A Linear Optimal Power Flow Model Considering

Nodal Distribution of Losses

Arild Helseth
SINTEF Energy Research
Trondheim, Norway
[email protected]

Abstract—This paper presents a method for accurately treating enables formulation of sparse nodal loss constraints, which
active power losses in linear (DC) optimal power flow models. are efficiently treated in Simplex-based linear programming
Quadratic nodal losses are approximated by iteratively adding (LP) algorithms. Second, the proposed method converges at
linear constraints. In each iteration a linear programming prob-
lem is solved, and constraints on the nodal loss functions are an arbitrary convergence tolerance within a finite number of
built as linearizations around the current system operating state. iterations. Convergence is ensured by introducing a fictitious
The performance of the proposed model – in terms of compu- dumping option at each transmission node.
tational time and convergence properties – is demonstrated on Several electricity markets are based on nodal (locational
the IEEE 118 bus test system. marginal) prices, and the DCOPF is extensively used for
analyses in these markets [5]. It is elaborated how to find
I. I NTRODUCTION
the reference-node independent nodal prices for the proposed
Optimal power flow (OPF) studies has many applications, DCOPF formulation.
and is frequently used in in the fields of expansion planning, This paper is organized as follows. Section II introduces
generation scheduling [1] and reliability assessment [2]. The the DCOPF problem and provides a brief review of relevant
basic principle of an OPF is to minimize generator operating methods to include losses while keeping the linear formulation.
costs. Although the objective coincide with that of an eco- The suggested DCOPF formulation and solution strategy is
nomic dispatch, an OPF model also treats the transmission outlined in Section III. Subsequently, the proposed method is
system connecting generators and loads. tested on the IEEE 118 bus test system, and its computational
Representing transmission system power flow by the AC efficiency and convergence properties are documented and
power flow equations in an AC optimal power flow (ACOPF) discussed. Section V concludes the paper.
will provide the most accurate solution. However, the ACOPF
problem is generally a nonlinear and nonconvex optimization II. T HE DCOPF PROBLEM
problem. Thus, the convergence properties, robustness and The DC power flow model (1) is an approximation of
computation time of the ACOPF will often be a challenge. the nonlinear AC power flow equations where reactive power
The AC power flow may be approximated by a linearized flows and active power losses are ignored.
(DC) power flow representation in a DC optimal power flow
(DCOPF) [1]. The linearized representation is considered a
P − D = BΔ (1)
fair approximation of the AC power flow equations for many
practical systems [3]. The major advantage of the DCOPF Where:
is its simplicity, robustness and computational speed. These P Vector of active power generations;
properties are important in models applied to large-scale sys- D Vector of power demands;
tems where the OPF problem is solved under various operating B Susceptance matrix;
conditions a large number of times. As the DC power flow Δ Vector of voltage angles.
representation by definition is lossless, it is challenging to
Once the voltage angles are found, line losses in a single line
accurately model and capture the marginal impact of losses
can be approximated as:
on the simulated power system. On the other hand losses will
impact generation schedules and power flows, and could be  2
loss 2 δi − δj
included to make DCOPF studies more accurate. p = R f = R (2)
X
This paper presents an accurate method for distributing
line losses to transmission system nodes in a DCOPF model. Where:
Quadratic line losses are allocated nodes and are approximated ploss
 Active power loss in line ;
by iteratively adding linear loss constraints. The method R Resistance of line ;
presented here is based on the work in [4], but differs in two f Active power flow through line ;
important details. First, a different solution strategy is adopted δi Voltage angle at node i;
in order to improve computational efficiency. This strategy X Reactance of line .
Equation (2) shows that the line loss can be approximated as a B. DCOPF Models Considering Losses
quadratic function of the difference in voltage angle between Incorporation of power losses as found in (2) in the linear
the two nodes i and j at each end of line . The standard optimization problem in (3a)-(3c) can be done by adding the
procedure is now to add half of the line losses to the demand sum of line losses to the right-hand side in (3b). Losses
at each of the two connected transmission nodes. A new vector are then balanced at the reference node, while in practice
of voltage angles can then be found in (1) and the losses in (2) losses are “consumed” by the transmission lines. An improved
can in turn be updated. This iterative procedure will normally solution is obtained by adding the estimated nodal loss to the
converge after a few iterations. price insensitive demand in (3c) to balance losses at individual
The next section introduces the lossless DCOPF. Subse- nodes. In the above-mentioned formulations losses are treated
quently a review of relevant published approaches for treating as parameters in the LP problem. Thus, the optimal solution
losses in the DCOPF is presented. may be far from optimal with respect to total losses and loss
distributions.
A. Lossless DCOPF Model Another approach is to incorporate a piecewise linearization
The lossless DCOPF problem can be formulated as a of line losses by discretizing the flow variables [7]. Each flow
minimum cost LP problem (3a) subject to a global power variable segment is then assigned a loss coefficient. The main
balance in (3b) and power flow constraints in (3c). disadvantage of this formulation is the predefined large number
of flow variables needed to accurately represent losses in large-
NB
 scale systems.
min Z = ci p i (3a) Authors of [8] describe a DCOPF model, as defined by (4a-
i=1 4d) below.
NB
 NB
 NB

