AN INTRODUCTION TO OPTIMAL POWER FLOW:

THEORY, FORMULATION, AND EXAMPLES∗

STEPHEN FRANK† AND STEFFEN REBENNACK‡

Abstract. The set of optimization problems in electric power systems engineering known collectively as Optimal Power
Flow (OPF) is one of the most practically important and well-researched subfields of constrained nonlinear optimization. OPF
has enjoyed a rich history of research, innovation, and publication since its debut five decades ago. Nevertheless, entry into OPF
research is a daunting task for the uninitiated—both due to the sheer volume of literature and because OPF’s ubiquity within
the electric power systems community has led authors to assume a great deal of prior knowledge that readers unfamiliar with
electric power systems may not possess. This paper provides an introduction to OPF from an operations research perspective;
it describes a complete and concise basis of knowledge for beginning OPF research. The discussion is tailored for the operations
researcher who has experience with nonlinear optimization but little knowledge of electrical engineering. Topics covered include
power systems modeling, the power flow equations, typical OPF formulations, and common OPF extensions. Supplementary
materials are available for this article. Go to the publisher’s online edition of IIE Transaction.

Key words. power flow, optimal power flow, electric power systems analysis, electrical engineering, nonlinear programming,
optimization, operations research

AMS subject classifications. 90-01, 90C26, 90C30, 90C90

1. Introduction. The set of optimization problems in electric power systems engineering known collec-
tively as Optimal Power Flow (OPF) is one of the most practically important and well-researched subfields of
constrained nonlinear optimization. In 1962, Carpentier [16] introduced OPF as an extension to the problem
of optimal economic dispatch (ED) of generation in electric power systems. Carpentier’s key contribution
was the inclusion of the electric power flow equations in the ED formulation. Today, the defining feature of
OPF remains the presence of the power flow equations in the set of equality constraints.
OPF includes any optimization problem which seeks to optimize the operation of an electric power system
subject to the physical constraints imposed by electrical laws and engineering limits. This general framework
encompasses dozens of optimization problems for power systems planning and operation [23, 61, 62]. As
illustrated in Figure 1.1, optimization of power system operation typically occurs via incremental planning:
long-term planning procedures make high level decisions based on coarse system models, while short-term
procedures refine earlier decisions using detailed models but more limited decision spaces. OPF may be
applied to decision making at nearly any planning horizon—from long-term transmission network capacity
planning to minute-by-minute adjustment of real and reactive power dispatch [26, 56, 62].
To date, thousands of articles and hundreds of textbook entries have been written about OPF. In its
maturation over the past five decades, OPF has served as a practical proving ground for many popular non-
linear optimization algorithms, including gradient methods [7, 19, 42], Newton-type methods [52], sequential
linear programming [6, 50], sequential quadratic programming [14], both linear and nonlinear interior point
methods [27, 53, 54], and semi-definite programming [31, 33]. These algorithmic approaches, among others,
are reviewed in several surveys [13, 30, 35, 38, 59, 61], including one recently published by the authors [23, 24].
Although OPF spans operations research and electrical engineering, the accessibility of the OPF litera-
ture skews heavily toward the electrical engineering community. OPF has become sufficiently familiar within
the electric power systems community that the recent literature, including survey papers, assumes a great
deal of prior knowledge on the part of the reader. Few papers even include a full OPF formulation, much
less explain the particulars of the objective function or constraints. Even introductory textbooks [46, 56, 62]
require a strong background in power systems analysis, specifically regarding the form, construction, and
solution of the electric power flow equations. Although many electrical engineers have this prior knowledge,
an operations researcher likely will not. We believe this accessibility gap has been detrimental to the in-
volvement of the operations research community in OPF research; our impression is that most OPF articles
continue to be published in engineering journals by electrical engineers.

