IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 61, NO.

6, JULY 2012 2421

Unified AC/DC Power Flow for Traction

Systems: A New Concept
Pablo Arboleya, Member, IEEE, Guzman Diaz, Senior Member, IEEE, and Manuel Coto, Member, IEEE

Abstract—This paper adds some new contributions to the uni- Superscripts

fied alternating current/direct current (ac/dc) power-flow method
that is applied to railway power supply systems. These contribu- T Transpose matrix.
tions are mainly focused on the way that the unified power-flow tt, ts, ss Train–train, train–substation, and substation–
problem is implemented. The authors propose a new technique substation.
based on graph theory to model the motion of the trains without DC Direct current (dc) system.
varying the system topology and dimensions, as well as the solution
vector. Furthermore, a new matrix formulation is developed to AC Alternating current (ac) system.
provide an easy way of setting out the problem. The combination L Links.
of these techniques makes the unified ac/dc power flow easier
to implement, and the comparison among different instants can
directly be done, representing the major contribution of this pa- Subscripts
per. Finally, as a minor contribution, a new technique based on
previous matrix formulation is developed to easily obtain all active t, s Train and substation.
and reactive power magnitudes in compact form. d, q Direct and quadrature components.
Index Terms—Combined alternating current/direct current B, N Branch and node.
(ac/dc) power flow, graph theory, rail transportation power i, j ith and jth elements.
system.

N OMENCLATURE I. I NTRODUCTION
Acronyms
RPSS Railway power supply system.
T HE RPSS represents an important part of the whole
traction system infrastructure. Light direct current (dc)
traction systems, such as the system modeled and analyzed
KCL Kirchhoff’s current law. in this paper, are based on a dc network (where the trains
are connected), coupled with the main alternating current (ac)
power grid through alternating current/direct current (ac/dc)
Variables conversion devices. The most common situation nowadays is
the use of noncontrolled six- or 12-pulse rectifiers combined
Λ Adjacency matrix.
with power transformers. However, the number of bidirectional
Γ Incidence matrix.
substations installed, permitting power flows in both ways, has
ω Synchronous reference frame rotational speed.
increased during the last years. With the use of this kind of
L, R Inductance and resistance.
devices, a more efficient energy management strategy can be
R, X Resistance and reactance matrix.
developed in the dc network, allowing the regenerated energy
v, i Voltage and current vector.
of the trains to be injected in the ac grid. To make a correct
I, S Identity and block diagonal matrix.
dimensioning, take correct decisions about future investments,
M Linear equations matrix.
or simply make an accurate estimate of the operation costs, a
z Current and voltage solution vector.
good estimation of energy consumption and peak power at each
Reqi Converter-equivalent resistance.
line and substation is needed.
P, Q Active and reactive power matrix.
The most used tool for this purpose is the power-flow
n Number of.
approach [1]–[3]. Based on this concept, there are different
tools that can work out the estimate. In [4], a neural-network-
based tool was used. Inputs such as power flows, power system
Manuscript received September 5, 2011; revised November 22, 2011, configuration, the distance between substations, average values
January 11, 2012, March 12, 2012, and April 2, 2012; accepted April 9, 2012. and standard deviations of the railway inclinations, and the
Date of publication April 25, 2012; date of current version July 10, 2012. This average number of trains and its average velocity, were used
work was supported in part by the Spanish Ministry of Science under Grant
MICINN-10-IPT-370000-2010-15. The review of this paper was coordinated to train the network.
by Dr. C. Chunting. As aforementioned, a number of dc traction systems are
The authors are with the Department of Electrical Engineering, University connected to the public grid through power transformers, and
of Oviedo, Campus de Viesques, 33204 Gijón, Spain (e-mail: arboleyapablo@
uniovi.es). many of the RPSS are dc systems that are connected to the
Digital Object Identifier 10.1109/TVT.2012.2196298 main power system through ac/dc converters. In this case, the
0018-9545/$31.00 © 2012 IEEE
2422 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 61, NO. 6, JULY 2012