pi = di (3b) min Z = ci p i (4a)
i=1 i=1 i=1

NB
 NB

NB
 pi − ploss = di (4b)
−fmax ≤ Θi (pi − di ) ≤ fmax  = 1, . . . , NL (3c) i=1 i=1
i=1
NB
 NB

Where: ploss − βi p i = α − βi di (4c)
Z Total cost of system operation; i=1 i=1

ci Cost of generation at node i;

NB

pi Generation at node i;
di Price insensitive demand at node i; −fmax ≤ Θi (pi − di − φi pLoss ) ≤ fmax  = 1, . . . , NL
i=1
Θi PTDF for bus i to line ; (4d)
fmax Maximum capacity at line ; Here βi = ∂ploss
and α is the estimated offset for the
∂pi ,
NB Number of transmission nodes; system loss constraint (4c). System losses are represented
NL Number of lines. by a variable ploss and a factor φi indicating the share of
For notational convenience the mathematical formulation the total losses being distributed to node i. Consequently,
used in this paper holds one generator with a single cost losses are balanced at each node and the impact on line flow
segment and a price insensitive demand at each node in the is represented in (4d). The distribution factors φ, marginal
system. However, the presented method is also applicable loss factors β and and offset α are predefined and serve as
to more complex system models, e.g. with piecewise linear parameters in the LP problem. The method presented in [9]
generator cost-curves and price sensitive demands. Variable extends the work in [8] by introducing line loss variables
bounds are omitted for brevity. and linearizing line losses around a base case system state
The power transfer distribution factors (PTDFs) used in (3c) to formulate equality constraints of type (4c). Thus, the nodal
are found from the DC power flow representation as described distributions can be seen as active parts in the optimization
e.g. in [1]. A PTDF for an individual node to a single line problem, but the representation of line losses through a static
gives the ratio between the change of flow on the line and linear constraint is not fully accurate.
the change of power injection at the given node, provided that The authors in [10] formulate a model similar to (4a)-(4d)
the change in injection is balanced by a predefined reference and apply successive linearization of losses by updating the
node. parameters β, α and φ iteratively based on information from
The power flow constraints in (3c) can be efficiently treated a full AC power flow solution. In reference [11] the ploss
by relaxation, as suggested in [6]. That is, the power flow variable is omitted by introducing information from the β and
equations are not explicitly expressed in the optimization α directly in the power balance equation (4b). Nodal losses
problem, but are iteratively added in case of overloads. are estimated and substitute the loss distribution terms in (4d).
 It should be noted that, in contrast to the approaches for
linear representation of losses reviewed above, one can also ploss
i
keep the quadratic representation of line losses in (2) and solve
the DCOPF as a nonlinear optimization problem, see e.g. [12].
r=1
III. M ODELLING A PPROACH
Distribution of losses is fully embedded in the proposed r=2
DCOPF, as expressed in (5a)-(5f) below.
NB