∗ This work is supported by the National Science Foundation Graduate Research Fellowship under Grant No. DGE-1057607.
† NationalRenewable Energy Laboratory, Buildings & Thermal Systems, 15013 Denver West Parkway, Golden, CO, 80401,
USA, E-mail: [email protected]
‡ Colorado School of Mines, Division of Economics and Business, 1500 Illinois St., Golden, CO, 80401, USA, E-mail:

[email protected]
1
2 STEPHEN FRANK AND STEFFEN REBENNACK

Planning Horizon
Years Months Weeks Days Hours Minutes Real-Time

n
lutio
del Reso
asin g Mo
Incre

Decr
easin
g Unce
rtain
ty

Long-term
Market Clearing / Economic
Generation
Unit Commitment Dispatch Automatic
Scheduling
Generation
Capacity Security- Control
Classic Optimal
Expansion Constrained Unit
Power Flow
Planning Commitment
Real Power Dispatch
Reactive Power Dispatch
Security-
Reactive Power Optimal Reactive Voltage
Constrained Unit
Planning Power Flow Control
Commitment

Fig. 1.1: Optimization and control procedures for incremental planning of power system operation. Bold
text indicates procedures which incorporate variants of optimal power flow.

What is missing in the literature—and what we provide in this introductory article—is a detailed in-
troduction to the OPF problem from an operations research perspective. Existing review articles and sur-
veys [13,59,60], including the authors’ recent survey [23,24], focus heavily on optimization theory and tailored
OPF solution algorithms. In contrast, this article emphasizes the electrical engineering theory and mechanics
of the OPF formulation. The goal of the this paper is to provide a bridge between OPF theory and practice:
in it we outline the tool set required to understand, formulate, analyze, and ultimately solve a typical OPF
problem. Therefore, we address many topics given little attention in other introductory materials, including
the construction of the admittance matrix for electrical power flow, treatment of advanced controls such as
phase-shifting and tap-changing transformers, a qualitative comparison of the various forms of the electric
power flow equations, and various practical considerations, such as the use of the per-unit system.
Because we have written this article for the operations researcher, we assume that the reader has signif-
icant experience with nonlinear optimization and advanced mathematical concepts but little background in
electrical engineering. Specifically, this paper requires a foundational understanding of
• linear algebra [28],
• complex number theory [25, 28],
• analysis of differential equations in the frequency domain [28], and
• linear and nonlinear optimization theory and application [39, 44].
We expect that readers may not possess a working knowledge of electrical circuit theory; we therefore
provide a brief introduction in Appendix B and recommend [40] for further reading. Readers interested in
the technical details of electric power flow should also consult a good power systems text such as [26] or [56].
We begin in Section 2 with a description of power systems models, including the classical formulation
of the OPF problem. Section 3 surveys some common applications of OPF and includes full formulations
 AN INTRODUCTION TO OPTIMAL POWER FLOW 3

