~

IEEE Transactions on Power Systems, Vol. 10, No. 2,May 1995 671

A THREE-PHASE POWER FLOW METHOD

FOR REAL-TIME DISTRIBUTION SYSTEM ANALYSIS

Carol S. Cheng Dariush Shlrmohammadi

Member Senior Member
Energy Systems Automation Group
Pacific Gas and Electric Company
San Francisco, CA

Abstract-This paper presents a three-phase power flow By comparing of these methods, it is seen that the method
solution method for real-time analysis of primary described in Tl] is able to take maximum advantage of the
distribution systems. This method is a direct extension of unique topological structure of a lstribution system, which
the compensation-based power flow method for weakly is radial or weakly meshed, to achieve high speed, robust
meshed distribution systems[11 from single phase to three- convergence and low memory requirement. In addition, the
phase, with the emphasis on modeling of dispersed V, I formulation used in [l] allows the extension from
generation (PV nodes), unbalanced and distributed loads, single-phase to three-phase to be straightforward.
and voltage regulators and shunt capacitors with automatic
local tap controls. The method proposed here is capable of In th~spaper we propose a three-phase power flow solution
addressing these modeling challenges while still method for real-time analysis of primary distribution
maintaining a high execution speed required for real-time systems. The focus of this paper is on several modeling
application in distribution automation systems. The paper issues related to distribution system operation incluhng
also includes test results from the application of a computer unbalanced multi-phase operation, dispersed generation (PV
program developed based on the proposed method to large nodes), distributed loads, and voltage regulators and shunt
primary electric distribution systems. capacitors with automatic tap controls. However, in this
paper we do not address other important issues related to a
Keywords: distribution power flow, distribution automation, real-time application such as data interfaces and
real-time power flow. requirements and user interfaces. The overall speed,
modeling detail and convergence performance of the
I. INTRODUCTION algorithm proposed here make it suitable for real-time
applications in distribution automation systems.
It has been realized that many real-time application
programs in the area of distribution automation require a 11. BASIC ALGORITHM FOR POWER FLOW
robust and efficient power flow solution method. Such a ANALYSIS OF WEAKLY MESHED
power flow solution method must be able to model the PRIMARY DISTRIBUTION SYSTEMS
special features of distribution systems in sufficient detail.
Some of the more prominent features of electric distribution The primary feeders of the lstribution system consist of
systems are: mostly three-phase overhead or underground line sections
(branches), and occasionally double-phase or single-phase
Radial or near radial structure line sections near the end of the feeder laterals. In the three-
Multi-phase, unbalanced, grounded or ungrounded operation phase power flow algorithm, we number each node or line
Dispersed generation section in the network by a single index, regardless of the
Multi-phase, multi-mode control distribution equipment number of phases of this node or line section. The series
Unbalanced distributed loads impedance of a line section, I, is represented by a 3x3
Extremely large number of brancheshodes matrix:
zaa.1 zab,l zac,l
Some power flow techniques that could deal with ~ o m eof
(1)
these features have already been developed [l-51. One type
of these algorithms is based on the system admittance matrix zac,l 4c,l zcc,l
and an iteration scheme similar to Newton-Raphson power
flow[4,5]. The other type is featured by the compensation- If any phase of the line section does not exist, the
based method with breakpoint impedance matrix[ 11, along corresponding row and column in this matrix contain all
with several efforts on efficient formation of the zero entries. Figure 1 shows line section 1 between nodes i
compensation matrix and handling of PV nodes[2,3]. a n d j with shunt admittances and loads attached to each
node.
root other line
94 SM 603-1 PWRS A paper recommended and approved node
by t h e IEEE Power System Engineering Committee of t h e sections
IEEE Power Engineering Society f o r p r e s e n t a t i o n a t
t h e IEEE/PES 1994 Summer Meeting, San Francisco, CA,
July 24 - 28, 1994. Manuscript submitted August 2, other linc
1993; made a v a i l a b l e f o r p r i n t i n g May 17, 1994. seclions

Fig. 1. A three-phase line section

0885-8950195/$04.00 0 1994 IEEE
672
To apply the single-phase power flow algorithm described in
[l] to three-phase we first convert the primary distribution
network to a radial system by breaking all the loops. Line
sections in the radial network are ordered by layers away
from the root node (substation bus) as shown in [I], a
diagram from [ 11 is reprinted below. where Jla, Jlb and Jlc are the current flows on line section
I and M is the set of line sections connected to node j .
Note that the negative sign in ( 3 ) is to keep consistent
with current injections as calculated in (2).

