Fast Decoupled Power Flow Unbalanced Radial Distribution Systems
Fast Decoupled Power Flow Unbalanced Radial Distribution Systems
4, November 1995
2045
Hsiao-Dong Chiang
School of Electrical Engineering
Cornel1 University, Ithaca, NY 14853 USA
K-r&
power flow, load flow, radial network, distribution system, fast decoupled
INTRODUCTION
Basic Concept
Our approach is based on the fact that, given the voltage and current at one end of a distribution feeder and the
currents injected into each sub-lateral, it is possible to
compute all voltages and currents in the rest of the feeder.
Since the current at the end of the feeder is zero, the end
voltage can be taken as the unknown. And since the voltage at the source is specified we have a boundary condition
we can use to compute a voltage mismatch.
Exploring the numerical characteristics of a distribution line, we can make decoupling approximations to the
Jacobian used to update the unknown end voltages. This
results in a very fast Newton-like solution to a reduced set
of load flow equations.
For comparimn, this fast decoupled algorithm was
implemented along with the standard Newton method, the
implicit Z b Gauss method, and the backlforward sweep
method of [9]. Each method was evaluated and compared
on a 292-bus and a 394-bus test system.
0885-8950/95/$04.00
Ca 1995 IEEE
2046
BUS & LATERAL INDEXING
In most typical load flow formulations, a set of equations and unknowns is associated with each bus in the network and these equations and unknowns are organized by
a particular bus ordering. Due to the radial structure of
the systems under consideration, the number of equations
and variables can be reduced so that each set of equations
and unknowns corresponds to a n entire lateral instead of
an individual bus. Our formulation therefore calls for a n
appropriate lateral indexing to order these equations and
variables.
Aradial system can be thought of as a main feeder with
laterals. These laterals may also have sub-laterals, which
themselves may have sub-laterals, etc. So first we define
the level of lateral i as the number of laterals which need
to be traversed to go from the end of lateral i to the source.
For example, the main feeder would be level 1,its sub-laterals would be level 2, their sub-laterals level 3, etc.
The laterals within level 1 are indexed according to the
order visited during a depth-first traversal of the network.
Each lateral can be uniquely identified by a n ordered pair
( k , m ) where 1 is the lateral level and m is the lateral
index within level 1.
Buses are also indexed within each lateral starting
with the first bus on the lateral, so that each bus is
uniquely identified by an ordered triple ( I , m, n) where R.
is the bus index. So ( I , m, n) refers to the nth bus on the
mth level 1 lateral. The source is given an index of
(1,1 , O ) . Figure 1 shows an example of this indexing
scheme on a sample 63-bus system. The boxed numbers
show the reverse breadth-first (RBF) ordering of the laterals found by sorting the lateral indices in reverse order,
r,,
(2.1.4
(2.1.3
(2.1.2
(2.1.1
wk-l
(2)
= gk(wk)
where wUR
is a 12 x 1vector containing the real and imaginary parts ofthe voltages and currents at bus k. The func-
cogeneratob
shunt1
capacitor
/////////////-
2041
I,' =
IGk+ICk+ILk-Ik+l-
Ij
je
Element
Update Equation
...................................................................................
(4)
(5)
(6)
(3)
line
I 'h-1 =
I I , = -I;
'h
I I, = Y p , - ,
+Y p ,
From (15),we can solve for the voltage and current at the
primary given the voltage and current at the secondary
resulting in (13) and (14).
PROBLEMFORMULATION
Here we present a new load flow formulation with a
reduced number of equations and unknowns. The load flow
problem is typically formulated as a set of non-linear
power mismatch equations as functions of the bus voltages. The number of equations and unknowns in an n-bus
all three-phase system is 6 (n - 1) ,since the source is the
only voltage specified bus in a distribution system. Our
formulation reduces the number of equations and
unknowns to 6 times the number of laterals in the system.
To illustrate our formulation we start with a system consisting of a single main feeder.
Single Feeder
Given the voltage and current at either end of the
feeder we can compute the remaining voltages and currents. Note also that throughout this formulation currents
could be replaced by complex power flows as in [131. We
have two boundary conditions, the current I,,+l at the
end of the feeder is equal to zero, and the voltage V , at
the source is a specified constant Bo. So we can take
l'kansformer Model
Three-phase transformers are modeled by the admittance matrix equivalent proposed in E151 for grounded-we
to grounded-wye connections.
vo = Po
/o+
"-------t-H-]
source
V"
-In+,
end bus
=0
2048
either the current I , at the source, or the voltage V, at
the end of the feeder to be the unknown. Unlike I , , an initial guess to V, is readily available without computation
(balanced 1 P.u.), so we choose V, , which we refer to as
the end voltage, as the independent variable.
= 0 we can compute
Using (2) and the fact that I ,
the source voltage (and current) as a function of V,.