for several of the decision processes shown in Figure 1.1. Section 4 describes the bus admittance matrix,
which is the foundation of the power flow equations. Section 5 reviews the various forms of the power flow
equations, with an emphasis on describing their relative advantages and disadvantages. Building on the
power flow equations, Section 6 introduces OPF solution methods and practical considerations. Section 7
then provides a worked example of a classical OPF formulation. Finally, Section 8 concludes the paper.
We also include four appendices which provide supplemental information regarding electric power systems
modeling and analysis. Appendix A documents the notation used throughout the paper, Appendix B reviews
fundamental electric power system concepts relevant to OPF, Appendix C summarizes the per-unit system,
and Appendix D describes common formats for the exchange of power system data.
Different readers may have different goals in reviewing this article. We recommend that all readers start
with the introduction and Section 2. Readers interested only in a brief OPF overview can subsequently read
Sections 3 and 6.2 and safely skip the detailed derivations in other sections. Conversely, readers attempting
to implement or test an OPF solution algorithm should read Sections 4–7 in detail. Readers interested
primarily in understanding linear approximations for the OPF problem should review Section 5, and in
particular Section 5.3 which discusses DC power flow.
2. Modeling of Power Systems. In this section, we introduce the classic network model for an electric
power system and describe conventional and optimal power flow. The development requires some basic
knowledge of electrical circuit theory, the frequency domain (phasor) representation of electrical quantities,
and the concept of complex electric power. Readers unfamiliar with these topics should consult Appendix
B first for a brief overview.
2.1. Notation. Throughout this article, italic roman font (A) indicates a variable or parameter, bold
roman font (A) indicates a set, and a tilde over a symbol (e a) indicates a phasor quantity (complex number).
Letter case does not differentiate variables from parameters; a given quantity may be a variable in some cases
and a parameter in others. Symbolic superscripts are used as qualifiers to differentiate similar variables, while
numeric superscripts indicate mathematical operations. For example, the superscript L differentiates load
power P L from net power P , but P 2 indicates (net) power squared. Where applicable, electrical units are
specified using regular roman font. The unit for a quantity follows the numeric quantity and is separated by
a space; for example 120 V indicates 120 Volts.
We use the following general notation for optimization formulations:
u vector of control variables (independent decision variables)
x vector or state variables (dependent decision variables)
f (u, x) objective function (scalar)
g(u, x) vector function of equality constraints
h(u, x) vector function of inequality constraints
In order to remain consistent with the existing body of OPF literature, this article uses notation that fol-
lows electrical engineering conventions rather than those of the operations research community. In particular,
the symbols e and j represent mathematical constants:
e Euler’s number (the base of the natural √ logarithm), e ≈ 2.71828, and
j the imaginary unit or 90° operator, j = −1.
This differs from the use of e as the unit vector and j as an index as is common in operations research liter-
ature. Appendix A includes a full listing of the notation used in this paper, including relevant commentary
on other notational differences and a listing of electrical engineering units used in power systems analysis.
2.2. Network Representation. Electric power systems may be modeled as a network of electrical
buses (nodes) interconnected by branches (arcs or edges) that represent transmission lines, cables, trans-
formers, and similar power systems equipment. Buses represent physical points of interconnection among
power systems equipment while branches represent paths for the flow of electrical current. The purpose of
an electric power system is to transfer electrical energy from generation (supply) buses to load (demand)
buses elsewhere in the network.
Buses are referenced by node with index i ∈ N, while branches are referenced as arcs between nodes
(i, k) ∈ L, where i, k ∈ N. The undirected graph (N, L) therefore describes the connectivity of the electrical
network. The number of buses and branches are N = |N| and L = |L|, respectively.
Each system bus i has an associated voltage Vei which is measured with respect to the system reference
(typically earth ground). When connected via the branch network, or “grid”, these voltages induce current
4 STEPHEN FRANK AND STEFFEN REBENNACK

in each branch in proportion to the branch admittance. (Admittance, which is a measure of how easily a
conductor permits the flow of electrical current, is described in Appendix B.) The most common and concise
mathematical description of this phenomenon is the matrix equation

Ie = Ye Ve , (2.1)

in which Ve = (Ve1 , . . . , VeN ) is an N -dimensional vector of phasor voltages at each system bus, Ie = (Ie1 , . . . , IeN )
is an N -dimensional vector of phasor currents injected into the network at each system bus, and
 
Ye11 . . . Ye1N
Ye =  ... .. .. 

