IEEE Transactions on Power Systems, Vol. 14, No.

3, August 1999
I

1063

RELIABLE LOAD FLOW TECHNIQUE FOR RADIAL

DISTRIBUTION NETWORKS
Antonio G6mez Exp6sito Esther Romero Ramos

Department of Electrical Engineering University of Sevilla Sevilla, Spain

Abstract - This paper presents a method to solve the power flow problem in radial networks. In the proposed formulation, the load flow equations are written in terms of new variables resulting in a set of 3N equations (2N linear plus N quadratic) for a network with N + l buses. A computationally efficient solution scheme based on the Newton-Raphson method is proposed, and possible simplifications are discussed. Finally, a comparative evaluation of the proposed methods performance against existing algorithms is provided. Keywords: Load Flow, Distribution Systems, Radial Networks.

Introduction

Although efficient general-purpose algorithms to solve the load flow problem are available since the early seventies [l, 21, there exist certain ill-conditioned cases for which those methods may present convergence difficulties [3,4]. Radial feeders typically used in distribution systems constitute a well-known example in this respect. Several power flow algorithms specially designed for radial distribution networks have appeared in the literature [5]-[16]. In addition to their improved numerical robustness, these techniques take full advantage of the radial structure of the networks to save computation time. The main ideas behind some of the most representative published algorithms are discussed below. In [6] the authors propose a compensation-based technique for weakly meshed distribution networks. The radial part is solved by a straightforward two-step procedure in which the branch currents are first computed (backward sweep) and then the bus voltages are updated (forward

PE-344-PWRS-0-12-1996 A paper recommended and approved by the IEEE Power System Engineering Committee of the IEEE Power Engineering Society for publication in the IEEE Transactions on Power Systems. Manuscript submitted July 26, 1996; made available for printing December 16, 1996.

sweep). Branch power flows, rather than branch currents, were later used in the improved version presented in [7]. Extensions of this method, with the emphasis on modeling unbalanced loads and dispersed generation have recently been published [12,14]. In [8],as a subroutine within the optimal capacitor sizing problem, a methodology for solving the radial load flow is proposed. For each branch, three non-linear equations are written in terms of the branch power flows and bus voltages. The number of equations is subsequently reduced by using the terminal conditions associated with the main feeder and its laterals, and the Newton-Raphson method (NR) is applied to this reduced set. In order to improve the computational efficiency, some simplifications are made in the jacobian. Numerical properties and convergence rate of this algorithm have been later studied in [ll]. In [9], the author makes use of the well-known biquadratic equation which, for every branch, relates the voltage magnitude at the receiving end t o the voltage at the sending end and the branch power flow. Given an estimate of the power losses, all branch power flows are found starting from the leaf nodes. Then, sweeping the tree in the opposite direction, voltage magnitudes are computed. Bus phase angles do not appear in this formulation which was used also in [13]. Aimed at studying the voltage stability of radial networks, a load flow technique which, for every branch, leads to a pair of quadratic equations relating power flows at both ends with the voltage magnitude at the sending end is proposed in [lo]. Finally, some modificatins to the NR method are explored in [15,16], taking advantage of the radial structure of distribution networks. In [16], a mismatch voltage equation is written for each lateral feeder, yielding a reduced set of equations which are solved by a decoupled methodology. The modified NR technique proposed in [15], on the contrary, makes use of the conventional unreduced set of equations, but the jacobian is not explicitly built. The design of most of those algoritms has been mainly guided by computational efficiency issues. This explains the interest in reducing the number of equations, approximating the jacobian, etc. However, as computers become increasingly powerful, aspects other than computational

0885-8950/99/$10.00 0 1997 IEEE

I064

effort may be more important in choosing a particular methodology. The algorithm presented in this paper follows a different approach aimed at enhancing the convergence rate. It is based on the intuitive idea that the more linear an equation system the better its convergence rate. To reach this goal, the load flow equations are written in terms of alternative variables leading to a set of 3N equations (2N linear plus IY quadratic) for a network with N + l buses. An efficient computational scheme, based on the NR method, is proposed to solve the resulting equation system, and possible simplifications are discussed. Finally, the proposed method is applied to several ill-conditioned networks appeared in the literature and the results are compared with those of some existing algorithms, both in terms of speed of convergence and computation times. Extension of the method to unbalanced (three-phase) networks, containing several PV buses and voltage-dependent loads, will be the subject of a future publication.