power-flow problem cannot directly be solved by conventional into stage or exits the system. In addition, the order of the
methods used in ac networks, and two main methodologies have variables in the solution vector is kept; therefore, different
been developed. instants can directly be compared. Second, the development of
The first approach is to consider only the dc subsystem a matrix formulation reduces the implementation complexity of
problem, considering the ac/dc substations as a dc voltage or these unified methods, overcoming the main drawback of the
current source, with a series or parallel connected resistance, extended variable methods. The use of matrices that describe
respectively. Most of the authors who use this simplification power flows in electrical networks is a widely accepted method
use linear and direct methods based on the Zollenkopf bifactor- of showing the final results [11], [12], and in this paper, we
ization [5] or the incomplete Cholesky conjugate gradient [1]. propose, as a minor contribution, a new method of computing
The use of this approach, however, does not consider the effect these matrices using the aforementioned matrix formulation.
of the ac grid voltage drops in the dc subsystem. Thus, in a In Section II, we will explain the proposed method to de-
real scenario, two identical ac/dc converters with the same load scribe the dc topology. The description complexity of this part
level and different voltage outputs can be found [6], but with the of the problem is higher, and we will explain how the topology
aforementioned methods, such a situation cannot be modeled. is kept constant even when the train motion is simulated. Sec-
Therefore, to get a more accurate solution, a combined tion III describes all linear and nonlinear expressions needed to
power-flow computation, including ac and dc subsystems, must solve the ac/dc combined power flow. In Section IV, we specify
be solved. Combined ac/dc power-flow methods are not new and explain the method used to solve the problem and the
[7], [8], and the following two representative trends have been proposed procedure for calculating the active and the reactive
adopted to solve ac/dc power flows: 1) the sequential method power. In Section V, a specific case of study is defined, stated,
and 2) the unified method. and solved using the proposed method, and the commercial
Sequential methods apply an iterative procedure between ac software (DIgSILENT). The obtained results are analyzed and
and dc systems [2], [3], [6]. These methods consider ac/dc compared to test the method accuracy. Finally, in Section VI, a
converters as voltage or current sources from the dc subsystem set of conclusions will be stated.
point of view and loads from the ac point of view. In most of
the cases, a plain voltage profile in the dc subsystem is adopted
for the first iteration in the dc part. Under this assumption, II. DC P OWER S YSTEM T OPOLOGY D ESCRIPTION
the power demanded by each substation from the ac system is
To properly analyze the dc traction network, two main prob-
computed. Thereafter, the ac power flow is solved to correct the
lems must be settled. The first problem refers to the number
initial dc voltage profile.
of trains within the system at a given simulation instant, which
Differently, unified methods simultaneously solve the whole
is not constant and implies a change in the problem dimension
system of equations. These methods have also been called the
over different simulation instants. The other problem that we
extended variable method, because the dc variables are added
have to deal with resides in the train motion, which produces
to the ac solution vector [7], [8].
changes in the relative position of the network nodes during the
The main advantage of sequential methods lies in their
simulation. Traditionally, to overcome these difficulties a new
simplicity of implementation. They present, however, some
problem is stated at every simulation step.
convergence problems under particular situations that were
No major inconveniences appear when a small system and a
derived from the discontinuities that emerged in the rectifier
short time interval are analyzed. However, the drawbacks that
interface quantities between successive ac/dc iterations and the
are derived from the traditional way of setting up the problem
nonlinear complexity of the rectifier model [3], [7]–[9]. Note
arise when a real system must be studied over a significant
that, in a particular case of dividing the track into a number of
period of time, causing a set of serious disadvantages, which
equal-length finite cells, as done in [6], the number of nodes
are summarized as follows.
drastically increased, and the impedances of the ac and dc
lines are spread too wide of a range. Under these particular • A procedure for determining which trains are in the system
circumstances, it could be preferable to use sequential methods. and their position must be developed. We need to define a
Conventionally, the stationary equivalent method for moving new topology at every simulation step, varying the number
loads, as proposed in [10], is assumed. This method assumes of nodes, their position, and the lines that connect such
that the speed of the trains is not very high to induce pro- nodes.
nounced electrical transients and the dc traction network slowly • Enumeration criteria must be designed to identify each
moves from one state to another as the locations and the network element (node/line) at every simulation instant.
input power of the trains vary. Using this stationary equivalent, Because the same criteria are applied to different instants,
temporal analysis of RPSS is computed by solving successive different numbers may be assigned to the same element at
instants. two different instants.
This paper contributes to existing unified ac/dc power-flow • Due to these last points, the solution vector changes in
methods, adding two main features. First, a graph-theory-based its dimension and its components, making it very hard
technique allows, in an easy way, space–time variation in to compare different instants. Although vectors that cor-
loads to simulate the motion of the trains. With the proposed respond to two different instants have the same dimen-
approach, train motion can be modeled, keeping the system sions, the same vector position could belong to different
topology and dimension constant, even if a new train comes variables.
ARBOLEYA et al.: UNIFIED AC/DC POWER FLOW FOR TRACTION SYSTEMS: NEW CONCEPT 2423

is a complete bipartite graph between trains and dc nodes.

A complete bipartite graph is a graph whose vertices can be
divided into two disjoint and independent sets, trains U , and dc
nodes V such that every vertex in U is connected to every vertex
in V [see Fig. 1(b)]. In this case, the number of edges can be
calculated by multiplying the number of trains by the number
of dc nodes (nt × ns ).
The third subgraph represents the real topology of the dc
system without trains; the vertices of this graph are the substa-
tions or even ramification points in a dc line. The edges in this
graph represent only the real connections between dc nodes;
therefore, the number of edges in this subgraph depends only
on the number of real lines that connect dc nodes.
Each graph can be described by using its associated ad-
jacency matrix [13], [14]. Our method requires oriented-type
graphs, which are directed graphs where only one edge exists
for each pair of vertices. In this case, the vertex enumeration
will determine the edge direction. The edge tail will be the
lower vertex, and the edge head will be the higher vertex.
Under this assumption, the adjacency matrix will be upper
triangular, and the enumeration of the edges will be easier. The
Fig. 1. Set of subgraphs that represent the dc network topology. (a) Subgraph
that describes all possible connections between trains (complete-type graph).
adjacency matrix that corresponds to the whole dc subsystem
(b) Subgraph that describes connections between trains and dc nodes (complete will be called Λdc and can be calculated as the sum of the
bipartite graph). (c) Subgraph that describes the real dc network topology. aforementioned subgraphs adjacency matrices as

