Mesh grid structure vs.

radial structure
Performance and perspectives of evolution

Tiago Manuel Ferreira dos Santos

[email protected]

Instituto Superior Técnico, Lisboa, Portugal

October 2013

Abstract
In this article several algorithms are studied with the objective of transforming a mesh grid in to
a radial grid. This is achieved through the process of opening meshes by successively disconnecting
lines between buses. Multiple criteria are analyzed for the removal of the connections based on their
expected impact on Joule’s losses in lines and transformers. Once chosen the best criteria for removing
connections between buses, analysis of various search algorithms are carried out to obtain the radial
grid with the lowest possible losses. To achieve this goal various Greedy approaches are tested together
with other more complex algorithms like the Branch and Bound algorithm.

Keywords::Electrical grids, Power-flow, Network optimization, Losses minimization

1. Introduction Engineering of the Washington University[3].

The liability of an electric grid is highly affected The 14 bus grid is composed by 20 connections
by lightning strikes, and a way to assure the liabil- and 5 generators, the 30 bus by 41 connections and
ity of a grid is by applying a safety criteria usually 6 generators, the 57 bus by 60 connections and 7
known as n-1. N-1 criteria simply means that in generators and, finally, the 118 bus grid has 186
case of disconnection of one line, the system con- connections that are powered by 57 generators. All
tinues to function normally. On another hand, this the grids have transformers with a fixed value of
also means that is required a major investment on nominal turns ratio.
design and construction of the electrical grid.
3. Description of the algorithms
There are two main topologies of structures for
In this section, it is done a brief description of
defining the layout of a distribution or transmission
how the used algorithms works. All the power-flow
grid, the radial grid topology and the mesh grid
of the grids are done using the Matpower [4].
topology.
The Matpower [4] is a package of Matlab, made
The radial grid is a tree shape topology where
available free on the site included on the bibliogra-
do not exist close loops. This means that you start
phy, as well as the manual and instructions of how
on one bus and deliver power to the next without
to use it and install it.
the possibility of finding the original bus, except
if you turn backward. This kind of topology is the
3.1. Greedy algorithm
simplest and cheapest topology for an electrical grid
The point of this algorithm is to choose the best
but, with this topology, if a line is disconnected
criteria to remove the lines. There are three criteria:
for some reason, all the lines downstream also lose
power. ˆ type = 1 choose the connection with less cur-
On another hand, in the mesh grid topology it rent
is possible to find close loops and the power is de-
livered through multiples lines, connected to each ˆ type = 2 choose the connection with lower
other, making a mesh. This kind of topology is losses
more reliable, have less losses but it means bigger
investments. ˆ type = 3 choose the connection with lower volt-
age drop between the bus
2. Description of the grids used
The grids were provided by the ’Ieee’, and can It starts by choosing and loading the data of the
be found on the site of the department of Electrical grid on which it will be run the algorithm. Then

