Abstract
This research discusses the application of dierent formulations to solve
the DC optimal power flow problem. The authors have programmed the
optimization algorithm using a primaldual interior point solver based on
the quadratic formulation and nonlinear constraints. The classical DC optimization problem is reformulated using the power transmission distribution
factors. The main advantage of this formulation is the problem dimension reduction. The formulation proposed has been implemented and exhaustively
validated with dierent power systems test and Chilean power system. The
results of the proposed algorithm demonstrate both the feasibility of applying
the methodology to the DC optimal power flow problem and, above all, the
potential applicability of the approach in largescale power systems. This
is a very important factor when the planners and operators need to solve
many simulations per hour considering dierent topologies and conditions in
real and largescale power systems. The large quantity of simulations conducted by the authors to solve the DCOPF problem validates the formulation
proposed in this study.
Preprint submitted to Elsevier
Nb
number of buses.
NG
NL
C(Pi ) :
Pi
Pimin
Pimax
rj
Pjloss
fj
fjmax
xij
Pk
Pkd
A, Ar
Xmin , Xmax
Variable bounds.
2. Introduction
The concept of optimal power flow (OPF) was first proposed by Carpentier [1] in the early 1960s based on the economic dispatch problem. The OPF
problem has been largely discussed in the specialized literature. A wide range
of models have been developed and adopted to formulate dierent kinds of
optimization problems through dierent objective functions, dierent sets of
decision variables and dierent kinds of constraints [25].
The OPF problem is concerned with the optimization of a steadystate
power system performance, subject to various equality and inequality constraints. The OPF analysis aims at determining the power system operation
state according to cost, planning or reliability criteria without violating system and equipment operating limits. It is one of the most intensely used
tools in many power engineering applications as far as network optimization,
voltage control, state estimation, generation and transmission planning and
markets studies are concerned [6].
2.1. Direct Current (DC) Optimal Power Flow
The problem involved in solving an AC power flow can be illustrated by
the use of direct current circuit. A DCOPF problem is an approximation
for an underlying full alternating current network (ACOPF) under several
simplifying restrictions regarding voltage magnitudes, voltage angles, admittances, and reactive power. When it is not feasible to run a full ACOPF
problem due to time constraints, computing power, or lack of a robust solu
DCOPF problem. The main dierence between the three formulations is the
number of decision variables in the optimization problem. The first formulation is the classical DCOPF formulation which uses the power generation
and the voltage phase angles as decision variable. Based on expansion planning formulations, the authors have proposed a second DCOPF formulation
using the power generation, the voltage phase angles and the power flows on
the transmission lines as decision variables resulting in a significantly larger
optimization problem, so it is used to compare and validate the performance
of the algorithm. Finally, the authors have achieved an ecient formulation
using the power transmission distribution factors (PTDF) and the Y bus
matrix, so the network balance constraints are transformed in a global power
balance equation, and the transmission limit constraints are a function of the
distribution factors and the net power injection. With this transformation,
there is a very important reduction in the dimension of the optimization
problem.
The DCOPF approach using the PTDFbased formulation solves the optimization problem taking into account only the active power generation of
each unit as decision variables; this formulation does not depend on the voltage bus angles. Therefore, the main advantage of this formulation is the
reduction of the solution space, so the algorithm can speed up the convergence process. This is a very important factor, because in each electricity
control room (ISO), the OPF problem and/or approximation must be solved
many times per day (for instance every 5 minutes), and the solution space
could be very complex considering the dimension of the transmission network
(number of buses) especially with largescale power system. An added bene
fit of using the distribution matrix would be loss estimation. The drawback
of this approach is that the solution does not compute the voltage bus angles
to draw some conclusions about the system, but using the nodal admittance
matrix Ybus , it is possible to quickly calculate these angles.
Advances in algorithms are changing linear programming and a new class
of algorithm is emerging. Changes began in 1984 when N. Karmankar published his paper [7] introducing interiorpoint methods to solve linear and
nonlinear programming problems. In [8], the authors compare the interiorpoint method with classic simplex method solved with M inos. The results
in that study show the interiorpoint method is not as eective on smaller
dimension problems, but as problem get larger, it becomes more eective
than M inos.