min Z = (ci pi − cy yi ) (5a) p̂loss,1
i

i=1

NB
   NB p̂loss,2
i
pi − ploss
i − yi = di (5b) 3
2
p̂loss,3
i
i=1 i=1
1
NB f
  
f = Θi pi − ploss
i − yi − di  = 1, . . . , NL (5c)
Fig. 1. Iterative linear approximation of nodal losses.
i=1


ploss
i − (R fr ) f ≥ −p̂loss,r
i r = 1, . . . , NR (i) (5d) in (5f). The resulting operating point given by flow f1 and loss
∈ωi ploss,1
i is indicated in Fig. 1 as an open circle with the number
1. Subsequently, the exact loss p̂loss,1 is found from (6),
−fmax ≤ f ≤ fmax  = 1, . . . , NL (5e) i
as indicated by a filled black circle in Fig. 1. The ploss,1 i
suggested by the DCOPF model is at its lower bound and far
ploss
i ≥0  = 1, . . . , NB (5f) from the exact value. Thus, a loss cut (5d) is added to the LP
problem at r = 1. When re-solving the updated LP problem,
In comparison with the formulation in (4a)-(4d), three ad- a new operating point defined by f2 and ploss,2i is found, as
ditional variable types are introduced in the DCOPF problem. indicated in the figure by the open circle number 2. Still the
Each node is assigned one loss variable plossi and an option yi optimal nodal loss differs from the exact loss p̂loss,2 , so a new
i
to dump excess power. These are added to the global power loss cut is added to the LP model at r = 2. This procedure
balance in (5b) and will be discussed in Sections III-A and continues until a convergence criteria p̂lossi − ploss
i ≤  is
III-B. In addition, all line flows are represented by a designate reached. In Fig. 1 convergence at node i is reached in the
variable f as expressed in (5c) and bounded in (5e). third iteration, after adding two loss cuts. The resulting lower
A. Iterative Linear Approximation of Nodal Losses bound of the feasible region for ploss
i is shown by the dotted
grey line segments in Fig. 1.
A standard procedure for distributing line losses in the DC The iterative procedure described above could be improved
power flow solution is to divide the losses equally to each of to avoid a myopic nodal loss solution (i.e. ploss,1 = 0) in
i
the line’s nodes. According to this procedure the loss at node the first iteration. One could e.g. start with the loss cuts from
i is found by summing losses over the set ωi of connected a previously solved case, but these cuts should be removed
lines: in the subsequent iterations in case they do not represent the
1  current system conditions.
ploss
i = R f2 (6) This method for approximating losses in a DCOPF was first
2
∈ωi
presented by [4], where line losses were approximated by line
The loss allocated node i is a quadratic function of power loss cuts. As thoroughly demonstrated and discussed in [4]
flows on lines connected to i, and can therefore be approxi- this method has several advantageous features. Compared to
mated by a finite number NR (i) of linear constraints as in (5d). models were losses are piecewise linearized according to
These constraints are referred to as loss cuts in the remaining predefined segments, loss cuts are only computed for relevant
text. A derivation of the loss cuts in (5d) is provided in operating points in this method, i.e. points that have been
Appendix A. suggested by the DCOPF. This allows for more efficient
Nodal loss cuts are added iteratively to gradually improve and accurate loss computations. Opposed to successive LP
the linear approximation of (6), as illustrated in Fig. 1. The approaches, potential problems with slow convergence are
power loss in node i is a quadratic function of flow in a avoided by keeping previously built loss cuts in the LP
line  connected to i. First, let the loss cut counter r = 0 problem.
and the DCOPF problem described in (5a)-(5f) is solved with The formulation of nodal loss cuts, rather than line loss
no constraints on nodal losses other than the lower bounds cuts, in the presented formulation is motivated by limiting the
 t=t+1
number of constraints, as normally NL > NB . As long as line r=k=0
losses are divided equally to each connected node, the optimal r =r+1
solution should be equal for the line loss and the nodal loss Solve DCOPF
k =k+1
cut approaches.

