Sie sind auf Seite 1von 8

Optimal Power Flow in Multiterminal HVDC

Systems Considering DC/DC Converters


Alejandro Garces, Danilo Montoya Raymundo Torres
Universidad Tecnologica de Pereira Norwegian University of Science and Technology
Pereira, Colombia Trondheim, Norway
Email: alejandro.garces@utp.edu.co, danilo.monotoya@utp.edu.co Email: raymundo@elkraft.ntnu.no

Abstract—This paper presents a non-linear model for the Area AC AC Area


a)
optimal power flow in multi-terminal high-voltage direct cur- 1 DC DC 2
rent transmission systems. The proposed methodology considers
exclusively the DC side of the grid and includes branches
with DC/DC converters. Losses of lines and DC/DC converters
are also considered in the optimization model. A semidefinite
aproximation is proposed in order to obtain an unique solution
for real time operation. The methodology is evaluated on a Area AC AC Area
b)
reduced version of the CIGRE B4 DC grid test system. Results 1 DC DC 2
demonstrate the proposed approximation is efficient and accurate
compared to the non-linear model solved in GAMS.
Index Terms—multiterminal HVDC, optimal power flow,
semidefinite programming

I. I NTRODUCTION Area AC AC Area


DC
3 DC
DC
DC 4
High voltage direct current transmission (HVDC) is a
mature technology that allows power transmission for long
Fig. 1. HVDC configurations. a) point-to-point configuration. b) Multiter-
distances as well as interconnection between unsynchronized minal HVDC system
networks. The most common HVDC configuration (i.e point-
to-point) comprises two converter stations connected by an
overhead power line or an undersea cable as depicted in Fig 1a.
DC/DC converters are included in the model since they involve
Nevertheless, future power systems will require meshed DC
additional control variables as well as non-linear equations
networks for applications such as SuperGrids [1] and offshore
related to power losses.
wind farms [2]. This type of DC grids, namely Multiterminal
Non-convex problems such as OPF lead to local optimums
HVDC systems (MT-HVDC), are viable due to the develop-
(i.e an optimum within a neighboring set of candidate so-
ment of high-power force-commutated semiconductor devices
lutions), in contrast to convex optimization problems which
and DC breakers [3].
Recent studies have demonstrated that a densely meshed allow to find the optimum among all possible candidates [6].
multiterminal HVDC grid provides advantages in terms of Former type of optimization problems include linear program-
efficiency and security [4]. However, these type of grids are ming, semidefinite optimization and quadratic optimization
more difficult to control than conventional point-to-point lines. with linear constrains.
They require a reliable communication between terminals and In particular, semidefinite programming (SDP) is concerned
optimization algorithms to find the best operating point. with problems in which the solution space includes cones of
In addition, highly meshed MT-HVDC systems require semidefinite matrices and the objective function is linear. The
DC/DC stations to interconnect HVDC systems with different optimal power flow is not a SDP problem. However, it is
nominal voltage or different configuration (e.g monopolar, bip possible to define an SDP approximation with a high degree
olar, homopolar). DC/DC stations can control active power on of accuracy. This approximation can be efficiently solved by
a particular HVDC line (see Fig 1b) and assist the MT-HVDC interior point methods. It also guarantee to find a global
during disturbances such as DC faults. optimum due to its convex nature. These two aspects (i.e
Classic methodologies for power system operation such efficiency and uniqueness in the solution) are relevant for real
as the optimal power flow (OPF), must be adapted to this time operation of multiterminal HVDC systems.
new context. Just like its counterpart AC [5], the optimal This paper studies a non-linear non-convex model for op-
power flow for multiterminal HVDC systems is a non-linear timal operation of multiterminal HVDC systems. It includes
and non-convex problem. This constitutes a challenge in both both AC/DC terminals and branches with DC/DC converters.
theory and practice. The degree of complexity is increased if A quadratic formulation for the losses in the DC/DC converters

