1

A new power flow method for radial networks

Manuel A. Matos, Member, IEEE

Abstract—The need of fast algorithms for radial power V0 V1

distribution networks that take advantage of their particular z1
structure has been increasing, namely due to the use of genetic
algorithms and meta-heuristics for optimization in planning and
operation.
-S1
In this paper, a new method for power flow calculation in
radial networks is presented. It uses an iterative process along
the branches, in a way similar to other methods, but the main Figure 1 - Network with a single branch
idea is very different from previous approaches, since it is based
on the exact power flow solution for a single branch and also Of course, we have:
because it provides a complete solution (not only voltage *
magnitudes). The method is fast and robust for different types of  S 
networks and loads, including heavy loads. V0 − V1 = z 1 . − 1 
 V1 
The paper includes the theoretical derivation of the method
an illustration example and tests with benchmarking networks. or
2
V0* .V1 − V1 + z 1* .S1 = 0 (1)
Index Terms—Power Distribution, Load flow analysis,
Iterative methods, Planning, Operation.
The analytical solution of this complex quadratic equation
I. INTRODUCTION may be found in a number of ways. For instance, changing to
rectangular coordinates and assuming that θ0=0 (so V0=e0),
T he special structure of radial networks has lead, in the
past, to a number of specialized algorithms that tried to
V1=e1+jf1, z1=r1+jx1 and S1=P1+jQ1, we get:
take advantage of the absence of meshes to simplify the
calculations and save memory [1]-[5]. In some cases, the ( )
e 0 (e1 + jf1 ) − e12 + f12 + (r1 − jx 1 )(P1 + jQ1 ) = 0
methods are extended to weekly meshed networks with some
success. and, after separating real and imaginary parts:
Calculation of the power flow in radial networks was not a
priority in the past, since approximate methods were sufficient ( )
e 0 e1 − e12 + f12 + (r1 .P1 + x 1 .Q1 ) = 0 (2a)
to have a general picture of the power flow and, if necessary,
a general-purpose method (like Newton-Raphson) could e 0 f1 + (r1 .Q1 − x 1 .P1 ) = 0 (2b)
always be used. Use of genetic algorithms and meta-heuristics
in the optimization of distribution networks, however, lead to Now, from (2b) we take immediately the value of f1, since
the need of fast calculation methods with some degree of all the other quantities are known, and then we find e1 by
accuracy [6]. On the other hand, more and more DMS substituting f1 in (2a) and solving the (real) quadratic
(Distribution Management Systems) [6] are being developed equation, where e1 will be the biggest solution. This analytical
and installed, and fast methods for radial networks are again solution of the power flow problem is well known, but
welcomed. Finally, dispersed generation connected to generally not used, since it is only applicable to the trivial case
distribution networks is growing, and adequate algorithms are of two buses.
needed to deal with it.

II. FOUNDATION OF THE METHOD

V0 V1 V2 Vn-1 Vn
The main idea of the method is to use the exact power flow z1 z2 zn
solution for one branch, when we know the voltage in the
sending end (V0) and the injected power in the receiving end -S1 -S2 -Sn-1 -Sn
(S1), z1 being the impedance of branch 0-1 (see Figure 1).
Figure 2 – Simple radial network model
M. A. Matos ([email protected]) is with INESC Porto – Instituto de
Engenharia de Sistemas e Computadores do Porto, Campus da FEUP, Rua Dr.
If we have now a radial network with successive branches
Roberto Frias, nº 378, 4200-465 Porto, Portugal. Phone: 22 2094230 Fax: 22 1, 2,..n, where node 0 is the root, with a specified voltage
2094050 V0=e0 (see Figure 2).
 2

In the first branch, (1) transforms to: Vpred(i) Vi

zi
*
 n Sk 
V0 − V1 = z1 . − ∑ 

 k =1 Vk  -S’i

or, multiplying by V1 and conjugating: Figure 3 – General branch model

With the proposed formulation, it is not necessary to
n
V  estimate initial values for the voltages, but only to consider
∑
2
e0 .V1 − V1 + z1* . S k . 1  = 0 (3)
k =1  Vk  that:
Vk( 0) = Vi( 0) all k∈succ(i) (7)
Similar expressions may be established for the remaining
nodes, in each case using the predecessor node voltage as a This corresponds to using, in the first iteration,
constant. The general expression (i=1..n) is therefore:
 
+ z *i . S i + ∑ =0
* 2
 V  
V pred ( i ) .Vi − Vi Sk
 
*
∑
* 2
V pred ( i ) .Vi − Vi + zi . Si + Sk . i  = 0 (4)  k ∈succ ( i) 
  Vk  
 k ∈succ( i )
instead of (6). Of course, if good initial values are known,
The idea is then to apply (4), beginning in the first node they may always be used with the normal version of the
after the root - which corresponds to (3) – in order to equation.
successively calculate the voltages, until getting the leaves of
B. Iterative process
the tree. This corresponds to an iteration that can be repeated
with the updated values of the voltages, until some As mentioned before, the method progresses, in each
convergence criterion is met. iteration, from the root to the leaves, with successive use of
(6). The updated values of Vi are immediately used in their
III. CALCULATION PROCESS DETAILS successors’ equations.
The sequence of calculations is very straightforward and
A. Equations similar to other forward sweep methods, so we’ll only sketch
In order to conduct the iterative process, (4) is conveniently it with the help of Figure 4.
transformed to:
0 1 2 3 4