B. Dumping Option and Convergence

Solve power flow
The opportunity to dump power through the variable yi is
in principle non-physical, but serves an important role in the
proposed DCOPF model. Consider formulating the DCOPF
model in (5a)-(5f) as a general optimization problem by Add power flow
Overloaded lines?
constraints (5c) yes
replacing the loss cuts in (5d) by the nonlinear representation
in (6), and omitting the dumping options y. As discussed
no
in [13] and [14], this problem is in general nonconvex, but
convexity can be guaranteed by ensuring non-negative nodal
prices. That is, the dual values for all nodal power balances Add power loss Losses within
cuts (5d) or (7) yes tolerance?
should be non-negative. This is ensured by introducing the
nodal dumping option.
no
Although negative nodal prices occur in practice, their oc-
currence would lead the model to suggest additional losses at Converged
the respective nodes, since increased losses will then improve Fig. 2. DCOPF solution strategies.
the objective function.
An important point here is to introduce the dumping vari-
ables as sales options with a small (negligible) income cy
and a large capacity to ensure that the LP solution dumps  NB
  
power rather than artificially increasing losses. Consequently, ploss
i − (R fr ) Θb pb − ploss
b − yb − db ≥ −p̂loss,r
i
the optimal nodal losses will always be located at the lower ∈ωi b=1
boundary of the feasible region defined by the loss cuts and the (7)
non-negativity constraint, and the loss estimation procedure In [4] the latter strategy was presented, i.e. treating both
will converge at an optimal solution in a finite number of power flow constraints and loss cuts by relaxation. A se-
iterations. Note that there will always be one or two binding quential logic for checking these relaxations was proposed,
loss cuts or a binding lower bound constraining the nodal considering power flow constraints first, as shown in Fig. 2.
losses. Given the large number of loss cuts needed to achieve
A different approach was presented in [4], introducing an accurate approximations of losses in large-scale systems, we
upper loss-cut to limit the overestimation of losses in case of believe that the explicit representation of flow variables in the
negative nodal prices. loss cuts will improve computational performance.
D. Nodal Prices
C. Solution Strategies
The nodal price (or locational marginal price) is the
There are several possible solution strategies applicable for marginal cost of supplying the next increment of energy
the DCOPF in (5a)-(5f). Two of these are discussed below and considering the generation marginal cost and the physical
illustrated in Fig. 2. aspects of the transmission system [15]. These prices can be
In the proposed formulation, all NL power flow constraints interpreted as the dual values of the nodal power balances, but
in (5c) are included in the initial LP problem. This was done to recall that nodal power balances are omitted in the presented
enable sparse representation of loss cuts in (5d). Each loss cut DCOPF formulation.
will only include a nodal loss variable and the flow variables The nodal prices obtained from the converged DCOPF
representing the lines connected to that particular node. The solution can be expressed as in (8). This expression is derived
dashed boxes in Fig. 2 are not visited when using this strategy. in Appendix B. These nodal prices are independent of the
Alternatively, the power flow constraints in (5c) can be choice of reference node.
treated by relaxation, as illustrated by the dashed boxes in
NR (i)
Fig. 2. In this case the loss cuts in (5d) will not be accurate, 
πi = τir + γimin (8)
as all flow expressions in (5c) are not introduced in the LP
r=1
problem. The power flow expressions in (5c) can then be
inserted in the loss cuts, as shown in (7). This type of loss cut Where:
involves all variables contributing to the nodal power injections πi Price at node i;
in all nodes, and are therefore much denser than the ones τir Dual value of loss cut r in (5d);
in (5d). γimin Dual value of loss variable bound in (5f).
 TABLE I TABLE II
S IMULATED C ASES C OMPUTATIONAL PERFORMANCE

Case Loss Tol. [%] Absolute Tol. [MW] Line Cap. [MW] Dense Strategy Sparse Strategy
1 10 0.50 – Case CPU time [sec.] NR NK CPU time [sec.] NR
2 5 0.10 – 1 0.53 6.26 0.00 0.35 5.26
3 1 0.01 – 2 0.60 6.40 0.00 0.41 5.72
4 10 0.50 100 3 0.96 8.05 0.00 0.50 6.67
5 5 0.10 100 4 0.75 7.44 2.24 0.40 6.51
6 1 0.01 100 5 0.89 8.21 2.27 0.44 6.97
6 1.89 10.92 2.31 0.55 8.35

