metered, and a
tree of the network is measured at only one end, the number of Fortunately, a digital data gathering station turns out to be
measurements required is 210. more economical than its analog counterpart if several measure-
ments are required at each substation.
However, this bus-line scheme would not be as reliable as
measuring all the line flows of the network. Once a digital installation has been justified, the errors are
reduced to:
Real Time Data Accuracy
* Analog to digital conversion
The components of the data acquisition system that in- (0.10% full scale).
fluence the accuracy of power system measurements are: * Watt and var transducers
(0.50-0.25% full scale).
Instrument transformers (CT's and PT's) * Instrument transformers
Watt and var transducers (2% of MVA measured).

BUSSES
.1 1 1
765KV 345KV

Figure 1. This figure shows the 42 bus system which was used in testing the algorithm.
1700 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, SEPTEMBER/OCTOBER 1970

The error bounds specified above, together with a random E2 = Vector of unknown complex bus
number generator to allocate errors to the measured quantities, voltages, N entries.
were used in testing this method. thus, _- AE =
(3)
Pseudomeasurements letting,
u = -A1E
An apparent disadvantage of the method described in this
paper is that pseudomeasurements cannot be used together with B = A2
the real time data as proposed in recent reports(2)-(7). These
publications advocated the use of pseudomeasurements together we have:
with real-time measurements to reduce the metering and com- v + u = BE2 (4)
munications costs. It must be emphasized that "good" bus
pseudomeasurements can be costly. To develop good pseudo- If measurements of the complex voltages across the net-
measurement models may require installation of off-line data work elements are available, the relationship between measured
recorders and recurring manual data processing to keep these Iand calculated values is given by:
load models up to date.
Vm + u = BE2 + r
(5)
If pseudomeasurement accuracies are not compatible with where,
real-time data accuracies "interaction," of these quantities may Vm = Vector of measured complex voltages
cause degradation of the state estimate, unless a large weighting across the network elements
factor is assigned to the pseudomeasurement to weight it "out" r = Vector of residuals.
of the weighted least squares estimate. If the pseudomeasured
quantity does not "interact" with teal-time data, then it will Minimizing rT; to find the least squares solution of the over-
not degrade the state estimate regardless of the weighting factor determined matrix equation (5) yields: (9)O(10) (See Appendix)
used. Thus, in the opinion of the authors, the doubtful eco-
nomic and computational advantages of pseudomeasurements
E G (vm ) (6)
do not merit an algorithm which is able to process this type of
data. where,
G = (BTB)YlBT
MATHEMATICAL FORMULATION and 'T' denotes matrix transposition.

A relationship between network element voltages and bus The vector of 'measured' voltages across the network ele-
voitages can be written as. ments can be approximated directly from measurements of the
real and reactive line flows. Assuming a voltage at bus p, the
N+K measured' voltage across line j connected at bus p is given by:
Vi - A2 E; (See Appendix)
i=1,2...M (1)
3=1
Vpm -J ]
(7)
where,
vi = ith complex element voltage
Ej = jth complex bus voltage with where,
respect to ground. zj = Complex series impedance of line j
Aij = Incidence of measured element i Pjp,Qjp = Measured watt and var flow on
on node j, (0,+l or -). line j at bus p
N = Number of unknown bus voltages. yjp = Complex shunt admittance of line
K = Number of known complex bus j at bus p
voltages, (K > 1).
M = Number of line (watt and var) and (*) denotes complex conjugate.
measurements.
It is interesting to note that:
Equation (1) can be rewritten in matrix form and parti-
tioned as follows: . The gain matrix G is a constant real matrix that depends
on network configuration and the lines which are mea-
v A1'1 + 2!2 (2) sured. It might be advantageous in an application to
exploit the fact that the matrix B is sparse.
Where, * The vector Gil is a constant complex voltage vector. When
El = Vector of known complex bus the complex voltage is known at only one bus, all the
voltages, K entries. elements of this vector are equal to the voltage at this bus.
This vector is also useful if several complex voltages have
DOPAZO et al.: STATE CALCULATION OF POWER SYSTEMS 1701

been previously determined from the voltage calculation of errors. However, detection and identification logic must be
an adjoining network. incorporated in the algorithm to safeguard against metering
malfunctions and rare communication errors.
ITERATION SCHEME
If no apparent data acquisition failures are present, the
Equations (6) and (7) are solved iteratively as follows; error analysis for the input data is based on the 'goodness' of fit
between measured and calculated line flows. The identification
(1) Set iteration count k= 1. logic begins with the removal of the data associated with the
(2) Estimate initial bus voltages, EiO, i= 1,2,...,N largest mismatch between a measured and calculated quantity.
(3) Calculate element voltages vmjk from equation (7), This identification scheme is based on the hypothesis that if the
j=1,2,...,M data point with the largest mismatch was indeed the bad one,
(4) Calculate bus voltages Eik from equation (6), the 'goodness' of fit would be improved. Conversely, the 'good-
i=l,2,..,N ness of fit would further deteriorate if the largest mismatch did
(5) Test for convergence IEik - Eik-l ( Tolerance not correspond to the bad data point.
i=1,2,...,N
(6) If step (5) is not satisfied, advance iteration count k State estimation errors should also be included in the
and return to step (3) algorithm to evaluate the sensitivity of the state calculation to
small measurement errors. Since the standard deviations of the
COMMUNICATION AND METERING OUTAGES errors in the state would be due mainly to current network
confilguration and the 'information content' of available data, a
Since loss of data has to be accounted for in the formu- method of obtaining these error sensitivities would be to
lation, an attractive feature of the proposed measurement perform a real-time Monte Carlo trial for present data acquisi-
scheme is that: tion and network configuration. Based on these results, data
communication outages could be evaluated without the need of
Modification of the matrix G is seldom necessary when constructing the approximate covariance matrix of the errors or
communication or metering outages occur. carrying it through the calculations as proposed in references
(2) thru (6). Thus, when undetected bad data is suspected
If both ends of all lines are measured, equation (4) can be because of one or more unusually large mismatches, a diagnostic
rewritten in partitioned form as in equation (8). To simplify subroutine could be automatically called and the bad data
notation, let: identified. This technique could also be applied to the approxi-
mate standard deviations of the line flows.
then,
e NUMERICAL EXAMPLES