• A tracking subroutine must be developed to search within Λdc = Λtt + Λts + Λss
the solution vector a determined variable to be compared  tt   ts   ss 
with the same one at a different instant. Λ11 Λtt12 Λ11 Λts
12 Λ11 Λss
12
= + +
One of the key features of the proposed approach is the use Λtt tt
21 Λ22 Λts ts
21 Λ22 Λss ss
21 Λ22
of a graph-theory-based method to define the system to over-  tt 
Λ11 Λts12
come the aforementioned difficulties that were derived from = (1)
the changing topology. Our method considers the trains and dc 0 Λss22
nodes as vertices and the catenaries as edges of a graph. To
keep the problem dimension constant, all the trains that appear where the following conditions hold.
in the temporal interval of study will be represented in the
• Λdc is the dc system adjacency matrix.
graph, regardless of whether they are physically in the system.
• Λtt is the adjacency matrix that represents the first sub-
The enumeration criterion for vertices starts with the trains and
graph (for connections between trains, see Fig. 1(a)).
follows with the dc nodes. With this system, a determined node
• Λts is the adjacency matrix that represents the second
always represents the same train or dc node, independent of the
subgraph [for connections between trains and dc nodes,
simulation instant.
see Fig. 1(b)].
Nevertheless, the train motion still causes changes in the
• Λss is the adjacency matrix that represents the third sub-
system topology. To construct an invariant dimension system,
graph [for real connections between nodes, see Fig. 1(c)].
we consider that all trains are connected among them and with
all dc nodes. The graph that represents the whole dc system is The matrices construction is simple. Due to the use of the
composed of three subgraphs that represent connections among proposed vertex enumeration criteria and the use of oriented-
trains, connections between trains and dc nodes, and the real dc type graphs, Λdc will be upper triangular; therefore, Λss tt
21 , Λ21 ,
topology, respectively (see Fig. 1). and Λ21 are null matrices. Furthermore, Λ11 , Λ22 , and Λ∗∗
ts ∗∗ ∗∗
12
In Fig. 1(a), the first subgraph is depicted. It represents all represent connections between trains, dc nodes, and train–dc
possible connections between trains. Note that the case where nodes, respectively, which means that only Λtt ss
11 , Λ22 , and Λ12
ts

all trains run between two dc nodes can exist. To cover all can be not null matrices.
possibilities, this graph is a complete-type graph, i.e., a simple Λtt
11 (nt , nt ) is pure upper triangular, because connections
graph in which every pair of distinct vertices is connected by a among trains are represented by a complete-type graph. A pure
unique edge. In this case, the number of edges can be calculated upper triangular matrix is an upper triangular with nonnull
using the following expression: nt (nt − 1)/2. elements above the diagonal. Λts 12 (nt , ns ) will be filled with
The second subgraph shows the possible connections be- ones (it represents a complete bipartite graph, with all trains
tween the trains and the dc nodes. Note that every train can connected with all dc nodes). Λtt 22 (ns , ns ) represents the real
be connected at different simulation steps with every dc node. topology of the dc network and, in this particular case of
In this case, the subgraph that symbolizes these connections study, is a pure upper triangular matrix, because we only have
2424 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 61, NO. 6, JULY 2012

three dc nodes (substations in this case) connected each other; the implementation procedure. The second group of equations
otherwise, it would be an upper triangular matrix. is formed by the nonlinear equations.
As observed in Fig. 1, we first enumerate the edges whose
tail (lower indexed node) is vertex 1 following an ascending
order as a function of their head (higher indexed node), and then A. Linear Equations
edges whose tail is vertex 2, and so on. With these criteria, the To model the whole power system, the ac lines and the dc
dc element node incidence matrix (Γdc ) [13], [14] can easily catenaries have been considered as RL branches. In the ac part
be obtained using the adjacency matrix Λdc with the following of the problem, the complex vector theory has been used. The
procedure. The nonzero Λdc elements are extracted row by row, voltage drop (Δvdq ) in a resistor–inductor (RL) circuit can dy-
obtaining their indices (i, j). For each pair of values (i, j), a namically be expressed using this theory as follows [17], [18]:
new row in the Γdc matrix will be added. The column i will be  
filled with 1, and the column j will be filled with −1. Therefore, d
Δvdq = R idq + + jω L idq (3)
the Γdc rows and columns will represent, respectively, the dc dt
subsystem graph edges and vertices, and its element Γij is
given as follows. where

• Γij = 1 when the tail of the edge i is vertex j. vdq = vd + jvq (4)
• Γij = −1 when the head of the edge i is vertex j.
• Otherwise, Γij = 0. idq = id + jiq . (5)

Γdc will be used in the following sections to set out a set Equation (3) is generic; it serves for either transient or steady-
of linear equations that correspond to the KCL and Kirchhoff’s state analysis, and it gives us insight to proceed to decoupling
voltage laws of the dc subsystem. Γdc could directly be con- the system into dq components. The system will be analyzed
structed without the use of Λdc . However, we first construct in steady state, and therefore, the derivative term is null. The
Λdc to simplify the data input. The construction of an upper presence of (3) is intended only to give support to (6)–(9),
triangular Λdc is easier than the construction of Γdc directly. which are (3) but particularized to the steady state.
The traditional formulation allows us to express in vector According to the aforementioned, in case of the ac lines, (3)
form the power-flow equations as follows: in steady state can be written as follows:

s̄ = v̄ Ȳ∗ v̄ ∗ (2) Δvd = Rid − ωLiq (6)