978-1-5090-0873-5/16/$31.00 ©2016 IEEE 1212


is included in the optimization model resulting in two possible for multiterminal HVDC systems [13], today consensus among
optimization problems (i.e with and without losses in the the scientific community is that VSC and specially MMC are
DC/DC converters). An SDP approximation is proposed to the most prominent technologies for multiterminal grids [14].
solve the first model. The second problem is also convex and One of the main components of an MT-HVDC system is
hence, it can be solved although it is not strictly an SDP. the AC/DC converter which integrates the DC grid with the
Different SDP approximations have been proposed before neighbor AC network(s). Each converter injects or consumes
for a wide range of problems in power systems applications power from the MT-HVDC grid and controls two variables
including economic dispatch [7], optimal power flow in power independently. This is a key feature given by the forced
distribution systems [8], unit commitment [9] and hydrother- commutation technology. Hence, different operation modes
mal dispatch [10] among others. A semidefinite relaxation for can be obtained, for example, it could be favorable to control
the basic optimal power flow in DC grid was presented in [11]. the magnitude and angle of the AC voltage in offshore wind
However, that model did not consider the control and the losses farms. Nevertheless, most of converters in MT-HVDC systems
in the DC/DC converters nor the particular characteristics control one variable in the DC side and one variable in the
of multiterminal HVDC systems. To the knowledge of the AC side. Usually a terminal controls the power factor in the
authors, no SDP approximation has been studied for the AC side and the voltage or the power in the DC side.
optimal power flow (OPF) in multiterminal HVDC systems Notwithstanding its high controlability, AC/DC terminals
under the modeling presented here. are nodal devices from the point of view of the MT-HVDC
SDP relaxation for non-convex models is usually studied grid. This means, power injection but not power flow is
under two main scopes: accuracy and computational time. This controlled in each line. For this reason, it is expected future MT-
paper studies only the fist aspect. The main contribution of the HVDC systems include DC/DC converters among some
paper can be summarized as follows: branches in order to enhance controllability. This new control
• A general model for the losses in DC/DC converters.
variable must be considered in the power flow formulation.
• A model for the optimal power flow in multiterminal
A. Modelling DC/DC converters
HVDC systems without consider the AC side.
• A semidefinite relaxation for the proposed model that
As aforementioned, DC/DC converters are relevant for MT-
guarantee a global optimum (at least for the relaxed HVDC system since they allow to control the power through
problem) branches. There are different topologies of DC/DC converters,
• A comparison between the proposed relaxation and a non-
most of them are high power devices and require a coupling
linear method implemented in GAMS (General Algebraic transformer in order to grant galvanic isolation [15]. They are
Modeling System [12]) in fact, a three-stage converters (i.e DC/AC, transformer an
The remainder of this paper is organized as follows: Sec- AC/DC). These converters with galvanic isolation allow high
tions II presents a general background of multiterminal HVDC efficiency at high-power/high-voltage levels.
systems. Model and implications of the DC/DC converters are Losses in this type of converters, as well as in any forced
also discussed there. Then, Section III reviews briefly the main commutated converter, are classified in conduction and switch-
aspects of semidefinite programming. Next, in Section IV the ing losses. Both are highly dependent of the intrinsic charac-
OPF model for MT-HVDC grids is analyzed. After that in teristics of the semiconductor devices and passive elements as
Section V, the proposed SDP model is developed followed by well as operative variables such as the switching frequency
results and conclusions (Sections VI and VII). and the type of modulation. For this study the total losses
are assumed to be quadratic related to the average transfer
II. M ULTITERMINAL HVDC SYSTEMS power. This assumption is based on the results presented by
An MT-HVDC is an efficient alternative for applications Erikson in [16]and Zhou in [17]. However, new topologies
such as offshore wind energy, supergrids and urban distribution for DC/DC converters are continuously proposed and hence, a
systems in high dense areas of the city. This technology study in depth of this issue is required in order to generalize this
not only allows cost and loss reduction but also enhances consideration.
reliability and security. Let us consider a DC/DC converter placed between nodes
HVDC systems can be classified according to the type k and m. Transferred instantaneous powers qkm and qmk are
of semiconductor devices used in the converters. There are also considered with the directions given in Fig 2. One variable
two main semiconductor technologies for HVDC applications, is positive and the other is negative, according to the direction
namely line and force commutation. The first technology is of the current.
used for the line commutated converters (LCC), also called For modeling proposes, only one direction of the power

current source converters (CSC)1 , while the second is used flow is used (i.e qkm ) since the other is equal in magnitude
for the voltage source converters (VSC) and the modular mul- and opposite in direction.
tilevel converters (MMC). Although LCC has been proposed
1 We prefer the term LCC instead CSC since the former can be confused qkm + qmk = 0 (1)
with the PWM-CSC which is a current source converter with forced commu-
tation Losses are modeled as two loads sk and sm with
1213
k DC m IV. OPF IN M ULTITERMINAL -HVDC SYSTEMS
qkm DC qmk Among this section, capital letters represent matrices or vec-
tors, calligraphic letters represent sets and lower-case letters
sk = f (qkm ) sm = f (qmk ) represent uni-dimensional constants or variables.