3 Forward sweep to update nodal voltage: starting from the

first layer and moving towards the last layer, the voltage
at node j is:

After these three steps are executed in one iteration, the

power mismatches at each node for all phases are calculated:

Branch numbering scheme for radial

distribution network from Fig. 2 of reference [ 11

Let the root node be the slack node with known voltage
magnitude and angle, and let the initial voltage for all other If the real or imaginary part (real or reactive power) of any
nodes be equal to the root node voltage. The iterative of these power mismatches is greater than a convergence
algorithm for solving the radial system consists of three criterion, steps 1, 2 and 3 are repeated until convergence is
steps. At iteration k: achieved.
1. Nodal current calculation Equations (2)-(5) provide the solution for a three-phase
radial network. In a meshed system, it is necessary to
simulate loops by injecting currents at both ends of all
breakpoints. Currents at breakpoints are calculated using
the compensation technique and a breakpoint impedance
matrix. This procedure is based on the same principle as
described in [I]. Figure 2 shows a three-phase breakpoint, j .
For th~sbreakpoint, currents must be injected to all three
where phases with opposite polarity at the two end nodes, j l and
~ I L ?11
~ 6. 11, are current injections at node i j2. These currents are determined in an iteration loop
corresponding to constant power load outside the radial power flow solution. At iteration p,
and shunt elements
S',,, SI,,
S,, are scheduled (known) power injections
at node i
Via, L$,, V,, are voltages at node i
Yla,Ylb' Y,, are admittances of all shunt elements at
node z

Later in the paper we will show how distributed loads

along line sections can be converted to lumped loads at
system nodes. We will also describe the special case
where loads are fed via ungrounded three wire systems.

2. Backward sweep to sum up line section current: starting

from the line section in the last layer and moving
towards the root node, the current in line section I is:
Fig. 2. Three-phase breakpoint representation
using nodal current injections
 673

The three-phase currents for all breakpoints are obtained by values. In order to obtain the scheduled voltage magnitude
solving the following complex linear equation: at a PV node, we need to determine the correct amount of
reactive power or reactive current injection generated by the
unit. Therefore, the problem of compensating PV node
(7) voltage magnitude becomes: Find the reactive current
injection, Z for each PV node so that the voltage
where [?Iis the vector of three-phase voltage mismatches magnitude, $I, of this node is equal to the scheduled value.
for all breakpoints and [ z,] is a constant complex matrix, Since the relation between Iq and IPl is nonlinear, Zq can
referred to as the breakpoint impedance matrix. The only be determined iteratively. A PV node sensitivity matnx
physical meaning and the formation of [ z,]for single-phase is introduced to approximate the nonlinear relation between
systems have been well explored in [ 11-[3]. For three-phase Ill and lq and is used to evaluate Zq iteratively as described
systems, since all the loops are likely to be among three- below.
phase line sections, the break point impedance matrix
mainly consists of 3x3 block submatrices. Numerically, the
diagonal submatrix Z,, is equal to the sum of the line section 3.1 PV Node Sensitivitv Matrix
impedance matrix in (1) for all the line sections in loop i.
The off-diagonal submatrix, Zq, is nonzero only if loop i and A PV node is modeled in a similar manner as in [5], i.e., the
loopj share one or more common line sections. The signs constants for a generator are the three-phase real power
of the off-diagonal submatrices depend on the relative output and the magnitude of the positive sequence voltage.
direction of the breakpoint current injections for loops i and The use of positive sequence representation for voltage
j. magnitude regulation makes it possible to properly represent
the automatic voltage regulation (AVR) mechanism of a
In our algorithm, [ Z,] is formed and factorized at the generating unit, where in most cases, the average of voltage
beginning before any iteration is performed. After each magnitude of all three phases is the voltage magnitude that
radial power flow is converged, the three-phase voltage is regulated.
mismatches at all breakpoints are compared with a
threshold. If the mismatch of any breakpoint of any phase is The incremental relation between the magnitude of positive
greater than the threshold, (7) is solved and the breakpoint sequence,voltage and the magnitude of the positive sequence
current [Jl is updated. This iteration process continues until reactive current injection is expressed as
all the breakpoint voltage mismatches are less than the
threshold.
[Z,][Z,l(y) = [AV](')