. . 
YN 1 . . . YN N
e e

is the N × N complex bus admittance matrix, which is described in detail in Section 4. At each bus i,
injection current Iei represents the net current supplied to the network: generation (supply) minus load
(demand). In this framework, voltages Ve are state variables which fully characterize system power flow for
a given matrix Ye .
It is more convenient to work with power flows than currents because (i) injected powers are indepen-
dent of system voltage angle while injected currents are not, and (ii) working directly with power allows
straightforward computation of required electrical energy via integration with respect to time. Therefore,
power systems engineers transform (2.1) by applying the definition of complex power S = Ve Ie∗ to each side of
the matrix equation, as described in Appendix B.4. (Here and elsewhere in this paper, the symbol ∗ denotes
complex conjugation rather than an optimal value; this use is typical in electrical engineering.) The result
is the complex power flow equation
 ∗
S = Ve ◦ Ye Ve , (2.2)

in which S = P + jQ is a vector of complex power injections at each bus and ◦ denotes element-wise vector
multiplication. At each bus i, the total injected power is the difference between the generation SiG and the
load SiL ,

Si = SiG − SiL Pi + jQi = PiG − PiL + j QG L

⇔ i − Qi .

For numerical analysis or optimization, (2.2) may be decomposed into a set of equivalent, real-valued,
nonlinear power flow equations by separating its real and imaginary components (see Section 5).
2.3. Conventional Power Flow. The conventional power flow (PF) problem seeks a deterministic
solution to network equation (2.2) using numerical analysis techniques. Conventional PF is a feasibility
problem: there is no objective function. Rather, the goal is to compute all system bus voltages and power
injections. Here, we summarize the PF problem with polar voltage coordinates, which is the classical
representation.
If each bus voltage is represented in polar form with magnitude V and phase angle δ, then (2.2) decom-
poses into the set of power flow equations

Pi (V, δ) = PiG − PiL ∀ i ∈ N, (2.3)

Qi (V, δ) = QG
i − QLi ∀ i ∈ N, (2.4)

in which net real and reactive power injections Pi and Qi are trigonometric functions of the system voltages.
(See Section 5 for the fully expanded equations.) Each system bus has four variables (net real power
injection Pi , net reactive power injection Qi , voltage magnitude Vi , and voltage angle δi ) and is governed by
two equations. Thus, a deterministic solution to the conventional PF problem requires fixing the values of
two out of four variables at each bus.
In conventional PF, all system buses are assigned to one of three bus types:
Slack Bus At the slack bus, or swing bus, the voltage magnitude and angle are fixed and the power injections
are free. The purpose of the slack bus is twofold. First, it provides a voltage reference (typically
 AN INTRODUCTION TO OPTIMAL POWER FLOW 5

Table 2.1: Power system bus types and characteristics for conventional power flow.

Bus Type Slack PQ PV

# of buses in system 1 M N −M −1
Known quantities δ, V P, Q P, V
Unknown quantities P, Q δ, V δ, Q
# of equations in conventional PF 0 2 1

V = 1.0 p.u. and δ = 0°) such that the remaining bus voltages are uniquely determined; we explain
the per-unit system “p.u.” in Appendix C. Second, because it is the only bus at which real power
is free to vary, the slack bus is required to ensure that the power flow equations have a feasible
solution. There is only one slack bus in a power system model.
Load Bus At a load bus, or “PQ” bus, the power injections are fixed while the voltage magnitude and
angle are free. There are a fixed number of PQ buses in the system; the symbol M denotes this
number.
Voltage-Controlled Bus At a voltage controlled bus, or “PV” bus, the real power injection and voltage
magnitude are fixed while the reactive power injection and the voltage angle are free. (This corre-
sponds to allowing a local source of reactive power to regulate the voltage to a desired setpoint.)
There are N − M − 1 PV buses in the system.

Assigning buses in this way establishes an equal number of equations and unknowns. Table 2.1 summarizes
the known and unknown quantities for each of the bus types.

Once all voltage magnitudes and angles in the system have been computed, then the remaining power
injections are trivial to evaluate via (2.3)–(2.4). Solving the power flow therefore requires determining N − 1
voltage angles (corresponding to the PQ and PV buses) and M voltage magnitudes (corresponding to the
PQ buses only). This is done by solving N + M − 1 simultaneous nonlinear equations with known right
hand side values. This equation set consists of the real power injection equation (2.3) at each PQ and PV
bus and the reactive power injection equation (2.4) at each PQ bus. Section 6.1 discusses solution methods
for these equations.