Measurement Error Simulation


where,
ei,ei' = Modified complex element voltages Real-time data measurements were simulated based on the
determined from line i measurements following error bounds for each metering point:
at busses p and q respectively.
2% of MVA flow, plus
The solution of equation (8) is given by: 0.3 5% of Full Scale.
E
E2 1(BTB)lBT(-
2BB l

Full scale for all measurements was taken to be 2000MVA. The


which can be rewritten as: simulated measurement error was determined by the product of
E lG(e e'l) the error bound for each point and a random number within
2 (9) limits of ± 1.

If only one end of the lines, is measured, equation (9) The 42 bus test system shown in Figure 1 was used in
becomes, testing this mnethod. To test the sensitivity of this method to
=G measurement error, fifty cases were run for each numerical
example with different random numbers. The standard devia-
If both ends of all lines are measured, and data at one end of tion for errors in voltages and angles was calculated for each
line i is lost, setting ei = -ei' is all that is necessary. Modification bus.
of the matrix G is required only when data at both ends of a
line are lost or when the number of lines with data lost at one Numerical Results
end is excessive.
CASE A: Watt and var flows at both ends of all lines and
Digital communication hardware provides excellent error the low voltage sides of all transformers are available as
checking codes that detect and identify data transmission measured data.
1702 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, SEPTEMBER/OCTOBER 1970

CASE B: Same as Case A without measurements 19-21,


19-20, 20-22, 10-6, 8-3, 37-34, 29-28 and 30-31. This case
simulates realistic operation of the data acquisition system.

CASE C: Same as Case B with the additional outages of


the data acquisition stations at buses 3 and 24. This simulates a
severe double contingency outage of the data acquisition
system. 10
0)
-C)
The results of these cases are summarized in Table I. Thus far '.4
there has been no evidence of error amplification effects in the 0)
calculated non-measured quantities.
0
.5
-rf-
Convergence and Timing Characteristics
.1-
The convergence characteristics of the method are excel-
lent. A typical convergence characteristic is shown in Figure 2.
Five iterations are usually sufficient for convergence from a flat N
start in bus voltages.
0
For the real-time application (tracking mode with moder-
ate data sampling requirements), a state calculation time is 4--4
estimated to take 0.5 sec on an IBM 360/50. This calculated 0
timing requirement is approximately five times faster than the C
time required to perform one iteration of a Newton - Raphson 0
load flow in a sophisticated production program. This timing
reduction is mainly attributable to the selection of measured o .
variables which are unfeasible to forcast in system planning
applications.
CONCLUSIONS
The proposed method for the solution of the real-time 2 3 1s 5
load flow problem utilizes a very simple algorithm to obtain the Iteration (k)
state vector of the system. Acceptable results were obtained Figure 2. Typical convergence characteristics of algorithm.
from simulation studies in which realistic accuracies as well as Busses (16) and (1 7) had the slowest convergence.
typical outages were assigned to the various components of the
data gathering system. The proposed measurement scheme vergence characteristics. No error amplification was experi-
yields a set of compatible data which satisfies economic con- enced, that is, differences between measured and calculated
siderations as well as the requirements to obtain an acceptable values for both measured and non-measured line flows were
solution in the event of outages of the metering system. The found to be within the error bounds used for the measured
computational process is very stable and has excellent con- quantities.

kBLE I

RESULTS FROM MONTE( CARLO TRIALS


Maximum Standard Average Standard
Deviation of Errors Deviation of Errors
Voltage Angle Voltage Angle
(Percent) (Degrees) (Percent) (Degrees)
Case A .28 .18 .18 .09
Case B .15 .09
Case C .38 .20 .26 .11
DOPAZO et al.: STATE CALCULATION OF POWER SYSTEMS 1703

ACKNOWLEDGEMENT The solution of the overdetermined matrix equation (2A) is


obtained by minimizing the length of the residual vector F.
The authors wish to acknowledge J. Provanzana, B. Renz Since the real and imaginary parts of r are uncoupled, the
and D. Weinreich of the AEP metering section for providing squared length of this vector is:
information with regard to the accuracy and relative costs of
the various components of digital and analog metering systems. Jr 1r21== (-mvm -- c)Tc~
vc) (-Vm - -vc)
The condition for the minimum of Irn2 is:
REFERENCES r-vT 6vCT
.i ox )vm
*.
- c) =°(3A)
1. G. W. Stagg, J. F. Dopazo, 0. A. Klitin and L. S. Van From equation (lA), F_)T
Slyck. "Techniques for the Real-Time Monitoring of Power c
BT (4A)
System Operations" Presented at IEEE Summer Power ax
Meeting, Dallas, Texas, June 27, 1969 (69TP657 - PWR) substituting equations (1A) and (2A) into equation (3A) and
rearranging terms,
2. F. C. Schweppe and J. Wildes. "Power System Static State
Estimation: Part I - Exact Model" Presented at PICA con- B (Vm Bx + u) = n (5A)
ference, Denver, Colorado, June 18, 1969. (69 C2-PWR) where - v T
n = im, -VC)
3. F. C. Schweppe and D. Rom, "-: Part II - Approximate Rearranging equation (5A):
Model" Presented at PICA conference, Denver, Colorado, x = (BTB) 1[BT(vm+ ii) nJ (6A)
June 18, 1969. (69 C2-PWR)
During the iteration process- is taken equal to the zero vector.
4. F C. Schweppe "-: Part III - Implementation Presented at PICA This assumption accelerates the numerical computation to ar-
conference, Denver, Colorado, June 18, 1969. (69 C2-PWR) rive at an approximate best fit solution.