111. COMPENSATION FOR PV NODES where [Zd is a constant real matrix, referred to as the PV
node sensitivity matrix. The dimension of [Zd is equal to
Since dispersed generators are becoming prevalent in the number of PV nodes. Column j of [Zv]may be
distribution systems, modeling them in a distribution power determined by applying r j = {0.,1.) to PV node j with all
flow algorithm has become an inevitable task. Depending loads and sources removed, and solving a positive sequence
on the contract and control status of a generator, it may be network with one back and forward sweep for the change of
operated in one of the following modes: voltage magnitudes at all PV nodes.

1. In "parallel operation" with the feeder, i.e., the generator Equivalently, [Zd can be formed by observing the following
is located near and designated to supply a large load with numerical properties of its entries. The diagonal entry, z,,,
fixed real and reactive power output. The net effect is in [Zdis equal to the modulus of the sum positive sequence
the reduced load at a particular location. impedance of all line sections between PV node i and the
2. To output power at a specified power factor. root node (substation bus). If two PV nodes, i andj, have
3. To output power at a specified terminal voltage. completely different path to the root node, then the off-
diagonal entry z,, is zero. If i a n d j share a piece of common
In power flow context, the generation nodes in the first two path to the root node, then z is equal to the modulus of the
cases can be well represented as PQ nodes, which require sum positive sequence impeciance of all line sections on this
little special treatment in the power flow algorithm. The common path. Based on these, [Zd can be formed by
generation node in the third case must be modeled as a PV identlfying the path between PV nodes and the root node.
node. Special procedures must be performed to maintain its When forming [Zv]for a group of feeders connected to
voltage magnitude as well as to monitor its reactive power different substations, the impedance paths will be between
capability. the PV node on a feeder and the substation bus (root node)
to which the feeder is connected.
We have developed a compensation method using a PV node
sensitivity matrix to eliminate the voltage magnitude In our power flow algorithm, [Zd is formed for all initial
mismatch for all PV nodes. The basic idea of t h s method PV nodes and factorized into LU before any power flow
can be explained as follows. Suppose a power flow as iteration is performed. Dependrng on whether there are PV
described in Section I1 has converged, and the voltage to PQ node conversions, [Zd and its factors may have to be
magnitude at PV nodes are not equal to the scheduled updated, as to be described in section 3.2.
674

3.2 Iterative Process for VoltaEe Magnitude Correction

Suppose there are n PV nodes in a system. The reactive

current injections at these PV nodes are determined through Qid in (1 1) is the scheduled reactive load at PV node i.
an iteration loop outside the breakpoint current
compensation. Each time after the breakpoint voltage 4. Qlg then is compared with the reactive power generation
mismatches are reduced below a threshold, the following limits. If Qig is within the limits, i.e.,
steps are performed to correct the voltage magnitude at all
PV nodes. At iteration y:

1. Calculate positive sequence voltage magnitude mismatch then the corresponlng reactive current, I,qa, I,,,, I,,, are
for all PV nodes injected to PV node i according to (11). In subsequent
iterations, these currents will be combined with other
i = 1,2, ..., n (9) nodal current injections. Otherwise, if Qlg violates any
reactive power generation limit, it will be set to that
Where v," is the scheduled voltage magnitude for node i. limit, divided by three for three phases and combined
If any of these mismatches is greater than a threshold, with the reactive load of each phase at this node.
then perform the next step: Subsequently, the row and column in the PV node
sensitivity matrix, [Zd, corresponding to this node are
2. Solve for PV node reactive current injection using (8). removed and the LU factors of [Zd are updated.
The solution provides a linear approximation of the
reactive current injection needed to eliminate the voltage The iteration described in steps 1-4 will continue until the
magnitude mismatch in this iteration. If the reactive voltage magnitude mismatches for all PV nodes as
power generations were unlimited, we would inject I,qn calculated in (8) become less than a threshold.