Substituting eqs. ( l ) , (2) into (3), (4) gives rise to the 2 N non-linear equations solved by conventional load flow tools. However, using Ui to denote K2, and introducing the following pair of variables per branch,
% j

= qy cosejj

(5)

Iij = V;. 5 sin O j j

(6)

eqs. (I), (2) appear as a system of 2N linear equations in 3N unknowns (2 per branch, 1 per node),

p.. - G..U. - G 13 . . &3 . + B..I.. 83 - 1 1 . 1 $3

Qt.3.

(7)
(8)

- B?h)Ui - B $3 . . Ra3 . . - G '3 . . Ir.i - (B.. 23 2 3

In order to have enough equations, the following N

quadratic constraints must be also added

UjUj = Raj + I2. '3

,Problem Formulation

(9)

Consider the line model and notation shown in figure 1 (note that, because of the minus sign in the series line susceptance, all line constants are positive).

Note that, when counting the unknowns, it should be = Rji and Iij = -Ijj. kept in mind that In summary, substituting eqs. (7) and (8) into (3) and (4), we have the following set of equations

: which, representing the branch and nodal unknowns by 1 and U respectively and the specified powers by p , can be written in compact form as

Figure 1: Line model and notation adopted Th.e power flows leaving the sending node are given by the following expressions, Az+Bu=p f(1:,u) = 0 (13) (14)

Qij

= (B..-BSF)K'-BijKQ '3
13

cos6ij-GijKQsinOij

(2)

Note that A is a non-singular, 2N x 2N matrix while B is rectangular (2N x N). Once the adopted unknowns are found, the conventional nodal variables can be obtained from,

By adding the flows for all nodes j adjacent to i ( j E i ) , the respective power injections are obtained,

Qi

Qij

(4)

Finding the individual nodal phase angles, which is seldom required, involves sweeping the tree branches from the root.

1065

Solution Methodology

Applying the NR method t o the resulting system, the following equations must be solved at iteration k,

AAz

+ BAu = p - ( A z + ~ B u ~ ) Apk F:A~ +F ~ A U = -f(tk, u t )

1

(15) (16)

where F, = [%I, F,, = It should be stressed that, since (13) is a linear system, the right-hand side of eq. (15) is always null after the first iteration. This means that the power mismatch vectors have to be computed at most once, the convergence being dictated by the largest component of f(tk, uk). There are at least two ways of initializing the itera; = 0 and U: = 1 tive process: (A) Setting Rrj = 1, 1 ( U t = V;Llaek is also possible), which is equivalent t o the usual flat start (6: = 0, = 1). In this case, f(x,, u o )= 0 and Ap, = p (in absence of shunt elements). Therefore, no computations are required t o obtain the right-hand side vector in (15)-(16). (B) Setting U t = 1 and obtaining Rtj, 1; from (13), i.e., I" = A-'(p - B u ) . In this case, Ap = 0 from the beginning, but f(x,,u,) has to be computed. Scheme (A) has been finally adopted, according to experimental results clearly showing its superiority. The iterative process is then composed of the following steps:

[E].

k = 0, ~ ( z , , u , ) = 0, Ap, = p (a term should be eventually added t o account for shunt elements).

Find F:, F," and solve (15), (16) for A x , Au.

The upper part corresponds to the 2N linear equations (15) and, therefore, it is constant, while the lower part arises from eq. (16) and should be recomputed at each iteration. Note that the upper 2Nx2N diagonal matrix is lower triangular in terms of 2x2 blocks, and that the lower NxN diagonal matrix is upper triangular. Although this is the most natural and straightforward way of arranging the equations, the fill-in which appears in the lower diagonal block when the first 2N equations are eliminated discourages its use for very large networks (this example is not large enough to illustrate this problem). The situation is even worse if the two sets of equations are interchanged, that is, if the unknowns corresponding to eq. (16) are eliminated first. The solution adopted to circumvent this loss-of-sparsity problem consists in grouping the 2 linear equations for each branch with its corresponding linearized equation for the sending bus, as follows (fill-in elements are also shown as @),

k = IC + 1. Update ck,uk.
Compute f(xk,uk). If max I f i j 1 5 E , then stop. Otherwise, set Apk = 0 and go t o step 2.