Δvq = Riq + ωLid . (7)
where Y is the admittance matrix, and v̄ and s̄ represent the
complex voltage and power, respectively. Using the per-unit (p.u.) system, (6) and (7) can also be applied
However, the use of a method based on the admittance matrix to the branches that include power transformers. The Γ matrix
presents some disadvantages for our particular purpose [15], notation allows us to express such equations in vector form as
[16]. In the admittance matrix construction, the information
related to the network topology and shunt, as well as the series ac T ac ac T
 
ac ac T
Γac (vN d ) = RB (iBd ) − XB iBq (8)
devices parameters, is merged. Once the admittance matrix is  ac T  ac T ac T
obtained, we cannot go back in a univocal way to the network Γac vN q = RacB iBq + XacB (iBd ) (9)
topology or parameters. In our case, the topology and the
parameters will change at every simulation step; therefore, the where the following conditions hold.
admittance matrix should be computed each time. For this • Γac is the element node incidence matrix of the ac
reason, we advise against the use of the admittance matrix, subsystem.
even if a sequential technique is implemented. With the use of • Rac ac
B and XB are the resistance and reactance matrices, re-
Γ, the information with regard to nodes connection (constant spectively, that represent the impedance between ac nodes.
during all problem) and the line impedances (updated at each They are diagonal matrices where rii and xii represent,
iteration) is not merged. The impedance matrices are diagonal, respectively, the resistance and reactance of branch i or the
as observed in the next section, and their updating is much short-circuit resistance and reactance of the transformer
easier than the admittance matrix update. placed in the branch in case of the ac branches that
connects the ac and the dc system.
ac ac
• vN d and vN q are the node voltage vectors in components
III. AC/DC S YSTEM M ODEL d and q, respectively, of the ac subsystem.
The elements that are involved in the ac/dc traction network • iac ac
Bd and iBq are the branch current vectors in components
model can be summarized in ac lines, dc catenaries, ac and dc d and q, respectively, of the ac subsystem.
buses, power transformers, power converters, ac static loads, The same procedure is applied to the dc network, and in
and trains. Our modeling procedure will divide the device this case, the voltage drop in vector form can be expressed as
equations into two main groups. The first group is formed by follows:
the linear equations. The proposed method will use the Γ matrix  dc T  dc T
to obtain such equations, making an enormous simplification in Γdc vN = RdcB iB (10)
ARBOLEYA et al.: UNIFIED AC/DC POWER FLOW FOR TRACTION SYSTEMS: NEW CONCEPT 2425

where the following conditions hold.

• Rdc B is the branch resistance matrix of the dc subsystem.
It is a diagonal matrix, and rii represents the resistance
of branch i. Note that the proposed model of the dc system
described in the previous section generates a set of dc lines
that are not simultaneously used in the same simulation
step. The use of this formulation allows us to assign an
infinite value to lines that are not used lines, and therefore,
they do not have any influence in the system. In case of
one or more trains situated in the same station, they are
automatically separated a negligible distance (10−3 times
the total distance of the line where the trains are located)
to guarantee that two nodes are not in the same position,
avoiding the existence of branches with null length. Rdc B
Fig. 2. Detailed topology of an ac/dc link.
is the only matrix that is affected by the train motion
Using this formulation, all aforementioned linear equations
and must be updated when a train changes its location. In
can be expressed in compact form as follows:
this case, to implement the infinite, a number that is high
enough number was used. We implemented infinite as 106 . MzT = 0 (14)
All calculations are expressed in the p.u. system, and thus,
such a number is a high-enough value to obtain negligible where z is the vector that represents voltage and current magni-
currents when a branch is not active, without incurring in tude that is constructed in the following form:
an unacceptable condition number. 
• vNdc
is the node voltage vector of the dc subsystem, and in z = idcB iac
Bd iBq
ac
idc
t iL
B iL
Bd ···
this case, a node can represent both a conventional dc node
· · · iac
Nd iL
Bq iac
Nq
dc
vN ac
vN d
ac
vN q . (15)
(substation or ramification) or a train.
• idc
B is the branch current vector of the dc subsystem. The construction of M is represented in (16), shown at the top
The use of Γ can also be applied to the formulation of the of the next page.
KCL at each node of the system. In case of an ac subsystem,
the KCL is expressed as follows:
B. Nonlinear Equations
T L ac T