ea,
I, b, I,,, at 90 degrees leading the corresponding voltage
V,& V,, at each PV node, i : The flow chart of overall power flow algorithm is shown in
Figure 3. It is seen that the algorithm consists of three
nested iteration loops for radial power flow, breakpoint
voltage compensation and PV node voltage magmtude
compensation, respectively. The termination of each
iteration loop is controlled by a threshold. Our experience
shows that once the breakpoint compensation starts after the
initial iterative radial power flow, subsequent radial power
flows always converge in one or-two iterations under the
, ,6are voltage angles of the three same threshold for power mismatches. The same pattern
' node. Since in reality the reactive has been observed between PV node compensation and the
power capability of a generator is always limited, the breakpoint compensation.
reactive power limits must be checked first to determine
whether the required current injections are available, as
in the next step:
Read data. convert to radial
3. Calculate the required reactive power generation Qjg for network and order by layers
all PV nodes:
c
Form breakpoint impedance
matrix & PV sensitivity matrix

where (2;' is the new reactive power injection at node i.

It is calculated using the PV node voltage and the new
current injection:
Solve for J

(PV node /VImismatch<e2? for Iq -

The new current injection at PV node i is a combination by E d 9 )
of the desired reactive current injection and load current
injection:

Fig. 3. Flow chart of the power flow algorithm

 675

IV. OTHER MODELING ISSUES FOR Figure 5 shows the variation of q with difFerent values of V,
DISTRIBUTION SYSTEMS and V2. It is seen that q varies around 0.5 in a fairly narrow
range as the voltages at the two end nodes assume Merent
T h s section addresses several practical modeling issues values. Since the voltage drop across a single line section is
involved in the three-phase power flow algorithm. usually small (VIgV2), it is reasonable to choose q to be 0.5.
Based on this observation, in our power flow algorithm, the
4.1 Distributed Loads total chstributed load on each phase of a line section is
lumped half-half at the line section's two end nodes.
In a lstribution system, a primary feeder supplies loads
through lstribution transformers tapped at various locations 1 2
along line sections. If every load point were to be modeled
as a node, the number of nodes in a system would become
prohibitively large. In ow power flow algorithm, we model
these distributed loads by approximating their effect on
nodal voltage magnitude via equivalent lumped loads.

Consider any phase of a line section of length L shown in

Figure 4(a). Suppose the load is uniformly distributed along
the line section with the total value of P+jQ. We want to
lump a portion of the total load at one end node and the rest 4
at the other, as shown in Figure 4@), such that the voltage
- 5
magnitude drop at node 2 is the same as that caused by the
distributed load along the line. In our derivation, we assume Fig. 4. Equivalencing distributed loads to lumped loads
that the voltage angles at both nodes 1 and 2 are small and
negligible. This assumption is based on the examination of
many numerical results whch have consistently shown that
the angle deviation for an entire feeder is usually no larger
than a few degrees. We also assume that there is a linear
drop of the voltage magnitude along the line section, i.e.,

(13)
where x is any point on the line section measured from node
1, as shown in Figure 4(c). The current flowing in the line
section is a function of x:

Fig. 5. Variation of q as a function of V

The voltage drop across the line section is

4.2 Load Representation for Three Wire Ungrounded
Systems

In practice, some lstribution feeders are constructed as

three wire systems without ground wire (a common practice
where rl+jxl is the impedance per unit length of the line in Pacific Gas and Electric Company). Loads of each phase
(ohms/meter). This voltage drop should be equal to that are supplied through distribution transformer connected
caused by the equivalent lumped load at node 2, q(P + j @ : between two phases of the primary feeder, as illustrated in
Figure 6.

where q denotes the portion of the line section total load

P+jQ that is lumped at the end node, 0.0 < q < 1.0, and R,
+ j X , is the line section impedance. Substitute (14) into
(16), integrate the left-hand-side and solve for q, we have