IV. C ASE S TUDY

is referred to as the dense strategy. The differences in sim-
The presented method was tested on the IEEE 118 bus ulated losses and minimum costs for the two strategies were
test system [16]. This system comprises 118 nodes, 186 lines negligible. This was as expected since they solve the same
(actually lines and transformers, but referred to as lines here), problem.
54 generators and an active power base load of 4242 MW Performance indicators for the 6 different cases are pre-
distributed at 99 nodes. sented in Table II. This table shows the average CPU times
and average number of maximum iterations per load period for
A. Testing Procedure the two strategies. NR indicates the average maximum number
The operation of the IEEE 118 bus test system was sim- of loss cut iterations and NK the average maximum number
ulated using random samples of load data. The base values of power flow constraint iterations for the dense strategy. Both
for active power loads are defined in [16]. Nodal demand approaches converged within a maximum of 20 iterations for
values were drawn from a uniform distribution in an interval all load periods and for all cases. The test results in Table II
covering +/-50 % from the corresponding base values. The indicate that the sparse strategy is computationally faster than
DCOPF problem was solved for 500 samples of load data, the dense strategy for all simulated cases.
referred to as load periods in the following. Generator cost The differences in CPU times are most pronounced in the
curves were represented by a single segment, and the marginal cases with low convergence tolerances. Increased accuracy of
generator cost coefficients were defined in inverse proportion loss estimation comes at a lower computational cost when
to the installed capacities given in [16]. using the sparse strategy. Comparing the high-tolerance cases
A total of 6 cases were simulated, as described in Table I. 1 and 4 with the low-tolerance cases 3 and 6 for the sparse
The system was simulated with and without capacity con- strategy, the increase in CPU time when improving accuracy is
straints on line flows. In the flow constrained cases, a limit of 43 % for the non-constrained cases and 38 % for the capacity
100 MW was used for all lines. Three different convergence constrained cases. The increases between the corresponding
criteria for the loss estimation were applied, each holding both cases for the dense strategy are 81 % and 252 %.
a relative and an absolute tolerance. The absolute tolerance By adding capacity constraints on line flows, the power flow
was included to avoid numerical difficulties when estimating constraint loop in Fig. 2 is on average visited 2 times for the
low losses. dense strategy. The CPU time increases on average 10 % in
The DCOPF model was implemented in C++, using the cases 4-6 compared to cases 1-3 for the sparse strategy. The
Coin CLP LP solver [17]. For each new load period, the corresponding average increase for the dense strategy is 69 %.
LP problem was built and solved from scratch. This was Recall that in the sparse strategy all power flow constraints are
done to resemble possible time sequential changes in market added to the LP problem initially; Thus, the computational
description and the physical power system. Whenever the LP performance is less dependent on the transmission system
problem was not solved from scratch, i.e. either one of the loss capacity limits.
cut (r) or the power flow constraint (k) iteration counters were C. Loss Estimation
greater than zero, the dual algorithm was applied warm starting The results presented in this section were obtained using
from the previous solution. All simulations were carried on an the sparse strategy. An average increase in operational cost
Intel Core i7-Q740 processor with 1.73 GHz and 16 GB RAM. of 5.5 % was found when including losses in case 4. This
indicates the importance of considering losses in the DCOPF
B. Computational Performance
problem.
In this section the computational performance of the two The computed system losses for each load period in case 4
different solution strategies discussed in Section III-C is com- are shown in Fig. 3. This figure illustrates how the different
pared. The suggested strategy, where all power flow constraints system conditions impact loss computation.
are added to the LP problem initially and loss cuts of type (5d) The average numbers of loss cuts added in each loss-
are used, is referred to as the sparse strategy. The alternative iteration in cases 4 and 6 are plotted in Fig. 4. A maximum
strategy, where power flow constraints are relaxed and the number of 8 and 10 iterations was registered for cases 4 and
flow variables are implicitly represented in the loss cuts (7), 6, respectively.
 120

12
Total Losses [MW]

10
110

8
Loss [MW]
95 100

6
90

4
0 100 200 300 400 500 8
Load period 27

2
42
54
92
Fig. 3. Total system losses for each load period in case 4.