5. R. E. Larson, W. R. Tinney and J. Peschon, "State Estima- Numerical results have shown that for practical purposes
tion in Power Systems - Part I: Theory and Feasibility." the approximate best fit solution is acceptable for the network
Presented at the IEEE Summer Power Meeting, Dallas, shown in Figure 1. Differences between true and calculated
Texas June 18, 1969. (69TP638 - PWR) values using this approximation are on the order of tenths of
one percent for the voltage magnitude and tenths of one degree
6. R. E. Larson, W. F. Tinney, L. P. Hajdu and D. S. Piercy, for the voltage angles. On the other hand, the approximate
"State Estimation in Power Systems- Part II: Implementa- solution deviates from the best fit solution by hundredths of
tion and Applications." Presented at the IEEE Summer one percent for magnitudes and hundredths of one degree for
Power Meeting, Dallas, Texas, June 27, 1969. (69TP639 - voltage angles. However, if the best fit solution is desired, it can
PWR) be obtained after arriving at the approximate solution. This
only requires the evaluation of the vector T once' or twice near
7. Otto J. M. Smith, "Power System State Estimation" Pre- the solution point.
sented at Summer Power Meeting, Dallas, Texas, June 27,
1969. (69TP640 - PWR) To evaluate the vector fi, the computation of the partial
derivatives of the 'measured' element voltages with respect to
8. G. W. Stagg and A. H. El-Abiad "Computer Methods in the state variables is required. Once these partial derivatives are
Power System Analysis." New York, McGraw - Hill, 1968. obtained, the relationship between the real and imaginary com-
ponents of the vector n and the real and imaginary components
9. B. Noble, "Applied Linear Algebra." Englewood Cliffs, of the residual vector r is given by:
New Jersey, Prentice Hall 1969.
bv m r\T
i T
.-vm
-

10. Paul M. DeRusso, Rob J. Roy and C. M. Close, "State -r


n
-r
r
Variables for Engineers," New York, John Wiley and Sons, 6x r bxr
1966. (7A)
APPENDIX -i
n
\ y, i
Im ( iJ T
-1
r
To simplify notation let x = E2. From equation (5):
vm + u = Bx + r This procedure has the advantage that there is no cross-
designating coupling between real and imaginary quantities during the itera-
Bx - U =
Vc (1A) tive process. The partial derivatives in equation (7A) can be
then, obtained from equation (7) by taking partial derivatives of real
Vm -
vc = r (2A) and imaginary components of 'measured' element voltages with
respect to the real and imaginary components of bus voltages.
1704 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, SEPTEMBER/OCTOBER 1970