Let us consider a multiterminal-HVDC system as a set of


Fig. 2. Model of a DC/DC converter connected between nodes k and m buses represented by N = {1, 2, ..., N } an a set of AC/AC
terminals T ⊆ N . HVDC lines are represented by a set E =
{(k, m)} ⊆ N × N and the DC nodal admittance matrix
1 G ∈ Rn×n which is symmetric and positive semidefinite with
· α + β · kqkm k + γ · kqkm k
2
sk = (2) entries [G]k,m = gkm . DC/DC devices are represented by their
2
where α,β and γ are loss factors calculated under the respective branches D = {(k, m)} ⊆ N × N where E ∩
following assumptions: D = φ. These devices must be consider apart from the AC/DC
terminals as they allow to control the power flowing through
• Frequency and modulation technique are constant
the branch. For each bus k, there are two decision variables
• The converter is operated in stationary state
vk and pk . The first variable (vk ) is the nodal voltage while
• The voltage is close to 1 pu
the second variable (pk ) is the power injected by an AC/DC
Notice that sk equals sm and represents half of the power terminal (k ∈ T ). There is a third decision variable (qkm )
loss (i.e sloss = sk + sm ≥ 0). This model is similar to a shift which represents the power controlled by a DC/DC device
transformer in conventional power flow for AC systems. This (km ∈ D). In addition, the power flow in each HVDC line
is because both type of devices have equivalent functions (i.e is represented for a double subindex fkm (with km ∈ E ). In
to control the flow of active power in the line km).
general, fkm = fmk .
III. S EM IDEFINITE PROGRAM MING The multiterminal optimal power flow can be formulated
Semidefinite programming (SDP) is a field of the mathe- as minimizing the power losses in the grid pL and the power
matical optimization which is growing in interest in recent losses in the converters (sloss ), considering the transmission
years due to its theoretical and practical implications. An constrains as follows:
SDP problem has similar characteristics as linear programming

in the sense that both have strong theoretical foundation P


n P
N P
and are efficiently solvable using interior point methods. In Minimize pL = gkm · vk · vm + sk (4)
p,q,v k=1 m=1 (k)∈D
addition, it is possible to solve problems with hundreds or
Subject to v1 = 1 (5)
thousands of variables in few seconds. Therefore, it is a
P
n P
n
suitable methodology for power systems applications where pk − sk − akm · qkm = gkm · vk · vm (6)
real-time tertiary control is required. It is reasonable to ensure m=1 m=1
P
that any SDP model is easier to solve compared to any non- 2sk = γkm · qkm
2
+ βkm · kqkm k + αkm(7)
(km)∈D
convex problem, no matter the number of variables. More 2

v −v ·v
importantly, being a convex problem, it is possible to claim fkm = k r kmk m (8)
the optimum is global. vmin ≤ vk ≤ vmax (9)
A semidefinite programming model is an optimization prob-
pk(min) ≤ pk ≤ pk(max) (10)
lem of the form given in (3)
−fkm(max) ≤ fkm ≤ fkm(max) (11)
Minimize T race(C · X )
X −qk(max) ≤ qk ≤ qk(max) (12)
Subject to A·X = B (3)
X D0 Equation (5) represents the voltage in the slack node2 while
(6) represents the balance of nodal power. Equation (7) takes
where X represent the decision variables, C the losses (or (i.e X ∈ Rn×n ). It must be also symmetric and positive
cost according the objective function) and A, B are matrices semidefinite. This is represented by the symbol D.
which represent particular operative constraints. Notice the It is possible to develop SDP approximations for non-
structure of this model is quite similar to a linear programming linear/non-convex problems as is the case of the OPF. This is
problem. In fact, a linear programming model can be rewritten presented in the following section.
as an SDP. The same is true for some quadratic problems
[18]. However, unlike linear programming and quadratic pro-
gramming, the optimization variable is a matrix and not a vector
1214
into account the balance of power in each DC/DC device
neglecting the losses in the converter (i.e the power qkm in
the primary of the DC/DC converter is equal to the power in
the secondary qmk ). Losses in the converters are given by the
aforementioned quadratic model. The rest of the constrains
represent the operative limits of voltages and powers. The ob-
jective function is basically power losses but other objectives
can be effortless included, for example operative costs. Active
power on the AC/DC terminals must be on an interval given
by (10). However, these limits are fixed in some cases. Two
2 We are considering master slave control and hence, droop controls are