Starting with w , we apply the appropriate branch update
function gk consecutively until we reach w ,
+
bus k = ( I , m,, n)
We will denote the voltagf part of the composite function g , ... g , - g , as Vo . At the solution the mismatch between the specified source voltage V O and the
computed source voltage Vo(Vn) must be zero. The power
flow equations can therefore be written as
f(V,,
V0(V,)-Vo
= 0
(17)
(20)
= Vk(x)-vk(x) = 0
F(x) = 0
We express the system Jacobian for this simple single
feeder case as
where the first term in the product is just the top half of
G , and the last term is the left half of G, ,
General Radial Structure
To generalize this formulation to handle an arbitrary
radial structure we first note that the voltage at the beginning of any lateral can be computed as a function of the
end voltage if the currents injected into each sub-lateral
are given. In a system with L levels, the level L laterals
have no sub-laterals and can therefore be computed first.
After all level L laterals have been computed, the currents
injected into the sub-laterals of each level L - 1 lateral are
known, hence the level L - 1 laterals can be computed.
Next the level L - 2 laterals are computed, and so on,
until the main feeder has been computed. This is the
(21)
Newton Method
2049
4zk.
(18).
Approximation of the branch Jacobian by the identity is
the equivalent of replacing loads, shunt capacitors, and
cogenerators by constant injected currents and neglecting
line charging and line impedance. The same approximation is used for sectionalizing switches. For transformer
sections, the upper left block of G, is approximated by the
identity divided by the tap ratio, and the lower right block
2050
step 5 has been essentially eliminated and step 6 is now a
simple back-substitution. In fact, it is not even necessary
to explicitly form and store the Jacobian since each row
has a 1on the diagonal and a -1in the column mmesponding to the appropriate phase of the parent lateral.
Here again, a 1 is replaced by the reciprocal of a corresponding transformer tap ratio where appropriate.
t.80
Newton
Implicit Z
2.04
Gauss
BacklFoward
sweep
1.04:
69
:
:
Fast Decoslpled
I
RESULTS
c-
The fast decoupled method described above was implemented in Matlab 4 along with the traditional fornulation
of the Newton [4] and implicit Zb,, Gauss methods [5] and
a backKorward sweep method [9]. The implicit z b , Gauss
method actually uses a n optimally ordered, factored Y b a
instead of forming z b , explicitly.
The test systems are 292-bus and 394-bus unbalanced
systems, each with 6 distribution transformers. According
to our formulation, the first system has a total of 85 laterals and the second has 108 laterals, each including some
single and 2-phase laterals. Three of the algorithms,
excluding the backlforward sweep, require the solution of
a large sparse system of linear equations. In the Newton
and fast decoupled methods the resulting matrix is a Jacobian and in the implicit Zba Gauss method it is Yb,.
Table I1 shows the sizes and types of these matrices for our
two test systems. If all buses were three-phase we would
expect the dimension to be approximately 6n for Newton,
3n for implicit z b , Gauss, and 61 for fast decoupled, where
n is the number of buses and I is the number of laterals.
Since there are some single and 2-phase buses the numbers are slightly smaller than these. Notice that in the fast
decoupled method the matrix is triangular, that is, already
in factored form. Unlike the fist two methods, the proposed algorithm does not require the formation of Y b , or
the factoring of a large matrix.
In the case of the 394-bus system, the number of real
equations and unknowns, and therefore the size of the
Jacobian, was reduced from 1996 for the traditional Newton formulation to 560 for the fast decoupled method. This
decrease in size was accompanied by a correspondingly
large decrease in the computation required to solve the
load flow from a flat start. Figure 6 shows the computational complexity for each of the four methods in megaflops. These figures are based on the total number of
floating point operations as reported by Matlab and should
be indicative of run-time for the algorithms in compiled
languages such as C or Fortran.
Our results show that the fast decoupled method is also
more efficient than the bacWforward sweep method based
TypedtStructure
CONCLUSIONS
& FrrmTRE WORK
In this paper we have exploited the radial structure
(physicalproperty) and the decoupling numerical property
of a distribution system to develop a fast decoupled Newton method for solving unbalanced distribution load flow.
It involves a reduced set of equations and unknowns proportional to the number of laterals in the network as
opposed to the number of buses. Due to the reduced number of equations and the fact that the Jacobian is approximated by a constant triangular matrix, it is significantly
faster than the implicit Zb,, Gauss method or the traditional Newton method based on U,,,. Since each function
evaluation involves updating each bus voltage and current
IB Newton
real, general
Newton
Dimension
292bus 394bus
OfMntrix
1472
factoring necessary
736
no matrix necessaw
rea1,triangulaF
426
1996
998
560
megaflops
I
I
Figure 7 Convergence for 394-bus System
205 1
ACKNOWLZDGMENTS
The authors would like to thank Gary Darling and
acknowledge the financial support of New York State Gas
& Electric. The valuable input of Dr.Jianzhong Tong was
also much appreciated.
REFERENCES
C13
[2]
[3]
[4]
[5]
[6]
W. H. Kersting and W. H. Phillips, A Radial Threephase Power Flow Program for the PC, Conference
paper, presented at 1987 Frontiers Power
Conference, Stillwater, OK, October 1987.
2052
DISCUSSION
[Cll W.H. Kersting and W. H. Phillips, A Radial hreephase Power Flow Program for the PC, Conference
presented at 1987 Frontiers Power
Conference, Stillwater, OK, October 1987.
puper,