1
it will be constructed the matrix of adjacency that gorithm to a branch and bound algorithm approach.
is defined by a vector of vector. Here, it will be To make that change, it is considered that the
defined the bus of the connections that can be re- maximum value of losses that is allowed at any given
moved without isolating some bus. When all the time, is the value of the losses of the best radial grid
elements of the first vector are empty that means found until that moment. As a result of this, ev-
that the grid is radial. ery time the algorithm removes a connection, the
Next, the algorithm starts a loop where the con- value of the losses has to be recalculated. In case
nections are removed in the matrix, one by one the value of the losses is higher than the best ra-
according to the criteria chosen, followed up by a dial stored, that grid is discarded and the algorithm
power-flow, using the Newton-Raphson method, of tries to remove the next one. As a consequence of
the resulting grid. In case of the method do not con- this approach, this algorithm is very identical to the
verge, it will be necessary to step back and choose “greedy n” algorithm 3.2 with the twist that now
another line keeping the criteria. This loop is in- the only children accepted are the ones with lower
terrupted once all the index of the first vector are losses that the best radial grid found until the mo-
empty, on that particular point, the grid as become ment.
radial. Every time this algorithm found a radial grid with
Once the loop is finished, the radial grid is finallylower losses that the one that was stored, it will
obtained and all that is left is calculating losses and update with the stored one with the new one.
making the graphics. At this time, it also was introduced the concept
of maximum current supported by the lines/cables
3.2. “Greedy n” algorithm and verify if all the connections of the radial grid are
By analyzing the results of Greedy algorithm 3.1 bellow that value. Unfortunately, that data about
it was clear that the best criteria is to remove the the lines could not be found so, instead, it was cal-
connections with less current. culated the current of the original grid and multi-
The point of this algorithm is to construct a tree plied the maximum value of all by 1.3. The point of
and realize a DFS (Depth-first search) with the pur- this is to simulate a limitation that real lines/cables
pose of finding the radial grid with less losses. The have of supporting only a certain amount of current
root of the tree is the original grid, and the chil- before permanent damage occurs.
dren are made by taking the grid and removing
the connections with less current, followed up by 3.4. “Side subst” algorithm
a power-flow. In case of failure of the power-flow This algorithm tries to implement the concept
in all the children, the father is wrong and there- of the search algorithm Best-first search [5] to the
fore must be discarded. With this procedure, the problem at hand, the algorithm is based on starting
leafs of the tree are composed only by radial grids. with a node already promising, the radial grid ob-
As the memory was limited, the tree is constructed tained by the classic method, and expand its varia-
at the same time that is required by the DFS, this tions until you find the desired node. This approach
assure that the tree will occupy less memory. make possible to move only between the leafs of the
Once the DFS finds a radial grid it compares it tree, without the need to move up and down the
with the other grids stored in a vector and, in case it tree, making this algorithm much faster to find the
is different from them, the grid is stored in the same best grid.
vector along with the path of removed connections Note that all the radial grids obtained by others
that was used for finding it. When the DFS finishes, algorithms have the same number of connections, so
the value of the losses is calculated for all the grids this algorithm only have to reconnect one line/cable
in that same vector and chosen the one with less and try to remove others parallel to that one.
losses. The algorithm starts by put on index 1 of a vector
Finally, the original grid is read again and, with the radial grid obtained from the algorithm greedy
the information of the path of the connections re- 3.1 with the criteria 1. The next step is to reconnect
moved, the sequence is again performed but this the first connection that was removed to obtain the
time the losses are calculated after every removed radial grid, make the respective matrix of adjacency
connection, so that the graphics can be created. and try to remove another connection instead. If
the resulting grid have lower losses that the starting
3.3. Branch and bound algorithm grid, then that grid is added to the vector and the
Taking into account that the simulations with the algorithm try to remove the next one.
“greedy n” algoritm 3.2 are very time consuming Once finished that loop, the algorithm will grab
and that occurs a tendency to increase losses every the original grid again. But this time it will recon-
time that a connection between two bus is removed, nect the next connection that was removed to ob-
so the next approach was to convert the existing al- tain the radial grid, and redo all the process again.