K K

K K

K K

x
K K K K K K

1
K K K

Clearly, from a computational point of view, step 2) is the most relevant. To illustrate how eqs. (15), (16) can be efficiently solved, the 5-bus network shown in figure 2 will be used.

Figure 2: 5-Bus sample system Adopting the well-known node ordering strategy for radial networks (i.e., from the leaf nodes to the root), the resulting jacobian structure is,

With this arrangement, thinking in terms of 3x3 blocks, fill-in elements are restricted to already existing blocks, the tree structure being preserved both in the upper and lower triangular factors (note that branches incident to the root bus do not appear as off-diagonal blocks). Although the dimension of the resulting equation system is larger than with existing algorithms, the overall computational effort can be significantly reduced if a customized code, taking advantage of the problem structure and using efficient data structures, is written. Without going into the details, whose description would be quite cumbersome, the solution procedure is based on the following basic guidelines,
0

Not all the elements of the jacobian are explicitly built. Rather, the tree structure is made use of throughout the whole solution process, which is

1066

equivalent to the well-known forward/backward tree sweeps.

a

Method
I

10-2

Constant elements are never recomputed. Furthermore, irrespective of a particular element being cons8tantor varying, off-line exhaustive algebraic manipulations/simplifications are carried out in order to reduce, as much as possible, the total number of computations per iteration.

Convergence Threshold II 10-3 I 10-4 II 10-5 I io-6

I

A straightforward simplification results from the jacobian being constant after the first or second iteration, although a certain degree of convergence deterioration should be expected in such a case.

aMaximum No. of iterations exceeded (50)

Table 1: Convergence rate for the 85-bus network.

Experimental Results
Method io-2 Convergence Threshold I lo- I 1 0 - ~ I 10-~ I 2 1 3 4 1 5 2 1 3 M I M 3 5 4 4
I

The proposed algorithm has been coded in FORTRAN and compared with the following algorithms: Conventional NR-based power flow [l], Luo-Semlyen [6,7], CCspedes [9], fjaran-Wu [8] and Zhang-Cheng [15]. Several illconditioned radial networks have been gathered from the literature to compare the algorithms. Their sizes are: 30 buses [9], 43 buses [3], 69 buses [17] and 85 buses [13]. The main difficulty in comparing the convergence rate of those algorithms lies in the fact that different convergence criteria are used, depending on the state variables adopted. Luo-Semlyens and Zhang-Chengs methods use the conventional power mismatch vectors to stop the iterative process. CCspedes compares the total power losses at two consecutive iterations, and Baran & Wu examine the powers at the terminal buses, which approach zero as iterations progress. As far as the proposed method is concerned, the squared voltage mismatches given by eq. (12) are evaluated since, as was discussed earlier, power mismatches are null after the first iteration. Tables 1 to 4 show, for the 4 tested networks and using single-precision arithmetic, the number of iterations required to achieve convergence when the convergence threshold ranges from to p.u. (such an accuracy is not needed in practice, but it is useful in this context to show the robustness of some methods). The rows labeled Constant-11 and Constant-21 refer to the proposed method when the jacobian is kept constant after the first and second iteration respectively. Except for the 43-bus network, which is specially illconditioned, the proposed method requires only 3 iterations to reduce the mismatch vector components to Baran-Wus and CCspedes methods present also a good behavior, but the former diverges and the latter requires a high number of iterations for the 43-bus network. Keeping the jacobian constant in the proposed method after the second iteration, and sometimes just after the first, does not significantly deteriorate the convergence, the 43-bus network being again an exception. Not surprisingly, the convergence of all algorithms, including the conventional NR method, is dramatically im-

Proposed 1 1 2 Constant-11 1 1 2 Constant-21 I 1 1 1 1 2 1 Conventional NR I M I M 1 M I Baran & Wu 2 2 3 CCspedes 2 3 3 ~UokSemlyen M j M j M j Zhane: & Chene: I M I M I M I