Discussion A. S. Debs and L. P. Hajdu (Systems Control, Inc., Palo Alto,


Calif.): The authors are to be complimented for developing a new
and computationally efficient technique to estimate the complex bus
D. K. Subramanian (Indian Institute of Science, Bangalore-12. voltages from line flow measurements. Their approach is not very
India): The authors have obtained a new model for load flows and dissimilar from that of static state estimation; it uses redundant in-
have shown that this model is more effective for on-line computation formation and implicitly considers the measurement process statistics.
of load flows than the usual method of calculating voltages using bus In the introductory part of the paper, certain disadvantages are
powers. The model is very simple to construct and as the coefficients of attributed to static state-estimation techniques in general; among
B are either 0 or ± 1, it is easy to perform multiplications. The
authors are to be congratulated. The discusser would like to know
others, reference is made to the algorithm of Larson et al. [6]. While
the computational difficulties cited are true in general for the textbook
the authors' opinions on the following points. estimation equations, the particular algorithm described in [6] con-
1) Although B is a sparse matrix, will (BTB) be sparse? If it is tains a number of computational simplifications with which the au-
sparse, (6) can be solved by the LU decomposition method instead thors may not have had direct experience. Using the static estimator
of direct inversion. algorithm exactly as described in [6], none of the claimed disad-
2) Is there any need for employing some kind of ordering in vantages exists. Specifically,
the numbering of buses and lines, like giving the number of the bus to 1) no numerical difficulties are encountered, regardless of the size
one of the lines connected to it, in order to make BTB nonsingular? of the problem or the amount of redundancy in the input data;
3) In the iteration scheme, what will be the effect of the following 2) the equations are always compatible, and the numerical com-
modification on convergence? putation is very fast and stable;
Instead of solving for fim and E2 separately from (6) and (7), sub- 3) pseudomeasurements do not have to be used, unless they are
stitute forvm in (6) from (7). Then (6) becomes necessary to obtain a solution.
While the authors' approach apparently works well on a specific
E = Gu + Gf(E2) (10) measurement system, the static state estimator of [6] handles any
measurement configuration. Since the cost of processing additional
wheref(E2) is a vector of order (N X 1). It is a function of voltage available measurements (e.g., bus voltages) is negligible, we would
variables E2. like to retain this flexibility in a real-time data-processing scheme.
Rewrite (10) as Our approach to the on-line power-flow problem is described in a re-
(BTB)E2 = BTfl + BTJ(2). (11) cent paper [14]; this tracking algorithm will be capable of processing
all types of measurements, with any specified sampling rate(s), and
Solve (11) by Newton's method for S2. its computation and storage requirements are much less than those of
This procedure replaces steps (3) and (4) of the iteration scheme the very efficient static estimator algorithm of [6].
mentioned in the paper. The advantage is the elimination of vm from REFERENCES
the iteration scheme.
4) What is the effect on convergence of incorporatingni (5A) in [14] A. S. Debs and R. E. Larson, "A dynamic estimator for tracking
the iteration scheme? the state of a power system." IEEE Trans. Power Apparatus
The discusser thanks Mr. G. W. Stagg for making available a copy and Systems, this issue, pp. 1670-1678.
of his paper.
Manuscript received February 13, 1970.
Manuscript received February 16, 1970.
G. I. Stillman (American Electric Power Service Corp., New York,
N. Y. 10004): The paper couples the use of load-flow procedures with
A. H. Jazwinski (Analytical Mechanics Associates, Inc., Seabrook
least-squared estimation techniques to arrive at a method for on-line
calculation of the state of a system during steady-state conditions.
Md. 20801): The authors have presented an interesting application The fact that well-known load-flow procedures lie at the heart of the
of least squares state estimation to power systems. It is an interesting method proposed is a strong point in its favor. In my opinion, it
combination of network analysis and least squares, resulting in a fast is highly desirable to extend the use of widely known and understood
and stable computational algorithm. techniques to new problem areas as long as these techniques remain
In choosing their present approach, the authors state that weighted valid and are efficient; among other things, familiarity with the solu-
least squares is limited by 1) limited range of values for the weighting tion techniques gives users of the solution output confidence in the
factors, and 2) requirement of considerable redundancy. I would like solution and its validity, and this is not a trivial point, particularly
to point out that the difficulties which the authors encountered with in an area like on-line monitoring.
weighted least squares were largely due to the way the authors Solution of the load-flow model proposed in this paper is much
formulated the least squares iteration. It is possible, by a trivial faster than in off-line production programs. This must be somehow
modification to their iteration, to enhance and even guarantee con- related to the fact that real and reactive powers are taken from
vergence in weighted least squares. This is done by merely adding to meter measurements, and it would be helpful if the authors clarified
the diagonal elements of the information matrix. This conditions the just what this relation is.
matrix and effectively places a constraint on the iteration step. With Redundant measurements, particularly useful should there be
this modification, the iteration converges to the same answer as the bad data or a loss of metered data due to malfunctioning of instru-
unmodified iteration. Insofar as redundancy is concerned, the authors' ment transformers or analog-to-digital converters or communication
present approach clearly requires more redundancy than weighted links, are obtained by measuring real and reactive power at both ends
least squares. Incidentally, the excellent convergence of the authors' of each transmission line. In addition, there are a number of pseudo-
technique with a constant gain matrix G implies that weighted least measurements (not explicitly labeled as such by the authors) in the
squares will also converge with a constant gain. The authors' analysis El
form of a set of known compex bus voltages, in (2). Presumably,
gives insight in choosing that gain. these bus voltages are known at generator buses from a predetermined
I should also like to point out that the authors' technique is limited voltage schedule, although generally, and I would appreciate the
to a very special and redundant telemetering configuration. In their authors' comment on this point, only the magnitude of such voltages
paper the authors are concerned only with the 765/345-kV AEP are known and not their complex components.
transmission system. How about the lower voltage network? Clearly What is left to calculate then is the set of unknown bus voltages
the voltages on the system are correlated. As a consequence, using and angles and this calculation is done via the algorithm consisting
weighted least squares, one can throw all the data in the pot and of (6) and (7). Use of (7) involves introduction of additional pseudo-
estimate the state of the whole bulk power system. The abundant data data, in the form of zj and yjp, the series impedance and shunt ad-
from the 765/34AkY system would enhance the estimation on the mittance of the line. These values are always calculated from theo-
lower voltage system. retical considerations and, to my knowledge, almost never measured.

Manuscript received February 9, 1970. Manuscript received F'ebruary 9, 1970.