Even though the power flow equations are nonlinear, there exists only one physically meaningful solution
for most power systems models given an equal number of equations and unknowns. Although other mathe-
matically valid solutions sometimes exist, they have no realistic physical interpretation. (An example would
be any solution which returns a negative voltage magnitude, as magnitudes are by definition nonnegative.)
Hence, in practice, conventional PF is an exactly determined problem.

2.4. Optimal Power Flow. Optimal power flow combines an objective function with the power flow
equations (2.3)–(2.4) to form an optimization problem. The presence of the power flow equations is the
feature that distinguishes OPF from other classes of power systems problems, such as classic economic
dispatch (ED), unit commitment (UC), and market clearing problems.

Most OPF variants build upon the classical formulation of Carpentier [16] and Dommel and Tinney [19].
The classical formulation is an extension of classic ED: its objective is to minimize the total cost of electricity
generation while maintaining the electric power system within safe operating limits. The power system is
modeled as a set of buses N connected by a set of branches L, with controllable generators located at a
subset G ⊆ N of the system buses.
The operating cost of each generator is a (typically quadratic) function
of its real output power: Ci PiG . The objective is to minimize the total cost of generation.
6 STEPHEN FRANK AND STEFFEN REBENNACK

The classical form of the formulation is

X
Ci PiG ,


min (2.5)
i∈G

s.t. Pi (V, δ) = PiG − PiL ∀ i ∈ N, (2.6)

Qi (V, δ) = QG
i − QLi ∀ i ∈ N, (2.7)
PiG,min ≤ PiG ≤ PiG,max ∀ i ∈ G, (2.8)
QG,min
i ≤ QGi ≤ Qi
G,max
∀ i ∈ G, (2.9)
Vimin ≤ Vi ≤ Vimax ∀ i ∈ N, (2.10)
δimin ≤ δi ≤ δimax ∀ i ∈ N. (2.11)

Constraints (2.6)–(2.7) are the power flow equations in polar form. The remaining constraints represent
bounds on the system voltages and powers. Typically, load real and reactive power are fixed while gener-
ator real and reactive power are control variables subject to minimum and maximum limits. The voltage
magnitude and angle at the system slack bus (by convention, bus 1) are also fixed, usually to Ve1 = 1.0∠0.
Early methods partition the decision variables into a set of control variables u (typically the controllable
bus power injections) and a set of state variables x (the voltage magnitudes and angles) [14, 19]. Under this
framework, the vector of control variables (independent decision variables) for the classical formulation is
G
, QG


u = Pi:i∈G i:i∈G

and the vector of state variables (dependent decision variables) is

x = (δ2 , . . . , δN , V2 , . . . , VN ) .