2
 If there is no more lines/cables to reconnect, the From To 4Loss[pu] From To 4Loss[pu]
algorithm removes the original grid from the vec- Orig 0.175569 29 30 0.003948
tor and moves forward, making the same process 6 10 0.000547 14 15 0.002313
all over again to the next grid, until the vector isn’t 6 9 0.022983 6 7 0.016557
empty. 28 27 0.049130 2 6 0.012975
21 22 0.000395 23 24 0.089935
Whenever the algorithm calculates the losses in
6 28 0.000448 3 4 0.211921
a grid it will compare the value with the losses of
19 20 0.004561 Final 0.591280
the best grid find until the moment, updating if the
Total 0.415711
new value is lower, assuring that way that it will be
a variable with the best grid store at the end of the
The results for the 57 and 118 bus grids couldn’t
loop.
be obtained, five days after started the simulation,
In a simple way, this algorithm is composed by the algorithm was still running and it was decided
three loop. One inside loop that test the multiple to interrupt the simulation.
possibilities of replacing one given connection. An- For the last criteria, the losses of the resulting
other in the middle, whose job is to reconnect the radial grid are only 0.235223[pu] but, in this criteria
lines/cables to a designated grid. And finally, the the losses of the resulting grid are 0.591280 [pu], a
outer loop, which function is to make across the value much higher. At this point it is believed that
vector with radial grids. this criteria is not the best one to transform a mesh
grid to a radial one.
4. Results
In this paper, it will be only presented the re- 4.3. Greedy algorithm simulation lower volt-
sults obtained for the 30 buses grid, but the final age drop
conclusions and analysis of the results are based on The last criteria to be tested was, removing the
all four grids 2 . The results for the others electrical connection with less voltage drop between the bus
grid, as well as the algorithms, will be included in at the ends. For this one the resulting grid was:
the thesis.
From To 4Loss[pu] From To 4Loss[pu]
Orig 0.175569 6 7 0.000348
4.1. Greedy algorithm simulation with less
21 22 0.000072 23 24 0.007821
current 6 8 0.009909 9 10 0.005891
With this criteria, the sequence of which the con- 4 6 0.034119 14 15 0.000245
nections were removed is: 19 20 0.000827 3 4 0.049763
16 17 0.010673 29 30 0.001686
25 27 0.000404 Final 0.297327
From To 4Loss[pu] From To 4Loss[pu] Total 0.121758
Orig 0.175569 29 30 0.001465
8 28 0.000251 10 17 0.000921 For this criteria the simulation time also was high
14 15 0.000071 6 10 0.000075 but, in this case, the results for the 118 bus grid
23 24 0.000148 5 7 -0.000020 could be obtained. Taking into account the simu-
21 22 0.000028 2 4 0.013560 lation time and the losses in the resulting grid, it
24 25 -0.000005 4 6 0.042813 was easy to conclude that the best option to choose
18 19 0.000347 Final 0.235223 which connection it should be remove was the first
Total 0.05965 one, removing the connection with lower current.

4.4. Greedy n algorithm simulation

In all the grids it is visible a tendency for the Running the algorithm with the grid of 30 buses
losses to increase every time a connection are re- and 3 children for every node, the final result was:
moved.
From To 4Loss[pu] From To 4Loss[pu]
4.2. Greedy algorithm simulation with lower Orig 0.175569 19 20 0.002716
23 24 0.000158 5 7 0.000051
losses
14 15 0.000047 6 10 0.000066
If the criteria chosen was to remove the connec- 24 25 -0.000008 4 6 0.041870
tion with lower losses the resulting grid will be: 21 22 0.000034 2 4 0.000404
29 30 0.001443 V adjust 0.000000
8 28 0.000283 Final 0.223687
10 17 0.0010536 Total 0.048118