Fig. 6. Load supplied via three wire system

676
In this case, the load cannot be decomposed by phases and
the following equations should be used in the back sweep to
replace equations (2) and (3):

where (21) is to simulate the line drop compensation

(LDC) scheme. This scheme requires a regulator to
maintain voltage at some point remote from itself by
means of the dial settings on the regulator’s control
panel. In (21), RF XT are the resistance and reactance
calculated approximately from the LDC’s R, X settings,
and J,@) is the secondary line section current obtained
during the back sweep.
where Slab, Slbcand S,,, are the total load located between
phases ab, bc and ca, respectively. Other distribution 2. If VsM is greater than the upper limit, set Vs@)to the
equipment connected to an ungrounded system are also limit, calculate the new tap position and round it up to
connected between phases and their V-I characteristics can the nearest lower ta position. If V” , is less than the
be directly modeled in this solution technique. P
lower limit, set v,@to the limit, calculate the new tap
position and round it up to the nearest upper tap position.
4.3 Voltage Remlators 3. Check if the tap position obtained in step 2 exceeds the
maximum boost or buck limits. If it does, set the tap at
Voltage along primary feeders are often controlled by the corresponding limit.
voltage regulators. These regulators are auto-transformers
with individual taps on their windings. Typically, the 4. Re-calculate Vs@jusing the new tap value and continue
regulator is used to boost (increase) or buck (decrease) the the forward sweep.
voltage by a variable amount up to 10 percent. A regulator
can be operated in manual mode or automatic mode. In the With these steps, the tap positions of regulators in automatic
manual mode, the output voltage can be manually raised or mode will be determined after each radial power flow
lowered on the regulator’s control board. In the automatic solution. This assures that the regulated voltage is witlun
mode, the regulator control mechanism adjusts the taps to the desired range. Since in primary feeders there are usually
assure that the voltage being monitored is within certain only a few voltage regulators with automatic tap-changing
range. mechanism, the addition of the above steps into the forward
sweep of the radial power flow algorithm has no noticeable
The proposed three-phase power flow algorithm is capable effect on the overall solution speed .
of modeling Y connected three-phase voltage regulators
operating in both modes. The regulator in each phase is
represented by a series impedance and an ideal transformer 4.4 Shunt Capacitors
with taps on the secondary. In the manual mode, the tap
position is known and so is the turns ratio of the ideal Shunt capacitors are often used in primary feeders to
transformer. Thus during the back-forward sweep in the improve voltage profile and reduce line losses. Most shunt
radial power flow solution, the turns ratio will be applied to capacitors are equipped with a local control scheme for
the line section current calculation and the node voltage switching on-line or off-line according to certain conltions.
calculation wherever the regulator is encountered. In the The most commonly-monitored parameters for capacitor
automatic mode, the tap position is unknown prior to the control are time, temperature, voltage, current, KVAR and
power flow solution. It will be adjusted within the radial power factor. In practice capacitor controls use combination
power flow algorithm in order to simulate the regulator’s of these parameters, such as time control with voltage
automatic tap-control mechanism. override or temperature control with voltage override.
Specifically, the tap positions of all regulators in automatic The three-phase power flow algorithm is capable of
mode are first initialized to the neutral position (neither handling each of these controls. The following general
boost nor buck) before power flow iterations. During the procedures are used to simulate the automatic local control
forward sweep in a radial power flow solution, the following scheme of capacitors:
steps are performed when such a regulator is encountered.
At iteration k: 1. For those capacitors with time or temperature control,
initialize their o d o f f statuses based on given time and
1. Calculate the secondary voltage, V,, of each phase of the temperature and keep these statuses fixed during the
regulator using the given tap values, and check if the power flow solution.
voltage is withm the lower and upper limits of the
regulated voltage setting:
 677

2. For those with other controls, initialize their on/& more dispersed generators and controllable devices are
statuses based on given time or temperature, or initialize expected to be connected to or installed in primary feeders.
their statuses to off if time or temperature is not
monitored.
Table 1 Execution Time in Seconds on SUN SparclO
3. In the outermost iteration loop, check the monitored
quantity against the setting. For example, the monitored
quantity is the node voltage at the capacitor bank, and
the system has PV nodes. In this case, the outermost
iteration loop is the PV node compensation loop. Once
the PV node compensation is finished in one iteration,
the monitored voltage will be compared to the capacitor
setting.