M M

I M

Table 2: Convergence rate for the 69-bus network.

proved in most cases if double-precision arithmetic is employed. A representative example of this good behavior is shown in table 5 for the 69-bus network. However, as can be seen in table 6, some algorithms still have convergence difficulties with the 43-bus network (compare tables 5 and 6 with tables 2 and 3 respectively). Another important aspect when comparing different methods refers to the computational cost. This is of particular interest in this case considering the enlarged model adopted by the proposed algorithm. Of course, the computational time required by any algorithm is strongly affected by the efficiency of the code, and every effort should be made in order to use the best programming techniques. Fortunately, for radial networks, the tree structure provides a common framework from which efficient data structures are generated for every algorithm. Table 7 shows, for every tested method and network, the time required per iteration on an IBM RISC/SOOO machine. Although initialization time and other overheads should also be taken into account, the time spent on the whole iterative process is the most, important. The following comments are in order:
0

The general purpose NR method is the most expensive, but it can run on meshed networks (no advantage is taken of the network radial structure)

1067

I
~~

Method Proposed Constant- 1I Constant-21 Conventional NR Baran & Wu Cbpedes Luo & Semlyen Zhane: & Chencr
~~

in-" _

Convergence Threshold I _. I 1w5 Iiw _ - b 1

Method
10-2

Convergence Threshold I io-" I 10-4 I 10-5 I 10-6

5 18 6 17 22

6
34 9 28 34

1a 6

I I

I 45 I M l M l M

I I

40 M

I I

49 M

"Diverged

Table 3: Convergence rate for the 43-bus network. Method Proposed Constant- 1 1 Constant-21 Conventional NR Baran & Wu CCspedes Luo & Semlyen Zhang & Cheng Convergence Threshold io-" 1 0 - ~ 1 0 - ~ io-b: 2 3 3 3 3 4 6 7 2 3 3 4 3 49 M M 2 3 4 5 2 M M M 3 M M M 4 6 M M

Table 5: Convergence rate for the 69-bus network (double precision). Method ProDosed Constant-11 Conventional NR Baran & Wu Chedes Luo & Semlyen Zhang & Cheng
A

Convergence Threshold 10-3 I 10-4 I 10-5 I 10-6 5 1 6 1 6 1 7

22

34

Table 4: Convergence rate for the 30-bus network. Computation times required by Bar an-Wu 's met hod strongly depend on the number of lateral feeders. For the tested networks, these times are similar or slightly better than those of Cbspedes' and Luo-Semlyen's approaches. The proposed method is not so expensive as might be thought a priori. This is mainly due to the fact that no power mismatches have to be computed. When a constant jacobian is adopted, computation times are comparable to the fastest algorithms. In general, it can be stated that all computation times are of the same order of magnitude and, consequently, choice of a particular algorithm should be mainly dictated by its reliability in solving difficult cases.

Table 6: Convergence rate for the 43-bus network (double precision). method is then applied to these equations, and the resulting linear system is rearranged so that its block structure resembles the network tree. This jacobian structure, along with the fact that power mismatches are zero after the first iteration, yields a computationally competitive methodology, particularly when the jacobian is not recomputed. The proposed technique has been compared with some of the most representative algorithms appeared in the literature. When applied t o several ill-conditioned networks, it has proved t o be more reliable than existing methods. Future efforts will be directed to include voltage-dependent and unbalanced loads in the model.

Acknowledgements

Conclusions

The second author acknowledges the financial support provided by Iberdrola under grant no. 95/193T.

In this paper an algorithm to solve the load flow problem for radial networks has been presented. Instead of working with the conventional state variables (complex bus voltages), it resorts to a pair of branch variables plus a bus voltage variable. This leads to a set of 3N equations, of which the 2N related to power injections are linear, and the remaining N related to bus voltages are quadratic. The NR

References
[l] W. F. Tinney, and C. E. Hart, "Power Flow Solution by Newton's Method", IEEE Transactions on Power Apparatus ans Systems, Vol. PAS-86, pp. 1449-1460, 1967.