3
 After a few days simulating with the 57 bus grid, 5. Analysis of the results
it was obvious that the algorithm was making repet- In the initial phase, it was chosen the best criteria
itive decisions and end up with the same radial that for removing the connections: removing the line/-
it was already been analyzed. So it was decided to cable with lower current circulating at that time,
stop the algorithm and make the analysis of the lower voltage drop between the bus or to remove
grids already obtained. the line with the lowest value of losses.
With this algorithm it is still visible the problem After some simulations, it was easy to conclude
with the memory and simulation time required. that removing the connection based on value of
losses was not the best choice. Despite only having
4.5. Branch and bound algorithm simulation the obtained results to the 30 bus grid, the value
The results for the 30 bus grid are: of losses in the resulting radial grid was the highest
of all. By comparing the amount of losses on the
From To 4Loss[pu] From To 4Loss[pu] initial 30 bus grid with the final, obtained with this
Orig 0.175569 10 17 0.000626 criteria, it is visible that there was an increase of
8 28 0.000251 9 10 -0.000018 237% in the value of losses.
14 15 0.000071 5 7 0.000139 Pf inal −Pinicial 0,591280−0,175569
Pinicial = 0,175569 = 237%
10 22 0.000892 4 6 0.031065
21 22 0.000998 2 4 0.001701
24 25 0.001069 V adjust 0.000000 When the option of removing the connection with
19 20 0.005270 Final 0.219095 lower voltage drop is tested to see if it is the best
29 30 0.001460 Total 0.043526 option or not, a problem immediately figures. The
simulation time was too high and as such, it is only
This algorithm was able to find the best radial possible to obtain results for the grids with 14 and
grid for the electrical grid with 14 and 30 bus. Also, 30 bus it also should be noticed that the losses of
in the 30 bus grid the increment in losses was only the resulting grid are superior to those obtained by
24,79%. removing the line/cable with the lower current.
Taking into account all the aspects mentioned
Pf inal −Pinicial 0,219095−0,175569
Pinicial = 0,175569 = 24, 79% previously, it was decided that the best criteria
would be, to remove the line/cable with the lower
. current at that moment because this method re-
The results have shown that does not exist a pat- sulted in radial grids with lowest losses with a less
tern in choosing the path through the tree so, it simulation time.
is impossible to predict the connection on the best Considering this option, for the electrical grid
radial grid without testing them all. with 14 bus, the increase in losses:
Pf inal −Pinicial 0,184876−0,133933
= = 38, 036%
4.6. “Side subst” algorithm simulation Pinicial 0,133933

Finally, with the “Side subst” algorithm it could

be obtained the best radial grid to all mesh grids And for the electrical grid with 30 bus was:
in very little time. In the case of the 118 bus mesh Pf inal −Pinicial 0,235223−0,175569
grid the resulting radial grid was the same as the Pinicial = 0,175569 = 33, 978%
one obtained by the classic method. For the 30 bus
grid, the connections that was removed were:
And finally, for the network with 57 bus :
From To 4Loss[pu] From To 4Loss[pu] Pf inal −Pinicial 0,3818−0,278638
Pinicial = 0,278638 = 37, 03%
Orig 0.175569 10 17 0.000626
8 28 0.000251 9 10 -0.000018
14 15 0.000071 5 7 When building a tree and performing depth-first
0.000139
10 22 0.000892 2 4 search on it, the obtained radial grids have lower
0.014897
21 22 0.000998 4 6 value of losses. However, to get the best radial net-
0.017869
24 25 0.001069 V adjust work, it would be necessary to increase the number
0.000000
19 20 0.005270 Final of children that each node of the tree can create.
0.219095
29 30 0.00146 Total Which meant, that the algorithm would became
0.043526
slower. Although, keep in mind that the order in
As it can be seen this algorithm simply change which connections are removed does not affect the
the order which by the two last connections are re- final radial grid (if the order in which the connec-
moved, the final grid obtained was the same that tions are removed would be switched the final grid
the branch and bound algorithm 3.3 . would be the same).