not considered in the optimization model

1215
simple examples are step node where pk(max) = pk(min) = 0 In general, the result of the optimization of this SDP model
and offshore wind farms where pk(max) = pk(min) = pwind generates a rank N matrix. Therefore it is required to find a
(i.e the main objective is to obtain the power available for the rank one matrix in order to recuperate the vector V . For doing
wind). In the former case, a DC/DC converter is important since so, an eigen-decomposition of the matrix X is used:
it is the only element available to control the power N

flow. X k
X = λk · Wk · W T (24)
This problem is basically non-linear and non-convex since
k=1
(6) and (7) are not-affine equality constrains. Therefore, a

global solution cannot be guaranteed by conventional algo- where λ are the eigenvalues and W are its respective
rithms. In addition, it is required a fast and reliable algorithm eigenvectors. It is expected that N − 1 eigenvalues are close
to be implemented in the real time operation of the MT-HVDC to zero. Therefore, a rank 1 approximation of the matrix can
system. be obtained as:

V. S EM IDEFINITE RELAXATION X ≈ λm · (Wm · W m


T
) (25)
Let us consider firt the case where losses of the DC/DC where λm is the maximum eigenvalue (i.e the only eigen-
converters can be neglected. In this case, the optimal power value that is not close to zero). Under this approximation it is
flow for MT-HVDC systems can be rewritten as a general SDP possible to recover the vector of voltages:
model by defining a matrix X as follows
p
V ≈ λm · Wm (26)
X = V ·VT (13)
The main advantage of this methodology is its efficiency
where V ∈ Rn is a column vector with entries vk . Evidently and accuracy as will be demonstrated in the next section.
X is a rank one symmetric positive semi definite matrix. With Losses in the DC/DC converter can be included directly in
this definition, quadratic constrains can be rewritten leading a the approximation since the objective function is still convex:
completely equivalent model as follows:

Minimize pL = T race(G · X ) + S T · êN (27)


Minimize pL = T race(G · X ) (14) P,X,Q

P,X,Q Subject to X1,1 = 1 (28)


Subject to X1,1 = 1 (15)
P − S − AT Q = Diag(G · X ) (29)
D ·
P − ATD· Q = Diag(G · X ) (16)
S ≥ 12 ATD (Γ · (Q ◦ Q) + B · Abs(Q)) (30)
F = M · vec(X ) (17)
F = M · vec(X ) (31)
(v 2 ) · JˆN ≤ X ≤ (v 2 ) · JˆN (18)
min max ) · JˆN ≤ X ≤ (vmax ) · JˆN (32)
(vmin
2 2
Pmin ≤ P ≤ Pmax (19)
P ≤P ≤P (33)
min max
−Qmax ≤ Q ≤ Qmax (20)
−Qmax ≤ Q ≤ Qmax (34)
−Fmax ≤ F ≤ Fmax (21)
−Fmax ≤ F ≤ Fmax (35)
X D0 (22)
X D0 (36)
Rank(X ) = 1 (23)
where Abs(Q) is the element-wise absolute value of Q
where P , Q and F are column vectors with entries [P ] = and (◦) is the Hadamard product (i.e the entrywise product
(pk ), Q = (qkm ) and F = (fkm ) respectively. AD and AE are or array multiplication). Γ = diag(γ) and B = diag(β)
reduced incidence matrix (i.e incidence matrix which includes are the loss factors of the DC/DC converters. ATD is the
exclusively the DC/DC converters or the transmission lines).
efficiently solved by the interior point method [6]. Notice that
JˆN ∈ Rn×n is an all-ones matrix and Diag(·) is the main the SDP model increases the number of variables. However, an
diagonal of the corresponding matrix. M is a matrix which is increased number of variables does not imply an increasing in
constructed in such a way that (8) is fulfilled. the computational time. In addition, the particular sparse struc-
At this point, both models are completely equivalent (both ture of the matrices involved in the model allows improvement
are non-convex). Notice the only non-convex equation is the on the efficiency.
rank constraint (23). Consequently, an SDP is obtained by
relaxing this constraint. The resulting SDP problem can be
1216
undirected reduced incidence matrix. The parameters α do
not affect the optimization model since they are constant. As
aforementioned, this model is still convex and hence it can
be easily solved by interior point algorithm with a global
optimum guaranteed.
VI. R ESULTS
The proposed models were tested on a reduced version of
the CIGRE B4 bipolar DC grid test system depicted in Fig
3. It consists on eight nodes, ten lines, five terminals AC/DC
and one DC/DC converter. Hence, it covers a wide range of
components present in an MT-HVDC system.
Parameters of the system were taken from [19] and [20].
For the sake of completeness, they are presented in Tables I