1068

Network (buses) 85 1 69 43 I 30

H.-D. Chiang, A Decoupled Load Flow Methos for Distribution Power Networks: Algorithms, Analysis and Convergency Study, Electrical Power & Energy Systems, Vol. 13. No. 3, pp. 130-138, June 1991. T.-H. Chen, et al., Distribution System Power Flow Analysis- A Rigid Approach, IEEE Transactions on Power Delivery, Vol. 6. No. 3, pp. 1146-1152, July 1991. D. Das, D.P. Kothari, A. Kalam, Simple and Efficient Method for Load Flow Solution of Radial Distribution Networks, Electrical Power & Energy Systems, Vol. 17, No.5, pp. 335-346, 1995. C.S. Cheng, D. Shirmohammadi, A Three-phase Power Flow Method for Real-Time Distribution System Analysis, IEEE Transactions on Power Systems, Vol. 10. No. 2, pp. 671-679, May 1995. F. Zhang, C.S. Cheng, A Modified Newton Method for Radial Distribution System Power Flow Analysis, Winter Meeting IEEE/PES, January 21-25 1996. R. Zimmerman, H.-D. Chiang, Fast Decoupled Power Flow for Unbalanced Radial Distribution Systems, IEEE Transactions on Power Systems, Vol. 10, No. 4, pp. 2045-2052, November 1995. M.E. Baran, F.F. Wu, Optimal Capacitor Placement on Radial Distribution Systems, IEEE Transactions on Power Delivery, Vol. 4. No. 1, pp. 725-734, January 1989.

CCspedes Luo & Semlyen Zhang & Cheng

1.14 1.38 2.67

1.05 .9 2.24

.64 .68 1.36

.47 .51 .89

Table 7: Computation times per iteration on an IBM RISC/6000 workstation (msec.).

H . Stott, and 0. Alsac, Fast Decoupled Load-Flow, IEEE Transactions on Power Apparatus ans Systems, Vol. PAS-93, pp. 859-869, 1974.
S.C. Tripathy, G.G. Prasad. O.P. Malik, G.S. Hope, Load-Flow Solutions for 111-Conditioned Power Systems by a N ewton-Like Method, IEEE Transactions on Power Apparatus ans Systems, Vol. PAS-101, No. 10, pp. 3648-3657, October 1982. D. RajiEiC, A. Bose, A Modification to the Fast Decoupled Power Flow for Networks whith High R/X Ratios, IEEE Transactions on Power Systems, Vol. 3. No. 2, pp. 743-746, May 1988. W.H. Kersting, A Method to Teach The Design and Operation of a Distribution System, IEEE Transactions on Power Apparatus and Systems, Vol. PAS-103. NO. 7, pp. 1945-1952, July 1984.

1). Shirmohammadi, H.W. Hong, A. Semlyen, G.X.

Luo, A Compensation-Based Power Flow Method for Weakly Meshed Distribution and Transmission Networks, IEEE Transactions on Power Systems, Vol. 3. No. 2, pp. 753-762, May 1988.

G.X. Luo, A. Semlyen, Efficient Load Flow for Large Weakly Meshed Networks, IEEE Transactions on Fower Systems, Vol. 5. No. 4, pp. 1309-1316, November 1990.
A I . E. Baran, F . F. Wu, Optimal Sizing of Capacitors Placed on a Radial Distribution System, IEEE Transactions on Power Delivery, Vol. 4. No. 1, pp. 735-743, January 1989.

Antonio G6mez Exp6sito was born in AndGjar, Spain, in 1957. He received his electrical engineering and doctor engineering degrees from the University of Sevilla. Since 1982 he has been with the Department of Electrical Engineering, University of Sevilla, where he is currently a Professor. From 1991 to 1994 he was Head of the Department. His primary areas of interest are sparse matrices, parallel computation, reactive power optimization, state estimation and computer relaying. Esther Romero Ramos was born in Palma del Rio, Spain, iii 1967. She received her electrical engineering degree in 1992. From 1992 to 1993 she worked for Sainco. Since 1993 she has been with the Department of Electrical Engineering, University of Sevilla, where she is currently a Lecturer. She is interested in State Estimation and Load Flow problems.