DOPAZO et al.: STATE CALCULATION OF POWER SYSTEMS 1705

In their evaluation of the accuracy of the monitoring scheme pre- appears to this discusser that the measurement and telemetering of
sented, the authors use the same simulated network for comparing bus voltage magnitudes may also be justified. In contrast to voltage
measured with calculated values. This simulated network contains phase angles, voltage magnitudes can be measuLred easily and
identical zj and yjp for the case involving measurements and the case economically. (Three phase unbalances could be averaged.)
involving calculations.' Consequently, differences between z and y Some benefits that might be achieved by additional voltage
in the measurement network and the calculation network do not exist. information are summarized as follows.
But in real life there will be differences between the z and y of the 1) The dimension of the problem is greatly reduced since the com-
real system on which measurements are made and the mathematical plex voltage across each network element can be replaced by a mea-
model on which the calculations are based. Some of these differences sured bus voltage angle across the network element.
may be small and some may be large, particularly if there are errors 2) The incidence matrix A becomes much smaller but retains the
in the line data for the model. Would the authors care to discuss these same properties as (1). vj is replaced by the measured angle across the
differences and their effects on solution accuracy? network element, and Ej is replaced by the corresponding bus voltage
The authors eschew the use of pseudomeasurements in favor of real phase angle (with respect to any selected reference bus).
measurements. This is fine when real measurements can be made, 3) Iteration is not required since the measured angle across a
but what does one do in the case of foreign systems? From the view- network element can be computed directly from measurements of real
point of the EHV network, foreign systems fall into two categories: and reactive line flows and bus voltage magnitudes. The following
1) networks of interconnected companies whose lines are inaccessible; equation could be chosen in place of (7):
2) lower-voltage networks of one's own system which are accessible
but whose lines are so numerous that measurement of powers at both
ends of each line becomes inordinately expensive. Is there a way to elom.i= lEvl (1 + zj*yjp*) -z*(PyP + jQjp) (12)
cope with this problem other than by the use of pseudomeasurements where 0m.j is the measured angle across the series impedance of line
and pseudodata? j and IEP[ and JE11 are the measured voltage magnitudes at each end
Another question relates to gain matrix G, whose derivation ulti- of the line.
mately depends on incidence matrix A. The authors make it clear Aside from these suggested possibilities, the authors' valuable com-
in the paragraph following (9), that when data at both ends of a line ments are desired concerning the detection of errors in the basic net-
are lost, G must be modified. Loss of data at both ends of a line can work data (zj and yjp).
result not only from communication system failure but also, of course,
from an outage of the line itself. It would seem to be necessary to
distinguish between these two causes of loss of data. How is this done?
Are breaker positions to be monitored? Also, is the gain matrix auto-
matically modified or is the intercession of an operator necessary? Fred C. Schweppe (Massachusetts Institute of Technology, Cam-
The measurement of power at the ends of each line means that there bridge, Mass.): This verv interesting paper adds a potentially im-
will have to be a digital-gathering station at each substation and a portant new approach to the field of state estimation for power sys-
communications link from each station to the central computer loca- tems. I would like to provide some additional comments on the com-
tion. From the standpoint of economy, it would be desirable to do with parison of the authors' state estimator to the weighted least squares
less than this. Case C, in the Section "Numerical Examples", allows state estimator.
the loss of data-gathering facilities at two buses above the loss of mea- Weighted least squares has a very simple physical interpretation in
surements of 8 lines. The loss of accuracy in voltage measurement that it provides the state estimate which minimizes the weighted sum
(maximum standard deviation of errors) is only 1/10 of 1 percent of the squared residuals (differences between the actual observations
compared to the case when all data are available. This seems to in- and the observations predicted by the state estimate through the
dicate that a number of data-gathering stations could have been mathematical model). The weights are determined by the accuracy
omitted from the original scheme while still obtaining answers of of the observations. The authors' approach also minimizes a sum of
sufficient accuracy. Are the authors considering this possibi]ity? certain squared residuals (and hence could also be called at least
The residual vector r is actually a sump for the measurement square estimate if desired). However, the physical interpretation
errors contained in the monitored power flows at the end of each line, of these residuals and hence the resulting state estimate is not obvious
for the errors that exist in calculated values of Zj and yjp as used in (to me). A physical interpretation of a mathematical technique is not
the model compared to the line parameters of the real system, for absolutely essential, but it is desirable. I hope that the authors will
calculation errors in measured element voltage P., and for errors in provide further discussion on the physical interpretation of their
the known voltage set £1. Generally, such a residual or error vector, technique.
and in particular its variance, tells us something about the difference Weighted least squares is definitely a more general approach in
between the estimate of a state and the true value of the state. This that pseudomeasurements and bus injection measurements can be
point is not discussed explicitlyr in the paper, and I suspect that in the combined with line flow measurements in any desired combination
lack of a need for such a discussion lie some of the advantages of (provided the overall metering configuration is reasonable). Such
the method proposed in here as compared to the methods proposed versatility can be important in practice. Weighted least squares can,
in [2]-[6]. Would the authors kindly amplify this aspect of the work? of course, handle a metering configuration of only line flow measure-
Finally, the authors indicate that a state calculation can be made ments and voltage magnitudes at buses.
in 0.5 second on an IBM Model 50. Presumably, with a more powerful Weighted least squares is also more general in the sense that it can
computer, calculation time could be reduced further. Suppose, for be applied to a formulation of the state vector in terms of either volt-
example, that calculation time were reduced to 0.2 second and that age real and imaginary parts or voltage magnitude and angle (or any
the time required for sampling, gathering and transmittal of data to other formulation). The authors' techniques are based directly on a
the master computer were approximately 1/2 to 3/4 second (a not real-imaginary formulation. To my knowledge, all weighted least
unreasonable number for the latest available equipment), then a square studies done so far have actually used a magnitude angle
complete updated calculation of the system state could be made formulation because it is somewhat easier to interpret physically and
every second. This would seem to indicate that with presently avail- because of the partial decoupling between watt-angle and var-
able technology, the method presented by the authors could be useful magnitude relationships. Do the authors feel that a real-imagin-
for tracking the dynamic performance of a system for slow oscillations ary formulation would be helpful in a weighted least squares ap-
such as occur between large systems, i.e., oscillations with periods of proach?
several seconds per cycle. Have the authors also been thinking along The performance of both weighted least squares and the authors'
these lines? technique can be evaluated by Monte Carlo techniques, but such
Many thanks to the authors for a well-written enjoyable paper. procedures require expenditures of large amounts of computer time.
Weighted least squares has the advantage of providing a simple ana-
lytical formulation for an error analysis, and such equations can be
very useful both in the design of the metering system and in actual
Walter 0. Stadlin (Leeds and Northrup Company, North Wales, operation.
Pa.): On the assumption that a sufficient number of communication Weighted least squares has the property of providing estimates
channels have been installed at each substation to satisfy case A, it which are optimum in a certain well-defined and physically inter-

Manuscript received February 5, 1970. Manuscript received February 12, 1970.


1706 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, SEPTEMBER/OCTOBER 1970