1217
AC DC
DC AC 1.6245 × 10−15
1 4
10−15
3.1638 ×
6.5970 × 10−15
λ= (37)
1.1783 × 10−14
DC 3.2187 × 10−14
AC
5 6.0009

Based on these results, it could be conjectured that the SDP


is an exact relaxation of the original model. However, further
research is required to generalize these results.
The original non-convex optimization model given by (4)-
6 3 7 (12) was solved in Gams [12] in order to evaluate the accuracy
AC
DC
of the approximation. The solution given by the SDP model
DC
DC
was used as initial point for the optimization solver. Results
are given in Table III
8
2 TABLE III
DC
C O M PA R I S O N B ET WE E N T HE NO N-LIN E A R M OD E L AN D T HE SDP
AC
A P P ROX I M AT I O N

Node V (SDP) V (GAMS) Error


Fig. 3. Reduced CIGRÉ bipolar DC grid test system
1 1.0000 1.0000 0.0000
2 1.0082 1.0080 0.0003
and II. The vectors Pmax and Pmin were obtained according 3 1.0029 1.0030 -0.0001
to the desired interchange with the neighboring AC grids. The 4 0.9934 0.9930 0.0004
5 0.9927 0.9930 0.0003
terminal 1 is the slack node (v1 = 1) while the available 8 1.0031 1.0030 0.0001
voltage in the other terminals are between 0.9 and 1.01 pu.
The maximum capacity of DC transmission lines were defined The optimal value of power loss in the relaxed model was
in 1.0pu. Nodes 6 and 7 were eliminated by a Kron reduction. pL = 0.0099 pu while for the model solved in GAMS was
TABLE I pL = 0.010 pu. The model in GAMS took 56 iterations
PA R A M E T E R S O F T H E HVDC L I N E S while the SDP model in CVX took 14. As expected, the
approximation is faster than the non-convex model with the
From To rkm (Ω)
additional advantage of guarantee a global optimum (at least for
2 6 3.42/2
1 6 5.70 the approximated model). The power flowing through the
6 3 2.28 DC/DC converter was 349 M W in the direction 3 → 8.
1 3 4.56/2
1 4 1.90 A second simulation was performed, this time considering
4 5 1.90 the losses on the DC/DC converter. Loss parameters were
5 7 2.85
8 7 1.90 assumed γ = 0.03, β = 0.05 and α = 0 for q given in
per unit. The model was solved in CVX and compared to the
non-linear model solved in GAMS. Results are given in Table
TABLE II IV
PA R A M E T E R S OF THE HVDC C O N V E RT E R S As in the previous case, the approximation is very accurate
compared to the non-linear model (assuming the solution
Node Type Pnom (M W ) Vnom (kV ) Pmin Pmax
obtained in GAMS correspond to the global optimum). Total
1 AC/DC 2400 400 100 300
2 AC/DC 2400 400 300 500 power loss in the grid was 0.0131 pu while loss in the
3 AC/DC 2400 400 400 600 converter was qL = 0.0068 pu. The new value of the power
4 AC/DC 800 400 -700 -500
5 1600 400 -600 -400 thorough the DC/DC converter was reduced to 13 M W due
AC/DC
3-8 DC/DC 2000 400 -2000 2000
TABLE IV
The semidefinite relaxation (27)-(22) for the case in which C O M PA R I S O N B ET WE E N T HE NO N-LIN E A R M OD E L AN D T HE CO N VE X
A P P ROX I M AT I O N C O N S I D E R I N G L O S S E S I N T H E DC/DC C O N V E RT E R S
the losses of the DC/DC converter are neglected was solved
using CVX, a package for specifying and solving convex Node V (SDP) V (GAMS) Error
programs [21] [22]. The resulting matrix X was full rank 1 1.0000 1.0000 0.0000
meaning it is not exact solution of the original problem. 2 1.0100 1.0010 0.0000
3 1.0053 1.0050 0.0003
Nonetheless, all except one eigenvalues are close to zero as 4 0.9908 0.9910 -0.0002
given in (37). Consequently, the approximation given by (25) 5 0.9877 0.9880 -0.0003
8 0.9918 0.9920 -0.0002
is very accurate.