In this study, the authors have decided to use the interiorpoint algorithm to solve the optimization problem, because this algorithm has no problems solving the quadratic objective function and nonlinear constraints. The
methodology has been implemented and exhaustively validated with dierent
power systems. In order to evaluate the solution obtained by the algorithm,
the authors have used M atP ower. Moreover, the DCOPF computed by
M atP ower uses the primaldual interior point as well [9].
It is quite evident from the comparisons that the proposed methodology provides the better formulation. The results of the proposed algorithm
demonstrate both the feasibility of applying the methodology to the DCOPF
problem and, above all, the potential applicability of the approach in medium
and largescale power systems. Finally, a realistic Chilean case has been successfully analyzed in this study.
Z = min
{N
G
}
C(Pi )
(1)
i=1
Pk Pkd B k = 0
(2)
Pimin Pi Pimax
(3)
4. Transmission limits: The power flow through the jth transmission line
must be within its capacity limits (thermal limit constraints).
fjmax
i j
xij
fjmax
(4)
The decision variables in this optimization problem are the voltage angles
and the active power generation of each unit. In the above formulation, the
voltage bus angles should be constrained.
3.2. Second DCOPF formulation
For instance in expansion planning problems, there is another formulation proposed in the technical literature to consider the network balance constraint. In that modeling, the network balance constraint is divided using the
Kirchos current law constraint, and the Kirchos voltage law constraint.
In the planning formulation the voltage and current constraint are nonlinear,
but in the DCOPF formulation these constraints are linear due to the integer
decision variables are not considered in the modeling.
This formulation using the current and voltage constraints has the same
objective function, and it may be represented by the following equations:
8
S f + Pk Pkd = 0
(5)
fj (i j ) = 0
(6)
fjmax fj fjmax
(7)
3. Transmission limits:
The decision variables in this optimization problem are the voltage angles,
the power flows on the transmission lines and the active power generation.
The main disadvantage for this formulation is the growth of the solution
space, so the authors use this formulation to compare the modeling proposed
in this study.
4. DCOPF formulation using the power transmission distribution
factors
The DC power flow model can also be used to compute the sensitivities of
branch flows to changes in nodal real power injections, called generation shift
9
(8)
(N
b
i=1
)
d
P +P
loss
NG
(
)
P D + P loss = 0
(9)
i=1
k=1
The power generated by the thermal units must be able to meet the
power load plus the transmission losses. In the DC model, the method
must estimate the system losses. Through the considerations made in
a DC power flow [11], equation (10) could be developed to estimate the
power losses using the power transmission distribution factors.
P loss =
NL
rj fj2 =
j=1
NL
rj (P T DF (P P d ))2
(10)
j=1
2. Transmission limits: The power flow through the jth transmission line
can be computed using the power transmission distribution factors, so
the line flow constraints are just a product of the PTDF matrix and
the net power injection.
10
fjmax P T DF (P P d ) fjmax
(11)
min f (X)
X
(12)
subject to:
G(X) = 0
(13)
H(X) 0
Xmin X Xmax
11
(14)
where the linear and nonlinear constraints are incorporated into G(X)
and H(X) constraints (13).