pretable statistical sense. In this sense, the authors' technique can tem, is solvable only through a large investment in real time measure-
never provide better answers, although they may be just as good ments or else through the use of very accurate pseudomeasure-
from an applications point of view. It would be very helpful if the ments.
authors compared their accuracies relative to weighted least squares We agree with the discusser that the well-known technique of
accuracy. damped least squares solves the problem. However, the implementa-
The detection and identification of bad data points and modeling tion of this technique requires additional computer logic to achieve a
errors are a very important aspect of state estimation and can be reasonable convergence to the solution in the real-time environment,
done with both techniques. Weighted least squares has the advantage i.e., changes in network configuration, loss of information, etc.
that the necessary algorithms are well understood and have been With regard to the amount of redundancy required, please refer to
proven successful in other applications. They have also been shown to our comments to Dr. Schweppe's discussion. Weighted least squares
be successful on power system simulations [111. works well with a constant gain matrix as long as the operating point is
All of the preceding discussions indicated advantages of the not far from the point where the computation of the matrix took place.
weighted least squares technique. However, a very important prop- However, in the real-time environment one must be prepared to take
erty is the amount of computer time and storage required for imple- this possibility into account. Again, this introduces additional over-
mentation. Here a comparison is very difficult. A tracking state head into the calculation of the estimate which is not required with
estimator, as proposed in [12], can be implemented with very limited our algorithm.
computer storage and computation requirements and has been made We do not agree with Mr. Jazwinski that our technique is limited
to work on simulated data [13]. On the other hand, the authors' ap- to a special metering configuration. If one were to agree that real-
proach also requires a very limited amount of computer time and time information is required to solve the estimation problem in
storage and has been shown to work. the power system, then one has to go to the network nodes to get
My personal opinion on the relative advantages of the two ap- either injections, voltage magnitudes or line flows. Conceivably one
proaches can be summarized as follows. A weighted least square state does not have to take measurements at every node of a power system
estimator is definitely preferable if the computation and storage re- (see numerical examples). However, it is good policy for reliability
quirements are less than or approximately equal to those of the and to avoid computational difficulties, as we have found from our
authors' state estimator. However, if weighted least squares turn out studies, to measure line flows at both ends; but this is not a stringent
to require appreciably more computer load, the choice depends on an requirement either. It also proves to be a more economical measure-
engineering trade-off of weighted least squares advantages versus ment scheme compared to bus injections and some line flow measure-
computer requirements. More experience is needed before the actual ments.
computer requirements (including detection, identification, lost data As far as the lower voltage network is concerned, our method is not
points, etc.) are known. limited to handle only the 345/765-kY system. If the appropriate
data for the low voltage sysuem is available, it can also be included in
REFERENCES our "pot."
[11] E. Hnyilicza, "A power system state estimation with applica- Mr. Debs and Mr. Hajdu are quite right when they say that
tions, " School of Engrg., Massachusetts Institute of Technology, the statistics of the measurement process are implicitly included in the
Cambridge, PSEG Rept. 12, April 1969. model. As a matter of fact, it can be shown analytically that
[12] F. C. Schweppe, "Power system static-state estimation, pt. III: the method presented in our paper is a very good approximation
implementation," IEEE Trans. Power Apparatus and Systems, to the static state estimation algorithm of [2].
vol. PAS-89, pp. 130-135, January 1970. The disadvantages mentioned in the introduction of our paper
[13] R. Masiello and F. Schweppe, "A tracking static-state estima- were of a practical nature regarding the application of state estima-
mation," Paper 70 TP 707, presented at the IEEE Summer
Power Meeting and EHV Conference, July 12-17, 1970, Los tion to power systems. If economics are considered, the cost of mea-
Angeles, Calif. suring bus inputs and some line flows in the system are seen to be as
expensive as measuring all line flows at both ends; this fact is implicit
in the topology of the network. With few line flows for redundancy,
loss of bus injections leads to numerical difficulties regardless of the
method used to process the data. The computational simplifications
mentioned by the discussers are reasonable and have been under our
consideration for quite some time. However, the diagonalization of
J. F. Dopazo, 0. A. Klitin, G. W. Stagg, and L. S. VanSlyck: The the inverse-covariance matrix, though obviously less storage demand-
authors would like to thank all the discussers for their complimentary ing than calculating a full covariance matrix, introduces several dis-
remarks, penetrating observations and pertinent questions, all of advantages: 1) without the use of the matrix Q in (6), the convergence
which enhance the value of the paper. Since few questions have been of the estimator is quite slow and oscillatory in nature. Thus, addi-
asked by more than one discusser, we will answer each individually. tional logic to damp the convergence characteristic is required; 2)
Prof. Subramanian's questions will be answered in order. rapid sampling requirements to achieve the desired convergence
1) Since B is sparse, BTB will also be sparse. In fact, it has the characteristic and accuracy are necessary. This, in fact, is a hard-
same sparsity as Ybus. In an application, it would probably be more ware requirement to avoid including additional logic in the algorithm;
efficient to use sparse matrix techniques rather than an explicit in- 3) the diagonalization procedure distorts the covariance calculation
version which was convenient for explanation of the algorithm. and burdens the error analysis algorithms.
2) The matrix BTB is always nonsingular if the complex voltage is The meter placement philosophy expressed in our paper is suitable
known or defined at one or more buses. That is to say that A1 has at to calculate an excellent state estimate at a moderate cost. It also
least one column. We plan to measure voltage magnitudes at three provides a reliable redundant metering system. This policy does not
buses to assure that at least one magnitude will be available. Complex have to be implemented in the strict sense of the word. As a matter
voltage can thus be defined at one of these buses. The error in these of fact there are a number of places in the AEP network where line
magnitudes of measured bus voltages is expected to be within one flow measurements will not be available. Practically, this does not
tenth of one percent. penalize the estimate. Thus we have found no need to process bus
3) As the discusser points out (6) and (7) express E2 for iteration k voltage magnitudes or inputs to obtain the excellent results reported
in terms of E2 for iteration k-i using vr as a parameter. There is in our paper.
essentially no saving to be realized in using (11) rather than the Mr. Stillman's discussion includes various interesting points and
procedure of the paper for a Gauss iterative scheme. Solving (11) by questions. He is right in his comment that the high computational
Newtons's method involves introducing a Taylor series expansion for speed of the algorithm is related to the fact that line flows are mea-
1(E2), and this generates a state vectorfordependent Jacobian. Avoiding sured. This is an important point. In fact, we believe that regardless
such difficulties was the raison d'etre the research that lead to the of the method of solution, appropriate selection of data to be mea-
method presented in the paper. sured is of major importance for the solution of the problem. Line flow
As far as Mr. Jazwinski's comments are concerned, we should like measurements provide a compatible set of input data and in turn a
to point out that the weighted least squares referred to in our paper well conditioned set of equations. In addition, this measuring scheme
was specifically applied to the measuring system with bus injections permits a simple formulation and a solution which requires a rela-
and line flow measurements. The problem, with this measuring sys- tively small number of arithmetic operations. The simplicity of
calculations is due to the fact that the matrix G in (6) is constant and
the real and imaginary components of the vector E2 and er are un-
Manuscript received April 13, 1970. coupled.
DOPAZO et al.: STATE CALCULATION OF POWER SYSTEMS 1707