4. Modify the odoff statuses of all capacitors according to

the comparison results obtained in step 3 and continue The hlgh execution speed may be attributed to two factors.
for the next radial power flow solution. First, the radal power flow solution algorithm with
backward and forward sweeps, successfully used with loop
The power flow is considered to have converged if all compensation as proposed in [l], is very efficient,
convergence criteria are satisfied, plus there is no further particularly for large systems, as shown in Table 2. Note
change in capacitors odoff statuses. The capacitor control that the solution time increases only linearly with system
logic actually adds one more condition for the termination of size. Second, only small and constant matrices need to be
the outermost iteration loop (see Figure 3). formed for the breakpoint voltage compensation and the PV
node voltage magnitude compensation. The dimensions of
these matrices are equal to the number of loops and the
V. RESULTS number of PV nodes, respectively, which are substantially
smaller than the dimension of impedance or admittance
Based on the proposed solution algorithm, we developed a matrices in traditional power flow solution techniques.
three-phase distribution power flow program, DISFLO, as These matrices need to be formed and factorized only once
one of the real-time applications for distribution automation for each power flow solution, and only the factors of the PV
systems. The program was tested using typical primary node sensitivity matrix need to be updated for PVPQ node
distribution feeders in PG&E. The test system presented conversions. These features not only make the execution
here consists of two large primary feeders supplied by a speed less sensitive to the number of loops and PV nodes,
21kv substation. These two feeders are of radial structure but also keep the requirement of computer memory
with 233 switches at various locations. In our test, these resources at a minimum.
switches were opened and closed to simulate Werent
system topologies and different number of loops. Also since
it is expected that in the near future increasing number of Table 2 Execution Time for Three-phase Radial Systems
dispersed generators will be directly connected to primary
distribution feeders, PV nodes were introduced at various
locations for testing purposes. Other major features of the
I No. of multi-phase 1 Actual No. of nodes I Time I
system are listed below: nodes (line sections) (line sections) (seconds)
690 1555 0.69
No. of line sections : 1418 (Multi-phase), 3368 (actual) 1419 3371 1.46
No. of nodes : 1419 (Multi-phase), 3371 (actual) 2479 5983 2.39
No. of capacitors : 20 (with adjustable odoff status)
No. of regulators : 3 (with adjustable taps)
The proposed three-phase power flow algorithm is also very
Table 1 shows the solution efficiency as the numbers of robust. Convergence has been reached for all the cases
loops and PV nodes are varied in the system. The elapsed tested. Monotonous convergence pattern has been observed
execution time is measured on a single CPU SUN SparclO for the radial power flow (innermost iteration loop) and for
workstation. This time includes the execution of all steps breakpoint voltage compensation (second iteration loop).
illustrated in Figure 3 except ''read data" in the first box, The PV node compensation converges almost
and "report" in the last step. monotonically. A typical convergence pattern for PV node
compensation for the 1419 node system (with 40 PV nodes
It is seen that the proposed solution algorithm is efficient and 30 loops) is shown in Table 3. The slight fluctuation of
despite the large system size and the complexity of system the maximum voltage mismatch is caused by introduction of
equipment. As the number of loops and PV nodes is other reactive power sources via capacitor switching, as well
increased, the execution time increases less than linearly. as by PVPQ node conversions.
This feature is favorable for large primary distribution
systems at present and in the near future, where more and
 678
Table 3 Typical ConvergencePattern for PV Node for Weakly Meshed Distribution and Transmission
Voltage Magnitude Compensation Networks", IEEE Trans. on Power Systems, Vol. 3, No.
(1419 nodes, 30 loops, 40 PV nodes) 2, May 1988, pp.753-762.