Firstly, slack variables Z are added to inequality constraints to convert
them to equality constraints. Then nonnegative constraints on slack variables
are eliminated by placing them in a barrier objective function:
[
f (X)
min
X
ni
]
ln(Zm )
(15)
m=1
subject to:
G(X) = 0
(16)
H(X) + Z = 0
Z>0
As the parameter of perturbation approaches zero. For a given value
of , the Lagrangian for this equality constrained problem is the following:
ni
ln(Zm )
(17)
m=1
Taking the partial derivatives with respect to each of the variables yields
the following equations:
T
LX (X, Z, , ) = fX + GTX + HX
(18)
LZ (X, Z, , ) = eT [Z] 1
(19)
L (X, Z, , ) = GT (X)
(20)
12
L (X, Z, , ) = H T (X) + Z T
(21)
LXX
(X, Z, , ) = fXX + GXX () + HXX ()
(22)
Z>0
>0
where:
LX T
[ ]
Z e
F (X, Z, , ) =
G(X)
H(X) + Z
fxT + T GX + T HX
[ ]
Z e
G(X)
(24)
H(X) + Z
13
[
FX FZ F
LXX
GX
HX
GTX
] Z
= F (X, Z, , )
F
T
HX
LXX T
X
[ ]
Z e
Z
Z
=
G(X)
0
H(X) + Z
0
(25)
(26)
(27)
(28)
Then, substituting (27) and (28) into the first row of (26) results in:
T
T
(LXX + HX
[Z]1 []HX )X + GTX + HX
[Z]1 (e + []H(X)) = LX T
(29)
Defining the matrix M and N as follows:
14
T
M LXX + HX
[Z]1 []HX
T
M = fXX + GXX () + HXX () + HX
[Z]1 []HX
T
N LX T + HX
[Z]1 (e + []H(X))
T
[Z]1 (e + []H(X))
N = fXT + T GX + T HX + HX
(30)
(31)
Combining (33), (34) and the third row of the (26) results in a system
of equations of reduced size (32), so the KarushKuhnTucker conditions can
be transformed into the following borderblocked system:
The KarushKuhnTucker conditions can be transformed into the following borderblocked system:
GTX
GX
{z
K
N
G(X)
(32)
The inverse of the matrix K is the main step to compute the decision
variables (X) and the Lagrangian multipliers (). Defining the matrix
M and N as follows:
T
M = fXX + GXX () + HXX () + HX
[Z]1 []HX
(33)
T
N = fXT + T GX + T HX + HX
[Z]1 (e + []H(X))
(34)
5.3. Algorithm
The Newton update can then be computed in the following three steps:
1. Compute X and using (32)
15
,1
m <0
Zm
(35)
(
(
) )
m
min
d = min m <0
,1
m
(36)
(37)
+ d
The parameter is a constant scalar with a value slightly less than one.
The scalar is a safety factor to guarantee that the next point meets the strict
positivity conditions, with typical value of 0.99995.
In the algorithm, during the Newtonlike iterations, the perturbation parameter must converge to zero in order to satisfy the first order optimality
16
ZT
ni
(38)
on distribution factors to model the transmission constraints. The solution for the optimization problem does not converge. In Table 1,
it is shown the convergence criterions used in the implementation of
the interiorpoint method: a) feasibility condition (f eascond), gradient condition (gradcond), complementary condition (compcond), cost
condition (costcond). The solution converges when these variables are
lower than 106 . The variables f eascond, gradcond, compcond and
costcond are used by M atP ower to analyze the convergence process as
well, so we can compare the quality of the solution (optimal solution).
In addition, the gamma value is included in the Table in order to give
information about the convergence process.
During the Newton iterations, the perturbation parameter called gamma
() must converge to zero in order to satisfy the first order optimality
conditions of the primal problem. In the algorithm programmed by the
authors, there is a nonconvergence condition (1/252 ). The limit
value represents the spacing of floating point numbers used by M atlab.
The process stops when, after searching in the 8th iteration, its value
is greater than the limit value.
The authors have verified the PTDFbased formulation increasing the
capacity of the transmission lines to 200M W . The optimization algorithm converged in 9 iterations (f eascond = 0.00, gradcond = 1.37 x
1016 , compcond = 2.86 x 1007 and costcond = 3.82 x 1010 ).
The optimal cost of both solutions is 2570[$/h], and the power generation is P1 = 150M W and P2 = 200M W . The results are also compared
18
with the interiorpoint algorithm used by M atP ower, and both solutions are the same  the optimal solution.
6.2. DCOPF formulation applied to Garvers system
This system represents the Garvers case in the ninth period [14]. The
transmission system is based on the actual topology, and the authors have
included in the transmission network one circuit in 35 rightofway, one
circuit in 46 , and three circuits in 26. The expansion formulation uses
artificial generation (Ai ) in bus 2, 4 and 5 to model the loss of load. Notice
that the optimization problem is always feasible due to the presence of these
generators, whose generation estimates the power loss of the customers.
In Table 3, it is shown the convergence values for each generator using the
classical formulation (M1) and the PTDFbased formulation (M3) to model
the DCOPF problem.