  V ( p −1 ) 
  = 0
∑ S k . i( p −1 )
*( p ) 2 i
V pred ( i ) .Vi − Vi + z *i . S i +  
(5)
 k∈succ( i )  V
  k  5 6

where (p) denotes the iteration. Once it’s clear that the
voltage of the predecessor node is always known when we Figure 4 - Example distribution network
calculate the updated value of Vi, we may write, simply (the
meaning of S’i is obvious): Since node 0 is the root (V0 is known), calculation of V1 is
the first step of an iteration. In the second step, V2 and V5 may
* 2
+ z *i .S' i = 0 be calculated independently. V3 and V6 will follow, and V4
Vpred ( i ) .Vi − Vi (6)
will be updated in the final step of the iteration. This shows
the possibility of partial parallel calculation in typical
corresponding to the model of Figure 3. As mentioned
distribution networks, with important time savings.
before, a formulation in rectangular coordinates is the best
The updated values of the voltages are then used in a new
way to solve the equation in order to get Vi, avoiding the need
top-down iteration, and the process is repeated until a
for trigonometric calculations with small angles.
specified tolerance on the voltages' successive values is met.
Note that other modified forms of (4) could be used, but
A final convergence test on the specified injected power is
our results shown that they are less efficient than (5).
recommended.
C. Branch model
Up to this point, a simple model for the branches has been
used, considering only the branch impedance, which is
common in distribution networks. However, a more detailed
model can be used if necessary, with minor changes in the
 3

equations. This feature may be important in underground IV. SOME ENHANCEMENTS

networks, where the capacitance of the cables is not
negligible. A. PV nodes
Although the method is primary intended to deal with PQ
Vpred(i) Vi (or impedance) nodes, it possible to consider PV nodes as
zi well. The existence of a PV node affects the calculations of
itself and of all its predecessors. Regarding the first issue, we
start by rearranging (4) in order to isolate the constant terms
ysi ysi -S’i (note that, to deal with the more general equation (9), a similar
process is possible):
Figure 5 - Detailed branch model
V 
∑
* * 2
V pred ( i ) .Vi + z i .S i − Vi + z *i . S k . i  = 0
Figure 5 shows the typical π model for a branch, where ysi  Vk 
k∈succ( i )
is the semi-admittance of the branch. The other variables are
the same of Figure 3. Because there is more than one branch
V 
connected to each node, it is now convenient to define a
∑
* * * 2
V pred ( i ) .Vi + z i . jQi + z i .Pi − Vi + z *i . S k . i  = 0
branch admittance for each node i: k∈succ( i )  Vk 

Y si = y si + ∑ y sk (8) *
V pred *
( i ) .Vi + z i . jQi + C = 0 (10)
k∈scon( i )

where C is a complex constant with obvious meaning.

where scon(i) is the set of successors connected to node i.
Now, we must solve (10) to calculate Vi and Qi. The best way
In this case, the total current through branch i includes terms
is to eliminate Qi from the two real equations that result from
related to all the admittances in i and in its succesors:
(10) and then use the fact that we know |Vi| to obtain the real
and imaginary parts of Vi. The, using (10) again, we’ll obtain
  S' *  Qi.
V pred ( i ) − Vi = z i .  − i  + Y si .Vi + Y sk .Vk  ∑
 Vi   For control purposes, it is convenient to calculate also the
 k∈succ( i )  generated reactive power, using:
and (6) transforms into
QiG = Qi + Qiload
 * 
V