I Iteration No. I max(lAV1) / Vrated I [2] G. X. Luo and A. Semlyen, "Efficient Load Flow for
Large Weakly Meshed Networks", IEEE Trans. on
( Vrated'
12b) Power Systems, Vol. 5, No. 4, November 1990,
y= 1 0.03443 pp. 1309-13 16.
v=2 0.04154
I v=3 I 0.04095 I [3] D. Rajicic, R. Ackovski and R. Taleski, "Voltage
Correction Power Flow", Paper 93 WM 060-4 PWRD,
I v=4 I 0.03878 I presented at the IEEERES 1993 Winter Meeting,
I v=5 I 0.006187 I Columbus, OH, Jan. 31 - Feb. 5 , 1993.
0.002135
0.0006860 [4] T.-H. Chen, M.-S. Chen, K.-J. Hwang, p. Kotas and E.
0.0003218 A. Chebi, "Distribution System Power Flow Analysis - a
Rigid Approach", IEEE Trans. on Power Delivery, Vol.
0.00008039 6, NO. 3, July 1991, pp.1146-1152.
In summary, the speed and robustness of this solution [5] W.Xu, J. R. Marti and H. W.Dommel, "A Multiphase
algorithm makes it suitable for real-time application in Harmonic Load Flow Solution Technique", IEEE Trans.
distribution automation systems. on Power Systems, Vol. 6, No. 2, Feb 1991, pp. 174-
182.
VI. CONCLUSIONS Carol S. Cheng (M'92) was born in Beijing, China. She
received her B.S. in Electrical Engineering from Northern
This paper presents a unique three-phase power flow
solution method for large distribution systems. The kernel Jiaotong University in Beijing in 1982, M.S. in Mechanical
of this method is an efficient three-phase radial power flow Engineering from the University of Cincinnati in 1986, and
with breakpoint voltage compensation, extended from its Ph.D. in Electrical Engineering from Georgia Institute of
single-phase version proposed earlier in [l]. The idea of Technology in 1991. From 1982 to 1985, she worked for
, compensation is also employed successfully to eliminate the Northern Jiaotong University as a Teaching Assistant.
voltage magnitude mismatches at PV nodes, thus making Since 1992 she has joined PG&E as a Consulting Engineer.
this method capable of handling dispersed generation in Her current research interests include steady state power
primary distribution systems. Other issues involved in system analysis methodologies and their application in
distribution system operation, such as multi-phase operation distribution automation systems.
with unbalanced and distributed loads, regulators and
capacitors with automatic tap controls, can also be addressed
within the algorithm. Test results show that the method is Dariush Shirmohammadi (SM'89) received his B.Sc. in
efficient and robust, and hence suitable for analyzing the Electrical Engineering from Sharif University of Technology
real-time operation of large primary distribution systems. in 1975 and M.A.Sc. w d Ph.D. in Electrical Engineering
We have developed the computer program DISFLO based on from the University of Toronto in 1978 and 1982
this algorithm for use in PG&E's distribution automation respectively. Between 1977 and 1979, he worked in Hydro
systems. Quebec Institute of Research (IREQ) on the subject of
external insulation. Between 1982 and 1985, Dariush
VII. ACKNOWLEDGMENT
worked in Ontario Hydro on the development and
application the EMTP. Dariush worked with the Systems
The authors would like to thank Mssrs. Brian Agnew, Engineering Group of the Pacific Gas and Electric Company
Robert Hartwell, Tom Hillesland, Bill Lemon, Paul (PG&E) between 1985 and 1991 where he developed
Mauldin, Farajollah Sou& and Jimmie Yee of PG&E for advanced methodologes and computer models for the
their useful comments in the development of this algorithm. analysis, optimization and costing of transmission and
PG&E R&D department funded the development of the distribution networks. Presently, Dariush is with the Energy
DISFLO program. Systems Automation Group of PG&E where he is
responsible for the development and implementation of
REFEENCES automation technologies in electric distribution system
operation. Dariush is a registered Professional Engineer in
[ l ] D. Shirmohammadi, H. W. Hong, A. Semlyen and G. a e Province of Ontario, Canada.
X. Luo, "A Compensation-based Power Flow Method
 679