Although not considered in the earliest papers, more recent OPF formulations may include branch
current limits and, if applicable, box constraints related to the operational limits of power flow control
devices. These and other side constraints are discussed in Section 5.5.
2.5. Challenges. Power systems have evolved significantly since the early days of OPF, in particular
through the addition of advanced controls. Modern grids include control devices that are difficult to incor-
porate into OPF formulations: on-load tap changers [3], phase shifters [37], series and shunt capacitors [36],
and flexible AC transmission systems (FACTS) devices [58]. Because such controls exert a large influence on
system power flows, they cannot be neglected in practical OPF formulations. Unfortunately, many academic
papers neglect the modeling of advanced controls and treat only the classical formulation, limiting their
application in a practical setting.
Nevertheless, some researchers have dedicated considerable effort to developing accurate but efficient
models for advanced controls [3, 32, 58]. Most such models employ auxiliary power injections at certain
system buses coupled with side constraints to enforce power balance. However, even a modest number of such
constraints can significantly increase OPF problem complexity [8]. Many devices also have discrete control
settings, which if accurately modeled create a large and intractable mixed-integer nonlinear programming
(MINLP) problem [48]. A typical approach is to model the control space as continuous and round the optimal
solution to the nearest discrete value [4], but this heuristic can yield suboptimal or infeasible solutions [15,48].
Even without variable phase angles, variable tap ratios, or branch current limits, the classical OPF
formulation is difficult to solve. The power flow constraints (2.6)–(2.7) are both nonlinear and non-convex,
and the presence of trigonometric functions complicates the construction of approximations. Moreover,
because OPF is tightly constrained, local optima often provide only incremental improvements with respect
to the system starting point. Thus, the optimum may only improve a few percent upon the base case, as
illustrated by numerical results in [7], [27], [52], and many other papers.
For these reasons, OPF problems have historically been solved using tailored algorithms rather than
general purpose solvers. Although effective for research problems, few such algorithms are considered suf-
ficiently reliable for industrial deployment. In particular, solution methods for both conventional PF and
OPF become significantly less reliable when the power system is under stress [13]. Algorithm robustness is
also a key concern: small changes in the power system state may lead to the loss of local feasibility (a critical
 AN INTRODUCTION TO OPTIMAL POWER FLOW 7

issue for local nonlinear solvers) or large changes in the optimal solution [5]. In addition, real power system
models include thousands of buses; problems of this size present computational challenges for state-of-the-art
nonlinear solvers.
In actual practice, almost all practical OPF problems are solved using a linear (DC) power flow approx-
imation (Section 5.3) [45, 51]. The results of the DC-OPF are then used either for contingency analysis with
conventional AC-PF or, less commonly, to initialize, or “warm start,” a nonlinear OPF (AC-OPF). Unfor-
tunately, the DC formulation can exhibit significant inaccuracy, particularly for heavily loaded transmission
lines [36, 51]. Nor is the DC model applicable to reactive power dispatch. The search for reliable AC-OPF
solution methods therefore remains extremely relevant to the power systems community.
3. Applications of Optimal Power Flow. Besides the classical ED formulation, several other OPF
variants are common in both industry and research. These include security-constrained economic dispatch
(SCED), security-constrained unit commitment (SCUC), optimal reactive power flow (ORPF), and reactive
power planning (RPP).
3.1. Security-Constrained Economic Dispatch. Security-constrained economic dispatch (SCED),
sometimes referred to as security-constrained optimal power flow (SCOPF), is an OPF formulation which
includes power system contingency constraints [7]. A contingency is defined as an event which removes
one or more generators or transmission lines from the power system, increasing the stress on the remaining
network. SCED seeks an optimal solution that remains feasible under any of a pre-specified set of likely
contingency events. SCED is a restriction of the classic OPF formulation: for the same objective function,
the optimal solution to SCED will be no better than the optimal solution without considering contingencies.
The justification for the restriction is that SCED mitigates the risk of a system failure (blackout) should one
of the contingencies occur.
SCED formulations typically have the same objective function and decision variables u as the classic
formulation, except that the slack bus real and reactive power are considered state variables because they
must be allowed to change in order for the system to remain feasible during each contingency. However,
SCED introduces NC additional sets of state variables x and accompanying sets of power flow constraints,
where NC is the number of contingencies. SCED can be expressed in a general way as
min f (u, x0 ),
s.t. g0 (u, x0 ) = 0,
h0 (u, x0 ) ≤ 0, (3.1)
gc (u, xc ) = 0 ∀ c ∈ C,
hc (u, xc ) ≤ 0 ∀ c ∈ C,