1218
TABLE V
C O M PA R I S O N O F P OW E R S FL OW S U S I N G D I FF E R E N T VA L U E S O F T H E
R EFERENCES
T R A N S M I S S I O N L I N E C A PAC I T I E S [1] D. V. Hertem and M. Ghandhari, “Multi-terminal vsc-hvdc for the
european supergrid: Obstacles,” Renewable and Sustainable Energy
Node F with fmax = 1pu F with fmax = 0.5pu Reviews, vol. 14, no. 9, pp. 3156 – 3163, 2010.
2-6(x2) 0.3000 0.3000 [2] T. Haileselassie and K. Uhlen, “Power system security in a meshed north
6-2(x2) -0.2991 -0.2991 sea hvdc grid,” Proc. of the IEEE, vol. 101, no. 4, pp. 978–990, Apr
1-6 -0.1916 -0.1247 2013.
6-1 0.1929 0.1253 [3] W. Wang, M. Barnes, O. Marjanovic, and O. Cwikowski, “Impact of dc
6-3 0.1062 0.1738 breaker systems on multiterminal vsc-hvdc stability,” Power Delivery,
3-6 -0.1060 -0.1733 IEEE Tran. on, vol. PP, no. 99, pp. 1–1, 2015.
1-3(x2) -0.3735 -0.1388
3-1(x2) 0.3755 0.1391 [4] M. Bucher, R. Wiget, G. Andersson, and C. Franck, “Multiterminal hvdc
1-4 0.7851 0.5000 networks: What is the preferred topology?” Power Delivery, IEEE Tran.
4-1 -0.7778 -0.4970 on, vol. 29, no. 1, pp. 406–413, Feb 2014.
4-5 0.2778 -0.0030 [5] J. Hu, J. Cao, and T. Yong, “Multi-level dispatch control architecture
5-4 -0.2768 0.0030 for power systems with demand-side resources,” IET Generation, Trans-
5-7 -0.1232 -0.4030 mission Distribution, vol. 9, no. 16, pp. 2799–2810, 2015.
7-5 0.1234 0.4059 [6] S. Boyd and L. Vandenberghe, Convex Optimization. New York, NY,
8-7 0.1236 0.4078 USA: Cambridge University Press, 2004.
7-8 -0.1234 -0.4059
[7] A. Jubril, O. Komolafe, and K. Alawode, “Solving multi-objective eco-
nomic dispatch problem via semidefinite programming,” Power Systems,
IEEE Tran on, vol. 28, no. 3, pp. 2056–2064, Aug 2013.
[8] E. Dall’Anese, H. Zhu, and G. Giannakis, “Distributed optimal power flow
for smart microgrids,” Smart Grid, IEEE Tran. on, vol. 4, no. 3, pp.
to the inclusion of the losses of the converter. This shows the 1464–1475, Sept 2013.
importance of including converter losses. [9] X. Bai and H. Wei, “Semi-definite programming-based method for
security-constrained unit commitment with operational and optimal
Line flow constraints have influence on the results. To power flow constraints,” Generation, Transmission Distribution, IET,
demonstrate this, a third simulation was performed with dif- vol. 3, no. 2, pp. 182–197, February 2009.
ferent capacity of the transmission lines. Results are given in [10] Y. Zhu, J. Jian, J. Wu, and L. Yang, “Global optimization of non-convex
hydro-thermal coordination based on semidefinite programming,” Power
Table V. In this case, fmax = 0.5 and the total losses became Systems, IEEE Tran. on, vol. 28, no. 4, pp. 3720–3728, Nov 2013.
0.0364pu. The flow in the DC/DC converter was increased to [11] L. Gan and S. Low, “Optimal power flow in direct current networks,”
0.4210pu. in Decision and Control (CDC), 2013 IEEE 52nd Annual Conference
on, Dec 2013, pp. 5614–5619.
All matlab simulations are available in [23]. [12] GAMS - The Solver Manuals, GAMS Release 24.2.1, GAMS
Development Corporation, Washington, DC, USA, 2013. [Online].
VII. C ONCLUSIONS Available: http://www.gams.com/dd/docs/solvers/allsolvers.pdf
[13] D. Chapman, J. Davies, J. McNichol, E. Gulachenski, S. Doe, and
A non-linear model for the optimal power flow in multi- N. Balu, “Test experience with multiterminal hvdc load flow and stability
terminal high-voltage-direct-current transmission systems was programs,” Power Delivery, IEEE Tran. on, vol. 3, no. 3, pp. 1173–1181,