Discussion parameters. The test system we described in the paper has an x/r
ratio ranging from 0.1968 to 5.0. Voltage angles up to 15" has been
Hormoz Kazernzadeh and Fan Zhang (ABB Automated Distribution
Division, 1021 Main Campus Drive, Raleigh, NC 27606-5202) observed (under loading conditions particularly set for testing
Congratulations to the authors for providing such a practical and well- purpose and multiple substations). Convergence has been reached
organized paper. It has been known for some time that some special power with approximately the same number of iterations for all the cases
flow methods are computationally more attractive than Newton-Raphson we tested.
type methods in solving the power flow of distribution systems without
considering PV nodes. This paper extends one of these methods by taking
the PV nodes into account. Authors opinions on the following issues would b) A good initial guess for the reactive power injections at PV
be appreciated: nodes will result in smaller i+tial voltage magnitude mismatches
and thus will speed up the convergence of the algorithm. It will also
a) What are the assumptions (e.g. voltage angles among all buses are lead to more accurate predictions of PVPQ conversions. In our
small, high d r ratio etc.) you make to derive the equation (8) from its
corresponding complex equation, such that the real power computed by the test, however, we used zero initial reactive power for all generating
equations similar to equations (11) and (12) is close to the desired real units, duc. to the simulated nature of the entire generation data. It is
power ? Did you encounter any divergent cases if the test system hardly definitely useful if some heuristics can be derived to predict the
satisfies the assumptions ? initial reactive power for all generating units, since real-time
b) How sensitive is the computation performance versus the initial reactive
measurements for co-generating units are less likely to become
power injection for PV nodes? What is your suggested initial value? Is it available.
taken as zero, based on an assumed power factor, or other ?
c) We tested an algorithm in which the PV node voltage magnitude
c) How is the computation performance changed (better or worse) if the
is compensated first and the breakpoint voltage mismatch is
iteration loops for breakpoint voltage compensation and PV node voltage
magnitude compensation are combined in one iteration loop, or if the order reduced later on a 35 node system with 4 PV node placed at
of these two iteration loops is reversed? various locations, and under different loading conditions. The
algorithm tends to diverge. We have not tried to combine the
d) Distribution systems are usually unbalanced due to unbalanced loading breakpoint impedance matrix and the PV node sensitivity matrix in
of the different phases. The authors may consider using a 4x4 line
impedance matrix (equation 1) to properly represent neutral current due to order to compensate both in one iteration loop, because the former
unbalance. is a complex matrix based on three-phase parameters and the later
is a real matrix based on positive-sequence parameters.
e) The method of backward and forward sweep (otherwise known as the
Ladder Technique) is based totally on node voltages. It seems more logical The reason that we use the proposed compensation sequence, i.e.,
to use a voltage mismatch criteria, rather than a power mismatch criteria
(step 3), at the root node. eliminating the voltage mismatch at breakpoints followed by
eliminating the voltage magnitude mismatch at PV nodes, is based
0 However infrequent, sources are sometimes tied together in distribution on the following observations: the reactive power injection at each
systems for short periods of time. A real-time power flow must be able to PV node and its voltage magnitude adjustment is mainly a local
model parallel sources. Authors experience in modeling parallel sourCes
using your own method would be appreciated. effect. The relation between PV node voltage magnitude and
reactive power injection is more nonlinear than the relation between
Manuscript received August 9,1994. breakpoint voltage and breakpoint current injection. Therefore the
calculation of PV node reactive current injection (i.e., solution of
equation (8) ) should be done using the best guess of the system
state in that iteration, where everything is satisfied for a solution
except the PV node voltage magnitude. In this way, the
Carol S. Cheng and Dariush Shirmohammadi: The authors nonlinearities involved in this compensation are confined to such a
would like to thank the discussors for their interests in the paper level that a linear equation (8) serves as a valid approximation of
and their thoughtful questions. Our opinions on each item are as the actual gradient leading to convergence.
follows:
d) We hlly agree with the discussors on using a 4x4 line impedance
a) We derived Equation (8) based on a sensitivity study of PV node matrix. We used the 3x3 line impedance matrix since most feeders
voltage magnitude with respect to PV node reactive current in our system are 3-wire feeders, and neutral impedance are not
injection at no load condition. No special assumptions were made readily available. The proposed algorithm can be readily extended
for this purpose. We neglected all shunt elements. During the PV to accommodate the 4x4 line impedance matrix. Preferably,
node compensation, this equation is solved in each iteration to standard :nethods of reducing 4x4 impedance matrices to 3x3
provide the correct magnitude of the positive-sequence reactive matrices can be used to account for ground wire effect.
current injection. The angle of the reactive current injection for
each phase is determined according to the voltage angle in each e) We use the power mismatch criteria at every node as condition
iteration, i.e., the reactive current for each phase is always injected of convergence. This is considered the only accepted criterion to
at 90" le,ding its node voltage of the same phase. In this way, the guarantee true convergence for power flow algorithms.
injected current contributes only to the reactive power component
at this node. The real power output of each unit remains unchanged f) The proposed algorithm is able to model multiple parallel
throughout the power flow solution. sources, typically multiple substations. Each substation is modeled
as a slack node with known voltage magnitude and angle.
Based on our test results, convergence characteristics of the
proposed algorithm is insensitive to loading conditions and network Manuscript received September 9, 1994.