(Γac )T (iac
Bd ) = I iBd iN d (11) The nonlinear equations are derived from the ac network P Q
  T  T and P V nodes models, train model, and converter model. P Q
(Γac )T iacBq = I iLBq iac
Nq (12) nodes in the ac network contribute to the following expressions:
where the following conditions hold.
di iN di + vN qi iN qi − Pi = 0
ac ac ac ac
vN (17)
• iac ac
N d and iN q are the injected currents in components d and
q in the ac system nodes.
ac ac
vN qi iN di − ac ac
vN di iN qi − Qi = 0. (18)
• iL L
Bd and iBq are the branch current vectors in components The equations that correspond to the P V nodes are
d and q, respectively, in the branches labeled as link
di iN di + vN qi iN qi − Pi = 0
branches. In Fig. 2, the detailed topology of an ac/dc link is ac ac
vN ac ac
(19)
depicted. As observed, branch j is a link branch, because
 2
it connects an ac node with a dc node using a power ac )2 + v ac
(vN di N qi − |vN
ac
i | = 0. (20)
converter. This kind of branches combines ac magnitudes
upstream the converter (iL L
Bd , iBq ) with dc magnitudes
L With this approach, a complex model of both trains and con-
downstream the converter (iB ).
verter could be implemented by simply adding the equations.
• I is the identity matrix.
However, in this case, we have implemented a simple train
The vector form of KCL in the dc subsystem is model as follows:
 T  T
(Γdc )T idc = S idc iL (13) dc dc
Pi = v N
B t B i iti . (21)
where the following conditions hold. In our case of study, the train power will be provided by
• idc
t represents the current absorbed by the trains. other complex software package that uses the rail and trains
• S is a block diagonal matrix. The first block is the identity parameters for a given unit and route. The output data of this
matrix of dimension nt , and the second block is a diagonal software are the power absorbed or injected by the train in the
matrix denoted as SL of dimension ns . Element sii , which dc network at each instant. Nevertheless, a complex train model
belongs to SL , is 1 if the dc substation i is connected with could be implemented using the same procedure. Something
the ac network, and sii is 0 when the dc substation i is not similar occurs with the ac/dc converter model. We propose
connected with the ac grid. here the use of a simple model of an ac/dc six-pulse diode
2426 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 61, NO. 6, JULY 2012

converter. Equations (22) and (23), shown below, represent an incidence node matrix (Γ) is constructed by joining Γdc , Γac ,
ideal lossless converter, and (24), also shown below, represents and ΓL , where ΓL is the node incidence matrix of the links, and
the voltage drop that was derived from the commutation and its structure can be observed in the following expression:
conduction losses ⎛ dc ⎞
Γ 0
⎜ ⎟
di iBdi + vN qi iBqi − vN i LBi = 0
L L L L L
vN (22) Γ=⎝ ΓL ⎠. (25)
qi iBdi − vN di iBqi = 0
L L L L
vN (23) 0 Γac

 L 2  L  2 Using Γ, we can define the ac and dc extended node inci-

i − 1.35 vN di + vN qi − Reqi iL
L
vN Bi = 0 (24)
dence matrices ΓDC∗ and ΓAC∗ as follows.
where Reqi is the equivalent resistance of the conversion unit in • ΓDC∗ is the extended Γdc . It is constructed using the first
the regular commutation range. A better comprehension of its (ndc
B + nL ) rows and the first (nt + ns ) columns of Γ.
calculation is found in [15] and [19]. Using the same procedure, • ΓAC∗ is the extended Γac . It is constructed using the last
complex models of noncontrolled or controlled converters can (nL + nacB ) rows and the last (nac ) columns of Γ.
DC∗
be implemented; for example, see [2] and [9]. Γ and ΓAC∗ are used to obtain three matrices called
extended power matrices PDC∗ , PAC∗ , and QAC∗ . Their cal-
culation can be observed in
IV. S OLVING P ROCEDURE
P DC∗ (i, j) = idc dc
B (i)vN (j)Γ
DC∗
(i, j) (26)
Once the problem is stated, we solve it using the well-known AC∗
P (i, j) = iBd (i)vN d (j)ΓAC∗ (i, j)
ac
trust-region dogleg algorithm [20]. All linear and nonlinear
equations are solved at the same time. The use of the graph + iac
Bq (i)vN q (j)Γ
AC∗
(i, j) (27)
AC∗ ac AC∗
theory and the Γ matrix to obtain the power flows and losses Q (i, j) = iBq (i)vN d (j)Γ (i, j)
through the branches and the injected or extracted power in the − iBd (i)vN q (j)Γ
ac AC∗
(i, j). (28)
nodes simplifies the calculus. Two matrices, which are called
Once the extended power matrices are calculated, P and Q
the active and reactive power matrices and are denoted as P
can be constructed as follows:
and Q, respectively, are computed. The P and Q dimensions  
and structure are the same as the whole system node incidence PDC∗ 0
matrix (Γ). Each row of the power matrices represents a branch, P= (29)
0 PAC∗
and it will have two nonnull positions that correspond to the  
outgoing and incoming power, respectively, in the nodes con- 0 0
Q= . (30)
nected by this branch. Applying this concept, the power losses 0 QAC∗
in a branch can be calculated as the sum of the terms of a row.
In the same way, each column represents a node, and the sum
V. R ESULTS A NALYSIS
of its elements is the injected power in the node represented
by that column. A complete power balance of the system can In this section, we will describe the specific case used to
be done by adding all the matrix elements. This explanation is validate the method. The ac/dc system is composed of an
also valid for the reactive power matrix. ac subsystem with six nodes, two generators, one load, and
The calculation of the power matrices is simple. The first three connections with the dc subsystem. The dc subsystem
step is the construction of the whole system node incidence is composed of three substations connected to each other and
matrix. Note that, in previous stages, we have worked with three trains, two of which are in the same line. A general view
Γdc and Γac to obtain the linear expressions that correspond of the proposed system is depicted in Fig. 3. The enumeration
to the KCL. To obtain the active and reactive power matrices, criteria were described in Section II, where we first enumerated
the nonlinearity conditions of the branches tagged as links the trains (nodes 1–3) and then the dc nodes (nodes 4–6). As
represents no inconvenience. For this reason, the whole system aforementioned, when a connection between the dc and ac
ARBOLEYA et al.: UNIFIED AC/DC POWER FLOW FOR TRACTION SYSTEMS: NEW CONCEPT 2427