0
100

1 2 3 4 5 6 7
Iteration number

Case 4
Fig. 5. Variation of nodal losses with loss-cut iteration number. The legend
20 40 60 80 100
Number of loss cuts added

Case 6 indicates associated node numbers.

the DCOPF with and without loss computation can be reduced

by warm starting the LP problem in new load periods. For this
case study, where we only change the demand between load
periods, the power flow constraints should be added initially
and should only have their right-hand sides updated according
to the new nodal demands in the subsequent load periods.
One could also take advantage of loss cuts from the previous
0

2 4 6 8 10 load period to improve the decision in the first iteration, as

discussed in Section III-A.
Iteration number

Fig. 4. The average no. of loss cuts added per iteration in cases 4 and 6.
V. C ONCLUSION

Fig. 5 shows how losses allocated some selected nodes in a

An iterative method for accurate approximation of quadratic
specific load period in case 6 vary with the iteration number.
nodal losses in a linear OPF model was presented. The approx-
The selected nodes are among those having the highest nodal
imation of nodal losses converges at an arbitrary convergence
losses in this load period. Note that for most nodes the
tolerance within a finite number of iterations.
estimated loss converges within a few iterations. This is the
case for node 8 in Fig. 5 for which only one loss cut was The method was tested on the IEEE 118 bus test system,
needed. and it was shown that losses significantly impact the total
operational costs. A set of arranged test cases demonstrated
D. Comparison with the Lossless DCOPF the computational efficiency and convergence properties of the
As expected, the accurate treatment of nodal losses in suggested method. It was found that the computation time
the presented method significantly increases computation time is only modestly affected by increased accuracy in the loss
compared to a lossless DCOPF. To quantify this increase, a estimation procedure and by significant changes in available
lossless DCOPF was run for the network constrained case. transmission system capacity.
Average CPU times of 0.015 and 0.063 seconds were obtained The suggested loss computation procedure significantly
when treating power flow constraints with and without relax- increases computation time compared to the lossless DCOPF.
ation, correspondingly. The large gap in CPU times between Some measures were suggested to limit this increase.
 A PPENDIX A the Lagrangian function with respect to the price insensitive
D ERIVATION OF N ODAL L OSS C UTS demand di [8].
NB N R (b)

By differentiating (6) with respect to the line flows, one gets: ∂ψ  
=π− τbr Θi R fr
∂ploss  ∂di
i r=1
b=1 ∈ωb
= R f (A.1)
∂f NL
  
∈ωi
− Θi ηmax − ηmin (B.2)
Next, a first-order Taylor series expansion can be formed =1
around an operating point described by p̂loss,r
i and f r : The Kuhn-Tucker conditions for the nodal loss variables can
 be expressed as:
∂ploss 
ploss ≥ p̂ loss,r
+ i  (f − f r ) (A.2) 
i i
∂f f =f r NB N
 R (b)
 NR (i)

∂ψ r
loss
= π − τbr Θ R f
i   − τir
Where p̂loss is estimated from (6) based on f r . Combin- ∂pi b=1 r=1 ∈ω r=1
i b

ing (A.1) and (A.2): NL

  
 − Θi ηmax − ηmin − γimin = 0 (B.3)
ploss
i ≥ p̂loss,r
i + R fr (f − fr ) (A.3) =1
∈ωi
Finally, by combining (B.2) and (B.3) to eliminate π:
Reorganizing (A.3) to collect the parameters at the right-hand NR (i)

∂ψ
side: = τir + γimin (B.4)
  ∂di r=1
ploss
i − (R fr ) f ≥ p̂loss,r
i − R (fr )2 (A.4)
R EFERENCES
∈ωi ∈ωi
[1] A. Wood and B. F. Wollenberg, Power Generation, Operation and
Which can be further simplified to: Control, 2nd ed. John Wiley & Sons, 1996.
[2] R. Billinton and W. Li, Reliability Assessment of Electric Power Systems