Using the PTDFbased formulation, the authors have solved the problem
in 14 iterations, meanwhile the classical formulation solved the problem in
15 iterations; the algorithm needs one more to achieve convergence. The
optimal cost of both solutions is 19139.75[$/h], which was validated with the
interiorpoint algorithm used by M atP ower. In the power flow solution, it
is verified is verified that none of the power units and transmission elements
are overloaded. In addition, there is no loss of load in bus 2, 5 and 5.
In the power flow solution, the power flow in the transmission line 23 is
100M W . In this case, the system has an active transmission constraint, so
the test is used to demonstrate the performance of the PTDFbased formulation when there is an active transmission constraint in the network.
19
considering the PDTFbased formulation ; and 4) 0.0067s considering the application of M atP ower. It is shown that the simulation time can be reduced
with the formulation proposed in this study.
One important characteristic of the interiorpoint method applied to these
power systems is that the number of iterations depends very little on the
problem size.
Lastly, in M atP ower, there is another power system case called case2383wp.m.
This case represents the Polish 400, 220 and 110kV networks during winter 19992000 peak conditions, and consists of three hundred twentyseven
generators and two thousand eight hundred ninetysix transmission lines.
The optimal cost of the solution is 1796588.57[$/h]. The authors have used
M atP ower to solve the DCOPF, and the optimal cost is the same.
6.4. Chiles Large Northern Interconnected System (SING)
The transmission network used in the study corresponds to a reduced version of the trunk transmission system. The transmission system is comprised
of two voltage levels, 220kV and 110kV , so the system has 17 generation units
(3 runofriver units and 14 thermal units), 76 bus, 83 transmission lines,
and 10 transformers. The technical parameters and the load of the SING
power system were obtained from the website of the Center for Economic
Load Dispatch of the Large Northern Interconnected System (CDECSING,
www.cdecsing.cl).
We have modeled the load of the power system on February 13, 2014
considering the peak load at 22pm (1854.80M W ). The thermal units for
the DCOPF problem have been selected based on the shortterm generation
scheduling given by the CDECSING.
22
Using the PTDFbased formulation, the authors have solved the DCOPF
problem in 27 iterations. The optimal cost of the solution is 100266.54[$/h],
which was obtained by M atP ower as well. In the solution, the transmission
line Atacama  Domeyko 220kV is used to its maximum capacity, so the
Lagrangian multiplier associated is 440.16[$/M W h].
In addition, the power system losses are estimated using equation (10).
The interiorpoint algorithm does not have problem to model the quadratic
transmission losses.
The PTDFbased formulation has been implemented and exhaustively
validated with dierent test power systems and the SING system. The results of the proposed algorithm demonstrate both the feasibility of applying
the methodology to the DCOPF problem and, above all, the potential applicability of the approach in large power systems.
Nowadays, the authors are implementing and including some constraints
in order to model and simulate the AC optimal power flow, and the transmission and generation expansion planning problem. The authors have applied
the DCOPF formulation using the PTDFbased formulation in a previous
study of transmission expansion planning [16], so the previous results were
very promising.
Nowadays, the authors are programming the interiorpoint method in
DIgSILEN T P owerF actory using the DIgSILEN T Programming Language (DPL). The DPL oers an interface to the user for the automation
of tasks in the P owerF actory program. The DPL script language uses syntax quite similar to the C + + programming language. Previous results are
very encouraging, so the DPL adds a new dimension to the DIgSILEN T
23
8. Acknowledgements
This study was supported in part by the Chilean National Commission
for Scientific and Technological Research (CONICYT) under grant Fondecyt
1130793.
[7] N. Karmankar, A new polynomialtime algorithm for linear programming, Combinatoria, vol. 4, no. 8, pp. 373395, 1984.
[8] G. Astfalk, I. Lustig, R. Marsten, and D. Shanno, The interiorpoint
method for linear programming, IEEE Software, vol. 9, no. 4, pp. 61
68, 1992.
[9] R. D. Zimmerman, C. E. MurilloSanchez, and R. J. Thomas, Matpower: Steadystate operations, planning, and analysus tools for power
systems research and education, IEEE Trans. on Power Syst., vol. 26,
pp. 1219, 2011.
[10] H. Dommel and W. Tinny, Optimal power flow solutions, IEEE Trans.
on Power Appar. Syst., vol. PAS87, no. 10, pp. 18661876, 1968.