TABLE IV
DC B RANCHES R ESISTANCE

TABLE V
N ODE VOLTAGES

Fig. 3. Proposed ac/dc system. The upper part of the system corresponds to
the ac subsystem, and the lower part corresponds to the dc subsystem. Trains 1
and 2 are located between nodes 4 and 5, whereas train 3 is located between
nodes 4 and 6.
TABLE I
AC B RANCHES E LECTRICAL PARAMETERS

TABLE VI
B RANCH C URRENTS

TABLE II
AC N ODES E LECTRICAL PARAMETERS

TABLE III
P OWER T RANSFORMER PARAMETERS

subsystem is activated, a new auxiliary ac node is included (in

this case, nodes 7–9). Finally, we assign numbers to ac nodes
(nodes 10–15). In Fig. 3, the active branches, defined by the real
positions of the trains in the dc network, are represented.
The unit resistance and reactance of the ac subsystem lines are, Ferrocarriles Españoles (RENFE) overhead contact system for
respectively, 0.09962 and 0.51442 Ω/km. The ac network has a maximum speed of 160 km/h. The resistance of the overhead
six branches with different lengths. The lengths of these branches conductor is 0.051 Ω/km, and the rails are 54 kg/m type with
and the total resistance and reactance are shown in Table I. a resistance of 0.007 Ω/km. In this case, we suppose that each
In the ac network, there exist different types of nodes, as de- train has a perfect connection with the ground; therefore, the
fined in Table II. The ac/dc links are composed of one transform- rail resistance is added to the catenary resistance. In Table IV,
er and one converter. In the case of study, the system has three the active dc branches resistance and length determined by the
links with the same converter and transformer. The convert- position of the trains are shown for this specific case.
er is a six-pulse noncontrolled converter according to (22)–(24); For the specific instant analyzed, trains 1 and 2 absorb 443
however, another type of converter can be implemented. The and 380 kW, respectively. Train 3 injects 260 kW. The voltage
power transformer characteristics are summarized in Table III. in all nodes and the current in all branches are shown in
In the dc subsystems, all the catenaries are CR160 type. The Tables V and VI, respectively. In these tables, the voltages
acronym CR160 refers to the concept of the Red Nacional de and currents (root mean square) obtained using the proposed
2428 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 61, NO. 6, JULY 2012

TABLE VII
ACTIVE P OWER M ATRIX ( IN K ILOWATTS)

TABLE VIII
R EACTIVE P OWER M ATRIX ( IN K ILOVOLT-A MPERES R EACTIVE)

method can be compared with the voltages and currents ob- VI. C ONCLUSION
tained using commercial software. In this case, the selected This paper has proposed a new approach for solving the
software used to validate the results was DIgSILENT Power- combined ac/dc power flow using an extended variable method.
Factory. The voltage error is limited to 0.3%, and the cur- Historically, extended variable methods have presented bet-
rent error is limited to 1.7%. The source of these errors can ter convergence properties; however, they are hard to imple-
be derived from the differences in power transformer and ment compared with the combined ac/dc power-flow sequential
power converter models used by the proposed method and methods. Moreover, the train motion increases the complexity
DIgSILENT. For example, in our case, the parallel impedance of setting out this problem using conventional extended variable
in the power transformer model and the power losses in the methods.
diodes are neglected, even when voltage drops were considered. In this paper, we have presented some contributions to over-
Tables VII and VIII represent the active and reactive power come these drawbacks and developed a procedure for setting
matrices described in (29) and (30), respectively, for the specific out, solving, and analyzing the problem in a really easy way.
case of study. As aforementioned, these matrices can easily be The first contribution is the use of the graph theory applied
obtained using the node incidence matrix, and they allow us to to the network description, making it possible to introduce
make a rapid power-flow analysis. space–time variation in loads in an easy way. The train move-
Note that special cases, considering trains in the same loca- ment is simulated by activating or deactivating edges or the
tion and nonactive trains, were simulated to test the reliability graph, but the dimension of the problem remains constant. As
of the method. Such cases were also validated using the same a consequence, solution vectors of two different instants can
aforementioned procedure. easily be created and directly compared without the need for a
ARBOLEYA et al.: UNIFIED AC/DC POWER FLOW FOR TRACTION SYSTEMS: NEW CONCEPT 2429