4
 With the objective of obtaining the result quickly, mesh grid in to a radial grid. This is achieved
the concept of the branch and bound algorithm 3.3 through the process of opening meshes by succes-
was applied, so that the radial grids with lower sively disconnecting lines between buses. To achieve
losses, corresponding to meshed grids with 14 and this goal various Greedy approaches are tested to-
30 bus, could be obtained. In the new radial grid gether with other more complex algorithms like the
obtained for the 14 bus grid the increase of losses Branch and Bound algorithm. This analysis was
was: based on four standard IEEE networks.
Pf inal −Pinicial 0,162303−0,133933 The Greedy algorithms led to very high simula-
= = 21, 18%
Pinicial 0,133933 tion times as a consequence, the radial grid of most
IEEE networks could not be obtained. which was
In the grid with 30 bus the losses increment was: mainly due to the high number of power-flows re-
quired. To find the radial network with lower losses,
Pf inal −Pinicial 0,219095−0,175569
Pinicial = 0,175569 = 24, 79% it is necessary to find all possible networks and run
a power-flow for each one of them, which makes the
simulation time for the Greedy algorithms simula-
Remember that the increments obtained by the tion became high.
classical method are 38.036% for the grid with 14 To reduce the computational effort it was applied
bus and 33.978% for the grid with 30 bus. As can a Branch and Bound algorithm to the networks.
be seen, in the case of the grid with 14 bus, the dif- With this approach the simulation time was drasti-
ference of the value of losses between the radial net- cally reduce and, the best radial grid for the IEEE
work obtained by the classical method and the ra- 14 and 30 grids could be obtained.
dial grid with lower losses are 16.9%. In the case of By analyzing the results obtained with the
30 bus grid, this same difference decreases to 9.2% Greedy algorithm it is clear that the radial grid ob-
so, in the case of 30 bus grid, the classic method tained not always corresponds to the best radial
gives a radial grid where the value of losses is close grid, from the standard of lower losses. However,
to the best radial grid. the losses of the resulting grids are very close to the
When running the simulation with the algorithm ones obtained by the Branch and Bound algorithm.
“side subts” 3.4 , the results for the 57 and 118 bus The process of making a grid become radial im-
grids could be obtained. In the case of the grid with plies an increase in its losses, the value of this incre-
118 bus the result obtained with this algorithm was ment depends from the grid chosen. The increments
the same that was achieved with the classic method that was obtained are:
of removing the connections with lower currents.
For the 57 bus mesh grid the best radial grid ˆ 21% for IEEE 14 grid
obtained corresponds to an increase in the value
of losses of 6.516%, this increment was very small ˆ 28% for IEEE 30 grid
when compared with the results obtained for the 14
and 30 bus grids. ˆ 6,5% for IEEE 57 grid
Pf inal −Pinicial
Pinicial = 0,296794−0,278638
0,278638 = 6, 516% ˆ 39,5% for IEEE 118 grid

As it can be seen, the value of the increment

Observing the results obtained, it is clear that changes significantly from grid to grid. Loss incre-
they depend highly on the network used. In the ments have a variation from 6 to 40%, demonstrat-
case of 57 bus grid the classic method gave a grid ing that the decision on changing a mesh grid into a
very close to the best radial grid. But, in the case radial one cannot be taken for the generality of the
of the grid with 14 bus, this argument is no longer electrical grids without significant risk of degrading
valid because the difference is much greater. Thus, efficiency.
the classic method of removing one by one the suc- Also worth noting is the fact that, in the IEEE
cessive line/cables with lower current not always re- 14 bus grid it was removed 7 connections from the
turns the best radial grid from the standard point initial 20, and in the IEEE 30 bus, was removed 12
off view off lower losses. Therefore, it will be needed from the 41 initials connections meaning that, in
to test one by one all possible grids, despite the fact both cases, exists a 30% reduction in the number of
that the time required for this operation could prove connections.
to be too high for networks with a high number of
References
bus.
[1] Redes de Energia Eléctrica - uma análise
6. Final conclusions sistémica; José Pedro Sucena Paiva; IST
As initially mentioned, in this several algorithms Press; Abril de 2005; ISBN: 972-8469-34-9
are studied with the objective of transforming a

5
[2] Digital Computer Solution of Power-Flow
Problems; J. B. Ward, H. W. Hale;
Power Apparatus and Systems, Part
III. Transactions of the American Insti-
tute of Electrical Engineers; Junho 1956
http://ieeexplore.Ieee.org/xpls/abs_
all.jsp?arnumber=4499318&tag=1

[3] Power Systems Test Case Archive; http://

www.ee.washington.edu/research/pstca/

[4] Matpower: A MATLAB Power System Sim-

ulation Package; Ray D. Zimmerman, Carlos
E. Murillo-Sánchez & Deqiang (David) Gan;
http://www.pserc.cornell.edu/matpower/

[5] Wikipédia page about Best-first search al-

gorithm; http://en.wikipedia.org/wiki/
Best-first_search