Jul 1988.
presented. The model included ohmic losses on the DC [14] W. Leterme, P. Tielens, S. De Boeck, and D. Van Hertem, “Overview
transmission lines and power loss on the DC/DC converters. of grounding and configuration options for meshed hvdc grids,” Power
Two convex approximations were developed for the non- Delivery, IEEE Tran. on, vol. 29, no. 6, pp. 2467–2475, Dec 2014.
[15] N. Soltau, R. U. Lenke, and R. W. D. Doncker, High-Power DC-DC
linear model, both considering and neglecting the power losses Converter. E.ON Energy Research Center Series, 2015.
on the converters. The first approximation was a semidefinite [16] B. Arbetter, R. Erickson, and D. Maksimovic, “Dc-dc converter design
programming model while the second is an optimization prob- for battery-operated systems,” in Power Electronics Specialists Confer-
ence, 1995. PESC ’95 Record., 26th Annual IEEE, vol. 1, Jun 1995, pp.
lem with semidefinite constrains but still a convex problem. 103–109 vol.1.
The non-linear models were also solved using GAMS. The [17] Y. Zhou, D. Macpherson, W. Blewitt, and D. Jovcic, “Comparison of
proposed approximations gave an error of less than 0.3% dc-dc converter topologies for offshore wind-farm application,” in Power
Electronics, Machines and Drives (PEMD 2012), 6th IET International
in voltages for the case in which the losses in the DC/DC Conference on, March 2012, pp. 1–6.
converter were neglected and 0.4% for the case in which they [18] Z.-Q. Luo, W.-K. Ma, A.-C. So, Y. Ye, and S. Zhang, “Semidefinite
were considered. For practical proposes, these results can be relaxation of quadratic optimization problems,” Signal Processing Mag-
azine, IEEE, vol. 27, no. 3, pp. 20–34, May 2010.
considered as the exact solution. [19] C. Gavriluta, I. Candela, C. Citro, A. Luna, and P. Rodriguez, “Design
Future research includes the approximation for the complete considerations for primary control in multi-terminal vsc-hvdc grids,”
network including the adjacent AC grids as well as an detailed Electric Power Systems Research, vol. 122, pp. 33 – 41, 2015.
[20] T. K. Vrana, Y. Y. D. Jovcic, S. Dennetiere, J. Jardini, and H. Saad, “The
analysis on the impact of other power electronic components cigre b4 dc grid test system,” in Documents related to the development
on the optimization model. Dynamic constraints, such as rotor of HVDC Grids, 2011.
angle oscillations could be included in that case. [21] I. CVX Research, “Cvx: Matlab software for disciplined convex pro-
gramming, version 2.0,” http://cvxr.com/cvx, Aug. 2012.
SDP relaxation for non-convex models is usually studied [22] M. Grant and S. Boyd, “Graph implementations for nonsmooth convex
under two main scopes: accuracy and computational time. This programs,” in Recent Advances in Learning and Control, ser. Lecture
paper was focused in the fist aspect but computation time is an Notes in Control and Information Sciences, V. Blondel, S. Boyd, and
H. Kimura, Eds. Springer-Verlag Limited, 2008, pp. 95–110.
important feature that must be studied in detail. Some authors [23] A. Garces. (2016, May 20) Matlab cen-
have claimed that SDP solvers are not very scalable. However, tral file exchange. [Online]. Available:
the proposed model has some characteristics that could be used http://www.mathworks.com/matlabcentral/profile/authors/3009175-
alejandro-garces
to create an efficient algorithm for this particular problem.

1219