[11] A. J. Wood and B. F. Wollenberg, Power Generation, Operation and
Control. John Wiley & Sons, 1996.
[12] S. J. Wright, Primaldual interiorpoint method. SIAM, 1997.
[13] N. Duvvuru and K. S. Swarup, A hybrid interior point assisted differential evolution algorithm for economic dispatch, IEEE Trans. on
Power Syst., vol. 26, no. 2, pp. 541549, 2011.
[14] V. H. Hinojosa, N. Galleguillos, and B. Nuques, A simulated rebounding algorithm applied to the multistage securityconstrained transmission expansion planning in power systems, Int. J. Elect. Power Energy
Syst., vol. 47, pp. 168180, 2013.
26
[15] PSERC,
Software
matpower,
[online],http
step
0.21
0.79
1.75
1.75
1.25
0.90
0.90
0.90
0.90
feascond
630
0.252
0.251
0.204
0.204
0.204
0.204
0.204
0.204
gradcond
4.481
0.159
0.309
0.286
0.002
4.90e09
8.98e11
1.20e11
9.52e12
compcond costcond
gamma
0
0
1
2.24
0.15
0.084
1.68
0.003
0.063
0.58
0.16
0.020
3.66
8.931e06
0.125
264.72
5.76e07
9.029
506305
1.14e09
17268.6
1.94e+13 5.67e14 6.61e+11
1.48e+25
0
5.06e+23
System
9 bus
14 bus
30 bus
39 bus
57 bus
118 bus
300 bus
M1
M2
M3
M atP ower
M1
M2
M3
M atP ower
M1
M2
M3
M atP ower
M3
M3
M3
M3
27
cost[$/h]
20393.15
20393.15
20393.15
20393.15
7642.59
7642.59
7642.59
7642.59
960.09
960.09
960.09
960.09
41263.94
41006.74
125954.42
706240.27
iter
0 M1
M3
1 M1
M3
2 M1
M3
3 M1
M3
4 M1
M3
5 M1
M3
6 M1
M3
7 M1
M3
8 M1
M3
9 M1
M3
10 M1
M3
11 M1
M3
12 M1
M3
13 M1
M3
14 M1
M3
15 M1
M3
P1
A2
75.0 120.0
75.0 120.0
148.1
0.0
147.0
0.0
154.0
0.0
152.5
0.0
159.5
0.0
159.3
0.0
174.7
0.0
174.5
0.0
174.0
0.0
174.0
0.0
165.7
0.0
160.8
0.0
165.7
0.0
160.8
0.0
150.0
0.0
150.0
0.0
150.0
0.0
150.0
0.0
150.0
0.0
150.0
0.0
150.0
0.0
150.0
0.0
150.0
0.0
150.0
0.0
150.0
0.0
150.0
0.0
150.0
0.0
150.0
0.0
150.0
0.0

P3
180.0
180.0
280.2
280.5
290.7
288.7
294.3
293.1
356.2
360.0
360.0
360.0
360.0
360.0
360.0
360.0
330.4
330.4
330.4
330.4
330.4
330.4
330.4
330.4
330.4
330.4
330.4
330.4
330.4
330.4
330.4

28
A4
A5
P6
80.0 120.0 300.0
80.0 120.0 300.0
6.1 14.4 438.4
8.1 11.1 444.6
20.0
0.1 450.1
20.0
0.0 453.7
19.8
0.0 441.0
20.0
0.0 442.4
19.2
0.0 362.0
19.8
0.0 358.6
12.5
0.0 353.5
19.2
0.0 353.1
12.5
0.0 298.0
8.7
0.0 265.6
0.0
0.0 298.0
8.7
0.0 265.6
0.0
0.0 223.2
0.0
0.0 223.2
0.0
0.0 223.2
0.0
0.0 223.2
0.0
0.0 223.2
0.0
0.0 223.2
0.0
0.0 223.2
0.0
0.0 223.2
0.0
0.0 223.2
0.0
0.0 223.2
0.0
0.0 223.2
0.0
0.0 223.2
0.0
0.0 223.2
0.0
0.0 223.2
0.0
0.0 223.2