tracking subroutine to determine which position of the solution idc

B (1, ndc
B ).
vector belongs to a given variable. iac ac
N d , iN q (1, nac
B ).
The second contribution lies in the use of a matrix formu- L L
iBd , iBq (1, nL
B ).
lation based on the element node incidence matrix (Γ). This idc (1, ndc
B B ).
matrix formulation allows us to express all linear equations iL (1, ns ).
B
of the system using a single matrix expression. The structure idc (1, nt ).
t
of this matrix (M) remains constant with the train motion,
because it includes all possible derived topologies. Only the
values of Rdc A PPENDIX B
B , i.e., a diagonal matrix, must be updated as a
function of trains position at each simulation step. Furthermore, I NCIDENCE M ATRICES C ONSTRUCTION
a set of nonlinear equations can be added to introduce complex FOR THE C ASE OF S TUDY
models of loads, generators, trains, and converters. In this We have
case, we have included P Q and P V node type models and a ⎛ ⎞
simple model of trains where the injected or absorbed power 1 −1 0 0 0 0
as a function of time is a given variable. We use this kind of ⎜1 0 −1 0 0 0 ⎟
⎜ ⎟
train model, because this program will be combined with a ⎜1 0 0 −1 0 0 ⎟
⎜ ⎟
complex train simulator that calculates the power consumed or ⎜1 0 0 0 −1 0 ⎟
⎜ ⎟
injected in the catenary as a function of multiple train and rail ⎜1 0 0 0 0 −1 ⎟
⎜ ⎟
parameters. However, a nonlinear complex train model could be ⎜0 1 −1 0 0 0 ⎟
⎜ ⎟
implemented, if necessary. With regard to the converter model, ⎜0 1 0 −1 0 0 ⎟
⎜ ⎟
six-pulse noncontrolled power converters were used, but again, Γdc = ⎜0 1 0 0 −1 0 ⎟
⎜ ⎟
different types of converters can be added. ⎜0 1 0 0 0 −1 ⎟
⎜ ⎟
Finally, as a minor contribution, we have included a new ⎜0 0 1 −1 0 0 ⎟
⎜ ⎟
method of obtaining all power flows through the lines, losses, ⎜0 0 1 0 −1 0 ⎟
⎜ ⎟
and injected/absorbed power in all nodes in a compact form. ⎜0 0 1 0 0 −1 ⎟
⎜ ⎟
All active and reactive powers are calculated using a matrix ⎜0 0 0 1 −1 0 ⎟
⎝ ⎠
formulation based on the element node incidence matrix (Γ) 0 0 0 1 0 −1
0 0 0 0 1 −1
and are represented in matrix form, allowing a rapid power- ⎛ ⎞
flow analysis. As observed, the results do not differ from the 1 0 0 0 −1 0 0 0 0
⎜0 1 0 0 0 −1 0 0 0 ⎟
results obtained with the available commercial software, but ⎜ ⎟
⎜0 0 1 0 0 0 −1 0 0 ⎟
the train motion implementation is easier, and the introduction ⎜ ⎟
⎜0 0 0 1 −1 0 0 0 0 ⎟
of complex nonlinear models of loads and network elements is ⎜ ⎟
Γac = ⎜0 0 0 1 0 0 −1 0 0 ⎟.
more flexible. ⎜ ⎟
⎜0 0 0 1 0 0 0 0 −1 ⎟
Further work could cover the dynamic simulation of the ⎜ ⎟
⎜0 0 0 0 1 0 0 −1 0 ⎟
system, and for this purpose, the dq formulation has served as a ⎝ ⎠
very useful tool. These transient simulations would be valuable 0 0 0 0 0 1 −1 0 0
for different purposes such as short-circuit simulations, protec- 0 0 0 0 0 1 0 −1 0
tion arrangements, or substation control design.
ACKNOWLEDGMENT
A PPENDIX A
M ATRIX AND V ECTOR D IMENSIONS The authors would like to thank the CAF Company for their
support.
The dimensions of the matrices and vectors are listed as
follows. R EFERENCES
Λdc , Λtt , Λts , Λss ((nt + ns ), (nt + ns )). [1] C. Pires, S. Nabeta, and J. Cardoso, “ICCG method applied to solve DC
Γdc (ndc dc
B , nN ). traction load flow including earthing models,” Elect. Power Appl., IET,
ac
Γ (nB , nac
ac
N ).
vol. 1, no. 2, pp. 193–198, Mar. 2007.
L [2] Y. S. Tzeng, R.-N. Wu, and N. Chen, “Electric network solutions of DC
Γ (nL
B , (n dc ac
N + nN )). transit systems with inverting substations,” IEEE Trans. Veh. Technol.,
ac ac ac ac
R B , XB (nB , nB ). vol. 47, no. 4, pp. 1405–1412, Nov. 1998.
RdcB (ndc dc
B , nB ).
[3] Y.-S. Tzeng, R.-N. Wu, and N. Chen, “Unified AC/DC power flow for
system simulation in DC electrified transit railways,” Proc. Inst. Elect.
I (2nN , 2nac
ac
N ). Eng.—Elect. Power Appl., vol. 142, no. 6, pp. 345–354, Nov. 1995.
S (ndc dc
N , nN ). [4] L. Abrahamsson and L. Söder, “Fast estimation of relations between
M ((ndc ac dc ac
B + 2nB + nN + 2nN ), (nB +
dc aggregated train power system data and traffic performance,” IEEE Trans.
ac dc ac dc Veh. Technol., vol. 60, no. 1, pp. 16–29, Jan. 2011.
2nB + nN + 4nN + nN )). [5] Y. Cai, M. Irving, and S. Case, “Modeling and numerical solution of multi-
Γ, P, Q ((ndc L ac ac
B + nB + nB ), (nN + nN )).
dc
branched DC rail traction power systems,” Proc. Inst. Elect. Eng.—Elect.
ac ac ac
vN d , vN q (1, nN ). Power Appl., vol. 142, no. 5, pp. 323–328, Sep. 1995.
[6] C. Pires, S. Nabeta, and J. Cardoso, “DC traction load flow including AC
iac ac
Bd , iBq (1, nac
B ). distribution network,” Elect. Power Appl., IET, vol. 3, no. 4, pp. 289–297,
dc
vN (1, ndc
N ). Jul. 2009.
2430 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 61, NO. 6, JULY 2012