Fe. CCspedes G., New Method for the Analysis of Distribution Networks, IEEE Transactions on Power Delivery, Vo1.5, No. 1, pp. 391-396, January 1990.
G.B. Jasmon, L.H.C. Lee, Distribution Network Reduction for Voltage Stability Analysis and Loadflow Chlculations , Electrical Power & Energy Systems, Vol. 13. No. 1, pp. 9-13, February 1991.

1069

Discussion
Ali Abur (Texas A&M University, CS, TX 77843) I wish to congratulate the authors for this interesting paper. As authors state in the Introduction, conventional load flow methods developed essentially for solving transmission networks may encounter convergence problems when a.pplied to distribution networks. In this discusser's opinion, this poor behavior can be essentially attributed to the jacobian ill-conditioning caused by the existence of very short branch sections along with longer ones, rather than to the non-linear nature of the resulting equations. While it is clear that the proposed formulation helps to avoid the latter aspect, it is not as apparent how the ill-conditioning problem will also be alleviated. For instance, using the proposed method, the jacobian condition number for the simple 3-bus network shown below, approaches lo7 (flat start). It would be appreciated if the authors could clarify why their proposed algorithm's convergence pattern would remain robust using single precision arithmetic even for this ill-conditioned example system. 3 b 2 a 1

elements. In our 3-bus example, the block jacobian reduces to:

1
and the only block elimination is
A& = A22 - AzIA;,'AIz

-2

It is easy to obtain symbolically that

2G, 2B,

-A,

where A, = G: B," (actually, due to the sparsity of A12, only : / would be needed). Proceeding in this the last column of A way, we also obtain:

Az1A;,'A12 =
Y, = 1000 - lO0Oj
yb=l-j

and, finally,

:]

Manuscript received June 12, 1997.

Antonio G6mez Exp6sito and Esther Romero Ramos. We thank the discusser for his timely question which gives us the opportunity to better explain certain implementation details that could not be properly addressed when writing the paper. Although not explicitly stated in the paper, the resulting equation system is solved by means of block arithmetic. That is, 3x3 block pivots, rather than scalars, are inverted and used to update the appropriate diagonal blocks below. Actually, the fill-in elements shown as @ in the blocked matrix of section 3 are misleading and should be ignored, since they only appear when scalar arithmetic is used. As the blocked matrix structure is that of a tree, and the minimum degree ordering is adopted, only diagonal blocks get modified when the following kernel is repeatedly applied:

Consequently, the condition number of individual block pivots, rather than the conventional condition number of the matrix as a whole, should be considered. In the example raised by the discusser, the worst block condition number is about lo3, which is not too bad considering the extreme line length ratio employed. Another key factor when implementing the algorithm, in terms of both computational efficiency and numerical robustness, refers to the off-line symbolic manipulation mentioned on the fly at the end of section 3, which is only possible in the radial case. The advantages of symbolically handling eq. ( C l ) can be better understood when the jacobian at the first iteration, i.e., at flat start, is considered in absence of shunt

Hence, we have been able to obtain AkZ without carrying out any arithmetic operation and, what is more important, without accumulating any round-off error. When Ai2 is obtained numerically, we have observed in certain ill-conditioned cases that the 2 upper elements of the last column are not exactly Gb, &. This is due to the fact that, on a computer, the expression (G, Gb) - Gb may differ from Gb if G, >> Gb. In the general case, i.e., at subsequent iterations, the jacobian contains more complex expressions and its symbolic manipulation is not so straightforward, but the conclusion is the same: the computational effort is reduced and potential numerical errors are prevented. In summary, the fact that the algorithm behaves so well when using single precision arithmetic is essentially due to combining block arithmetic with symbolic simplifications. However, in those cases where reliability is more important than speed of response, the use of double precision is strongly recommended. This way, any line length of practical interest can be handled without losing accuracy [A].

[A ] D.J. Tylavsky, P.E. Crouch, L.F. Jarriel, J. Singh, R. Adapa, "The effects of precision and small impedance branches on power flow robustness", IEEE Transactions on Power Systems, Vol. 9, No. 1, pp. 6-14, February 1994. Manuscript received September 23, 1997.