The discusser states that a number of pseudomeasurements are more elements and,more seriously, the real and imaginary variables
used to form the vector El in (2). We wish to clarify that the elements are no longer uncoupled.
of this vector are not pseudomeasurements. It is mentioned in the Mr. Stadlin is apparenitly suggesting that the measured voltage
paper that £1 is a vector of known complex bus voltages. To establish magnituides be treated as the solution values. If this were done, then
(12) would apply with
the system voltage level, at least one bus voltage (a reference bus
voltage) must be known. In this case, the vector E1 has only one
entry whose value is determined by measuring the voltage magnitude Om amq m,p - (13)
and setting the phase angle arbitrarily. However, the vector 51 may being a measured angle difference for line j. The matrix A remains pre-
have more than one entry if complex voltages at various buses are cisely the same as in the paper, and
known. This general formulation of (2) permits inclusion in the state
calculation for one area, the effect of complex bus voltages that have a = G0m (14)
been previously determined from calculations of an adjoiniing area. is a fit of the measured line flows. Since the bus voltage magnitudes as
We agree with the discusser that the values for the network param- measured are intrinsically assumed to be the bus voltage magnitudes,
eters zj and yjp are not exact. Analysis of the effect of random errors in the problem would be linear and iterations would not be required as
these values was not reported in the paper because this effect has been he points out.
found from our studies to be equivalent to that of increasing the The results of the numerical test of this proposal are shown in
error bound of the measured quantities. Nevertheless, we would like
to mention that a number of cases were run in which error bounds of Table II. In this example, it was assumed that voltage magnitude
measurement accuracy was 40.1 percent. This resulted in an rms
up to i 10 percent of value were used to simulate true values of zj and
yjp. These random errors did not deteriorate results to the extreme of voltage measurement error of 0.063 percent as shown in the table.
being unacceptable. We do not expect errors of this magnitude to result Regarding errors in Zj and yjp, please refer to our answer to Mr.
from the line impedance formulas. An important consideration how- Stillman's discussion.
ever, is the appropriate representation of individual transmission Dr. Schweppe's comments regarding our development are most
lines. For instance, a nominal pi representation for the line from interesting.
Dumont to Kammer (approximately 300 miles) was not adequate As regards the physical interpretation of ourmethod, consider the
and hyperbolic correction factors had to be used. Another source of following derivation. The weighted least square residual of measured
error is the effect of temperature on conductor resistance. Variations minus calculated quantities can be written for line flows only and in
of conductor resistance could be larger thani 10 percent. However, complex form as
such error would affect the line impedance and the state vector very J(x) = [Sm -J(g)]*T W[Sm -J(i)] (15)
slightly in the EHV system. Moreover, the effect could be eliminated
for low voltage network applications. where
Regarding Mr. Stillman's statement that we eschew the use of
pseudomeasurements in favor of real measurements, we believe that fj(9) = T [xp - xF] + xpxp*yjp (16)
the virtues and economics of pseudomeasurements are only apparent. zj
Pseudomeasurements can be used only for those quantities which are and
practically independent of network configuration and system condi-
tions, such as injections at conformable load buses. Associated with j measurement index (watts and vars)
each piece of pseudodata, appropriate weighting factors are necessary. p,q bus indices (associated with measurement j)
We have found that pseudodata of poor quality may result in ill- Sm,j measured complex volt-amperes, (P -jQ)
conditioning of the equations when there is interaction between this xp complex voltage at bus p
data and real time measurements if these quantities are properly yjp line charging admiittance, at bus p
weighted. To determine acceptable values for pseudodata and their w a diafgonal weighting matrix corresponding to measurement
associated weights, equipment must be intalled and procedures estab- accuracies.
lished that result in additional expenditures without economic ad- (An advantage to our measurement scheme is that W can
vantage. be modelled by a scalar matrix since all measurements have
With respect to the discusser's question about what to do in the approximately the same accuracy.)
case of foreign systems where measurements can not be made, we wish
to point out that this information is not required for monitoring the Equation (16) can be written in matrix form for all measurements
internal system. Knowledge of the external systems is necessary for as
contingency studies. This problem is outside the scope of this paper. (i) = H[Vc + k] (17)
We do believe however, that pseudomeasurements in foreign sys-
tems would not help much in the solution of this problem. In fact, we where
think that the configuration of the foreign systems and their reaction
to system changes are the important factors to be considered. = Ag vc
Regarding modification of the matrix G due to transmission line
outages, we wish to point out that a function of the data acquisition
H diagonal comnplex matrix with Hi =
c complex vector with kj = xpzjyjp
xp*/zj
system includes monitoring of the status of all EHV circuit breakers.
This information will be used to automatically determine the net- substituting (17) into (15) we have
work configuration and thus the incidence matrix A from which the
matrix G is obtained without operator intervention. J(9) = [Sm - H(vc + k)]*W[SM - H(V3 + k)].
The discusser comments on the good answers obtained in case C Factoring H and regrouping terms gives
and the possibility of reducing the number of data gathering stations.
We feel that the proposed metering scheme is justifiable to maintain
low probability of data being unavailable from adjacent stations.
J(x) = [(H-1Sm -k) -Vc]j*TH*WH[(H-lSm
- k) -Vcj. (18)