*
pred ( i ) − z i . ∑ 
(
Y sk* .Vk*  .Vi − 1 + z *i .Y si* ⋅ Vi ) 2
+ z *i .S' i = 0 If QiG is outside its limits, the adequate limit must be used
 k∈succ( i )  instead, while the bus is temporarily classified as a PQ bus.
The values of Qi obtained in the process will be used in the
In order to maintain the structure of the equation, and since next iteration, in the calculation of the predecessors of node i.
1 + z *i .Y si* is a constant, we may now write: However, this is not sufficient for the algorithm to work, since
in the first calculation of a node, we need the values of Qk for
*
V pred *
( i ) − zi . ∑ Y sk* .Vk* z *i
all its successors. We may get ahead of this problem by using
Q k(0 ) = 0 as initial values for all PV nodes or, if non-trivial
k∈succ( i ) 2
* *
.Vi − Vi + .S' i = 0 (9)
1+ z i .Y si 1 + z *i .Y si* initial values are available for the Vk, by using (10) from the
leaves to the root (that is, in the opposite direction of the
With this formulation, the resolution process in rectangular normal algorithm) before beginning the iteration process.
coordinates is not altered. Obviously, if all the Ys are We summarize now the inclusion of PV nodes in the
negligible, (9) reduces to (6). general algorithm:
a) Estimate initial values for Qk of all the PV buses;
D. Node admittances
b) When reaching a PV node in the iterative process, use
It is also easy to include constant node admittances Yi, i.e., (10) to obtain Vi and Qi);
capacitors or reactors connected to node i, or loads c) Save the value of Qi to be used in the next iteration.
represented by a constant impedance. In fact, it is sufficient to
change (8) in order to include Yi, and then use (9) as the B. Dispersed generation
general expression of the algorithm: Dispersed generation is now frequent in distribution
networks, namely by means of asynchronous machines.
Y si = y si + ∑y sk + Yi (8’) Modeling these nodes as traditional PQ or PV nodes has been
tried, but that approach doesn’t capture correctly the behavior
k∈scon( i )
 4

of asynchronous generators [7]. [8] J. C. Pidré, J. M. Velasco, J.P. Lopes, F. M. Barbosa, “Modeling of non-
linear nodal admittances in load-flow analysis”, Proc. IFAC Symposium
We’ll implement here the idea of “PX node”, developed in on Power Plants and Systems, Munich, March 1992.
[7] and [8]. Briefly, the generated reactive power of one such
unit can be approximated by QiG = − Vi 2 X i (negative, since VIII. BIOGRAPHY
the machine actually gets reactive power from the network), Manuel A. Matos (El. Eng., Ph.D., Aggregation)
with X i = f (Vi ) . was born in 1955 in Porto (Portugal). He is presently
Full Professor at the Faculty of Engineering of the
In order to integrate this kind of node in our method, we University of Porto, Portugal, and the Manager of the
may use tabulated values for X i = f (Vi ) , or a simple model Power Systems Unit of INESC Porto. He also
collaborates with the Management School of the
like X i = (2 − Vi ) X i0 , where X i0 is the value of the University of Porto. His research interests include
fuzzy modeling of power systems, optimization and
magnetizing reactance at nominal voltage [8]. In any case, decision-aid methods. He is a member of IEEE.
we’ll use |Vi| to get Xi and then include it in the expression of
Ysi:
1
Y si = y si + ∑ y sk + Yi − j (8”)
k∈scon( i ) Xi

In this case, (8”) must be updated for all the PX nodes in

each iteration (for the remaining nodes, it is a constant). On
the other hand, if a battery is installed in the node, as it usually
happens, it may be included as Yi in (8”), as explained before.
No other changes in the algorithm are necessary.

V. ILLUSTRATION EXAMPLE
The performance of the method was tested with case
studies found in the literature. The results for two of them are
reported here: a 12-bus network [1] and a 85-bus rural
network [2]. In both cases, detailed data can be found in the
references and will not be repeated here.

VI. ACKNOWLEDGMENT
The author would like to thank students Agostinho Sousa,
and Nuno Ribeiro, both from the Faculty of Engineering of
the University of Porto. Agostinho developed a test program
for the new method, and Nuno helped in the comparison
studies.

VII. REFERENCES
[1] D. Das, H. S. Nagi and D. P. Kothari, "Novel method for solving radial
distribution networks", IEE Proc.Gen. Transm. Distrib., Vol. 141, pp.
291-298, July 1994.
[2] D. Das, D. P. Kothari and A. Kalam, "Simple and efficient method for
load flow solution of radial distribution networks", Electric Power &
Energy Systems, Vol. 17, pp. 335-346, 1995.
[3] A. G. Expósito and E. R. Ramos, “Reliable Load Flow Technique for
Radial Distribution Networks”, IEEE Trans. Power Systems, Vol. 14,
No. 3, pp. 1063-1069, August 1999.
[4] S. F. Mekhamer, S. A. Soliman, M. A. Mostafa, M. E. El-Hawary, “Load
Flow Solution of Radial Distribution Feeders: A new Approach”, in
Proc. 2001 IEEE Porto Power Tech., Vol.3 , Porto, September 2001.
[5] A. Augugliaro, L. Dusonchet, M. G. Ippolito, E. R. Sanseverino, “An
efficient iterative method for load-flow solution in radial distribution
networks”, in Proc. 2001 IEEE Porto Power Tech., Vol.3 , Porto,
September 2001.
[6] V. Miranda, M. Matos, J. P. Lopes J. T. Saraiva, J. N. Fidalgo, M. T.
Ponce de Leão, “Intelligent tools in a real-world DMS environment”,
Proc. IEEE Power Engineering Society Summer Meeting, Vol: 1, pp:
163-168, 2000.
[7] J.P. Lopes, F. M. Barbosa, J. C. Pidré, “Operation simulation of MV
distribution networks with asynchronous local generation sources”,
Proc. MELECON’91, XXXXXX, May 1991.