(R fr ) f ≥ −p̂loss,r
Using Monte Carlo Methods. New York: Plenum Press, 1994.
ploss
i − i (A.5) [3] B. Stott, J. Jardim, and O. Alsaç, “DC power flow revisited,” IEEE
∈ωi Trans. Power Systems, vol. 24, no. 3, pp. 1290–1300, 2009.
[4] T. N. dos Santos and A. L. Diniz, “A dynamic piecewise linear model for
A PPENDIX B DC transmission losses in optimal scheduling problems,” IEEE Trans.
D ERIVATION OF N ODAL P RICES Power Systems, vol. 26, no. 2, pp. 508 – 519, 2011.
[5] E. Litvinov, “Design and operation of the locational marginal prices-
based electricity markets,” IET Generation, Transmission and Distribu-
The Lagrangian function ψ for the LP problem in (5a)-(5f) tion, vol. 4, no. 2, pp. 315–323, 2009.
can be expressed as: [6] B. Stott and J. L. Marinho, “Linear programming for power-system
network security applications,” IEEE Transactions on Power Apparatus
and Systems, vol. PAS-98, no. 3, pp. 837–848, 1979.
NB
N
  B
 
[7] B. F. Hobbs, G. Drayton, E. B. Fisher, and W. Lise, “Improved
y
ψ= (ci pi − c yi ) − π pi − ploss
i − yi − di transmission representations in oligopolistic market models: Quadratic
losses, phase shifters, and DC lines,” IEEE Trans. Power Systems,
i=1 i=1
NL
N vol. 23, no. 3, pp. 1018–1029, 2008.
  B
  [8] E. Litvinov, T. Zheng, G. Rosenwald, and P. Shamsollahi, “Marginal
− μ Θi pi − ploss
i − yi − di − f loss modeling in LMP calculation,” IEEE Trans. Power Systems, vol. 19,
no. 2, pp. 880–888, 2004.
=1 i=1
R (i)
NB N
 [9] V. Sarkar and S. A. Khaparde, “DCOPF-based marginal loss pricing
  with enhanced power flow accuracy by using matrix loss distribution,”
− τir ploss
i − (R fr ) f + p̂loss,r
i
IEEE Trans. Power Systems, vol. 24, no. 3, pp. 1435–1445, 2009.
[10] Z. Hu, H. Cheng, Z. Yan, and F. Li, “An iterative LMP calculation
i=1 r=1 ∈ωi
method considering loss distributions,” IEEE Trans. Power Systems,
NL
 NL
   vol. 25, no. 3, pp. 1469–1477, 2010.
+ ηmax (f − fmax ) + ηmin −f + fmin [11] F. Li and R. Bo, “DCOPF-based LMP simulation: Algorithm, compari-
son with ACOPF, and sensitivity,” IEEE Trans. Power Systems, vol. 22,
=1 =1
no. 4, pp. 1475–1485, 2007.
NB
 [12] H. Stigler and C. Todem, “Optimization of the Austrian electricity sector
− γimin ploss
i (control zone of VERBUND APG) by nodal pricing,” Central European
i=1 Journal of Operations Research, vol. 13, no. 2, pp. 105–125, 2005.
(B.1) [13] A. Philpott and G. Pritchard, “Financial transmission rights in convex
pool markets,” Operations Research Letters, vol. 32, no. 2, pp. 109–113,
2004.
Where π, μ, τ , η and γ are dual values for the corresponding [14] R. Palma-Benhke, A. Philpott, A. Jofré, and M. Cortés-Carmona,
constraints and variable bounds. “Modelling network constrained economic dispatch problems,”
Flow variables in (B.1) can be substituted by the linear Electric Power Optimization Centre, 2010. [Online]. Available:
http://www.epoc.org.nz/
combination of nodal injections in (5c). An expression for the [15] M. Shahidehpour, H. Yamin, and Z. Li, Market Operations in Electric
nodal price at node i is obtained by finding the derivative of Power Systems. New York: Wiley, 2002.
[16] IEEE 118 bus test system. [Online]. Available:
http://www.ee.washington.edu/research/pstca/
[17] Coin-CLP. [Online]. Available: http://www.coin-or.org/projects/Clp.xml