[7] D. Braunagel, L. Kraft, and J. Whysong, “Inclusion of DC converter and [19] P. Pozzobon, “Transient and steady-state short-circuit currents in recti-
transmission equations directly in a Newton power flow,” IEEE Trans. fiers for DC traction supply,” IEEE Trans. Veh. Technol., vol. 47, no. 4,
Power App. Syst., vol. PAS-95, no. 1, pp. 76–88, Jan. 1976. pp. 1390–1404, Nov. 1998.
[8] T. Smed, G. Andersson, G. Sheble, and L. Grigsby, “A new approach to [20] T. Steihaug, “The conjugate gradient method and trust regions in large-
AC/DC power flow,” IEEE Trans. Power Syst., vol. 6, no. 3, pp. 1238– scale optimization,” SIAM J. Numer. Anal., vol. 20, no. 3, pp. 626–637,
1244, Aug. 1991. 1983.
[9] Y.-S. Tzeng, N. Chen, and R.-N. Wu, “A detailed RL-fed bridge converter
model for power flow studies in industrial AC/DC power systems,” IEEE
Trans. Ind. Electron., vol. 42, no. 5, pp. 531–538, Oct. 1995. Pablo Arboleya (M’09) received the M.Eng. and Ph.D. (with distinction)
[10] S. Talukdar and R. Koo, “The analysis of electrified ground transportation degrees in electrical engineering from the University of Oviedo, Campus de
networks,” IEEE Trans. Power App. Syst., vol. PAS-96, no. 1, pp. 240– Viesques, Gijón, Spain, in 2001 and 2005, respectively.
247, Jan. 1977. He is currently an Associate Professor with the Department of Electrical En-
[11] J. Bialek, “Tracing the flow of electricity,” Proc. Inst. Elect. Eng.—Gener., gineering, University of Oviedo. For several years, he worked on the design of
Transm. Distrib., vol. 143, no. 4, pp. 313–320, Jul. 1996. electrical machines and fault detection. His research interests include microgrid
[12] G. Strbac, D. Kirschen, and S. Ahmed, “Allocating transmission sys- modeling and operation techniques and combined ac/dc power flow.
tem usage on the basis of traceable contributions of generators and
loads to flows,” IEEE Trans. Power Syst., vol. 13, no. 2, pp. 527–534,
May 1998.
[13] P. Murty, Power Systems Analysis. Hyderabad, India: BSP, 2009. Guzman Diaz (SM’12) received the M.Eng. and Ph.D. degrees in electrical
[14] J. Gross and J. Yellen, Handbook of Graph Theory. Boca Raton, FL: engineering from the University of Oviedo, Campus de Viesques, Gijón, Spain,
CRC, 2004. in 1993 and 2001, respectively.
[15] F. Milano, Power System Modelling and Scripting. New York: Springer- He is currently an Associate Professor with the Department of Electrical
Verlag, 2010. Engineering, University of Oviedo. His research interests include the design
[16] G. Diaz, C. Gonzalez-Moran, J. Gomez-Aleixandre, and A. Diez, “Com- and evaluation of induction motors, where he has worked on several research
posite loads in stand-alone inverter-based microgrids—Modeling proce- projects for ABB Automation Products, and distributed generation.
dure and effects on load margin,” IEEE Trans. Power Syst., vol. 25, no. 2,
pp. 894–905, May 2010.
[17] N. Pogaku, M. Prodanovic, and T. Green, “Modeling, analysis and testing Manuel Coto (M’11) was born in 1983 in Oviedo, Spain. He received the
of autonomous operation of an inverter-based microgrid,” IEEE Trans. M.Eng. degree from the University of Oviedo, Campus de Viesques, Gijón,
Power Electron., vol. 22, no. 2, pp. 613–625, Mar. 2007. Spain, in 2007. He is currently working toward the Ph.D. degree with the
[18] F. Briz, M. Degner, and R. Lorenz, “Analysis and design of current reg- Department of Electrical Engineering, University of Oviedo.
ulators using complex vectors,” IEEE Trans. Ind. Appl., vol. 36, no. 3, His research interests include power-flow and ac/dc power system modeling
pp. 817–825, May/Jun. 2000. and simulation.