Communication links will be installed from the remote stations to The vector of "measured" element voltages, as defined in the body of
the paper, is given in matrix form as follows:
data concentrators in each of three areas of the system and from
these points to the central computer system, rather than assigning Om = H'iSm - k. (19)
individual communication links to each remote station. Substituting (19) into (18) gives
Our answer to Dr. Schweppe's discussion covers Mr. Stillman's
comment on the properties of the residul vector r. J(x)
Km- DCi*TH*WH[VM - c].
= (20)
As far as the possibility of using the algorithm for tracking slow The solution that approximately minimizes (20) and thus (15) is
oscillations, we have not investigated this problem. Although the given by
algorithm is quite fast, other factors to be considered include the
time constants of the measurement system. x = [ATY2WA I -1AFTYyWpm (21)
The authors agree with Mr. Stadlin that voltage magnitudes can be
measured easily and economically. However, their inclusion as part of where y2 is a real diagonal matrix withy_.2 -= lz-2.
The numerical results verifying this approximation for various error
the measurements to be fitted requires the use of a state vector de-
pendent gain matrix. This proper inclusion greatly increases the bounds are given in Table III where they are cormpared with the
dimension of the problem because the measurement vector now has classical weighted least squares algorithm with no approximations.
1708 1IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, SEPTEMBER/OCTOBER 1970

TABLE II
RMS VALUES, STADLIN'S METHOD VERSUS METHOD OF AUTHORS

Stadlin Authors
(MW) (Mvar) (MW) (Mvar)
Measured measured - true 3.15 3.96 3.15 3.96
line ends measured - calculated 2.98 7.67 2.16 2.94
true - calculated 2.95 7.16 2.27 2.65
Nonmeasured true - calculated 4.24 6.34 3.31 3.68
line ends
Magnitude Angle Magnitude Angle
(percent) (degrees) (percent) (degrees)
Bus voltages true - calculated 0.038 0.090 0.030
measured - true 0.063

TABLE III
V/2 RMS VALUES, COMPARISON OF ACCURACY OF METHODS

True True
True Measured -Calculated - Calculated
- Measured -Calculated Fitting to Fitting to
Measurement Fitting to True Flows True Flows
Error Bounds Method t Error* Measured Flows (Measured) (Nonmeasured)
0.35 percent PS 1 5.06 3.81 3.40 4.37
+ 2 percent Reading 2 5.06 3.83 3.40 4.45
3 5.06 4.97 3.97 5.76
0.5 percent FS 1 7.50 5.60 5.02 6.46
+ 3 percent Reading 2 7.50 5.61 5.03 6.62
3 7.50 7.32 5.87 8.44
1 percentlFS 1 13.47 10.2 8.86 11.4
+ 5 percent Reading 2 13.47 10.2 8.88 11.6
3 13.47 13.3 10.5 15.3
2 percentFS 1 27.0 20.3 17.7 22.9
+ 10 percent Reading 2 27.0 20.4 17.7 23.3
3 27.0 26.5 21.0 30.7
* Measurement errors modeled as:
Sm.,j = Si + ki(FS);jc + k21Sj1c2
where
Smwj,
Sh measured and true complex volt-ampere flow
(FS)j errorfull-scale meter reading, measurement j, (reading = -FS, +FS)
ki,k2 bounds, FS and reading, respectively
Cl,C2 complex random numbers within unit disk.
tMethods:
1) Classical least squares without simplifying approximations.
2) Method of authors with squared line admittance weighting.
3) Method of authors without weighting.

Note that the algorithm of the paper is specified without weights, policy of line flow measurement at both ends is not as strict as it may
however without loss of generality and without additional computa- seem. There are many line flows in the AEP .345/765kV network
tional effort they can be included. where metering facilities are not available. Practically this does not
When one compares the algorithm of the paper to the tracking degrade the state estimate as simulated usinig Monte Carlo trials.
estimator of [12], several important factors come into play. Equation In comparing the accuracies of weighted least squares with our
(21) has an intrinsically constant gain matrix. The tracker of the cited method it should be pointed out that the errors simulated were not
reference has a time varying gain matrix which is used as a constant taken as Gaussian. Because of the physical characteristics of metering
for a specified period of time. However, this gain matrix must be devices, i.e., finite full scales, a rectangular probability density func-
evaluated under certain system conditions to maintain good perfor- tion was used for error analysis. This also maximizes the uncertainty
mance and numerical stability. This requirement introduces additional of the errors in the measurement system [15]. For this set of condi-
specialized computer logic into the tracking algorithm. Assuming that tions Table III compares both methods as a function of the errors
this problem is solved, it seems that at best the tracker of [12] would in the measurements. It is evident that there is no degradation as the
require twice the storage and twice the time required by the solu- error in the measurements increases. Only if the measuring system is
tion of (21) in the tracking mode. characterized by a Gaussian process, do we agree with Dr. Schweppe
Dr. Schweppe is absolutely right in that weighted least squares is a that, "weighted least squares is optimal in a well defined statistical
more general approach: it can process any combination of data, it can sense" [15], [16].
be formulated in different coordinate systems, it can use pseudo-
measurements, etc. A careful study of the advantages arising from the REFERENCES
use of pseudomeasurements, injections, line flow measurements and [15] S. Lees, "Uncertainty and imprecision," Trans. ASME, J.
voltage measurements compared with those obtained through process- Basic Engrg., June 1966.
ing compatible redundant data with its associated computational [16] R. E. Kalman, "A new approach to linear filtering and predic-
simplifications made us eschew such versatility. The measurement tion problems," Trans. ASME, J. Basic Engrg., March 1960.

Das könnte Ihnen auch gefallen