where C = {1, . . . , NC } is the set of contingencies to consider. Each contingency has a distinct admittance
matrix Yec , typically with less connectivity than the original system. Apart from the contingency index, f , g,
and h are defined as the objective function, equality constraints, and inequality constraints in the classical
OPF formulation (2.5)–(2.11), respectively.
For each contingency c ∈ C, the post-contingency power flow must remain feasible for the original
decision variables u:
(i) The power flow equations must have a solution,
(ii) The contingency state variables xc must remain within limits, and
(iii) Any inequality constraints, such as branch flow limits, must be satisfied.
One very common contingency set considers independently the loss of each generator and each non-radial
transmission line. The resulting SCED solution is said to be N − 1 reliable, since for every possible single
element contingency c the network has a feasible solution xc given the optimal dispatch decisions u.
Typically, the limits on the contingency-dependent state variables xc and other functional inequality
constraints are relaxed for the contingency cases compared to the base case. For example, system voltages
are allowed to dip further during an emergency than under normal operating conditions. The relaxation of
system limits is justified because operation under a contingency is temporary: when a contingency occurs,
operators immediately begin re-configuring the system to return all branches and buses to normal operating
limits. For the same reason, the SCED objective function considers only the base case: contingency cases
are improbable and transient and do not contribute to the expected cost function.
8 STEPHEN FRANK AND STEFFEN REBENNACK

SCED has interesting connections to other areas of optimization. The motivation for SCED is theoreti-
cally similar to that of Robust Optimization (RO) [12], although RO typically addresses continuous uncertain
parameters rather than discrete scenarios. Additionally, because the constraints are separable for a fixed u,
SCED lends itself well to parallelization and decomposition algorithms [43].
3.2. Security-Constrained Unit Commitment. In electric power systems operation, unit commit-
ment (UC) refers to the scheduling of generating units such that total operating cost is minimized. UC differs
from ED in that it operates across multiple time periods and schedules the on-off status of each generator
in addition to its power output. UC must address generator startup and shutdown time and costs, limits
on generator cycling, ramp rate limits, reserve margin requirements, and other scheduling constraints. UC
is a large-scale, multi-period, MINLP. Many UC formulations relax certain aspects of the problem in order
to obtain a mixed-integer linear program (MILP) instead—for instance by using linearized cost functions.
If the power flow equations are added to the UC problem, the formulation becomes security-constrained
unit commitment (SCUC). In SCUC, a power flow is applied at each time period to ensure that the scheduled
generation satisfies not only the scheduling constraints but also system voltage and branch flow limits. In
other words, SCUC ensures that the UC algorithm produces a generation schedule that can be physically
realized in the power system. Because of its complexity, research on SCUC has accelerated only with the
advent of faster computing capabilities and algorithmic advancements.
In SCUC, we introduce a time index t ∈ T and a set of binary control variables wit to the OPF
formulation. Each wit indicates whether or not generator i is committed for time period t. The modified
formulation becomes
XX 
wit Ci PitG + CiSU wit (1 − wi,t−1 ) + CiSD (1 − wit ) wi,t−1 ,


min (3.2)
t∈T i∈G

s.t. Pit (Vt , δt , ϕt , Tt ) = PitG − PitL ∀ i ∈ N, ∀ t ∈ T, (3.3)

Qit (Vt , δt , ϕt , Tt ) =QG
it − QLit ∀ i ∈ N, ∀ t ∈ T, (3.4)
wit PiG,min ≤ PitG ≤ wit PiG,max ∀ i ∈ G, ∀ t ∈ T, (3.5)
wit QG,min
i ≤ QG G,max
it ≤ wit Qi ∀ i ∈ G, ∀ t ∈ T, (3.6)
Vimin ≤ Vit ≤ Vimax ∀ i ∈ N, ∀ t ∈ T, (3.7)
δimin ≤ δit ≤ δimax ∀ i ∈ N, ∀ t ∈ T, (3.8)
ϕmin
ik ≤ ϕikt ≤ ϕik
max
∀ ik ∈ H, ∀ t ∈ T, (3.9)
min max
Tik ≤ Tikt ≤ Tik ∀ ik ∈ K, ∀ t ∈ T, (3.10)
max
Iikt (Vt , δt ) ≤ Iik ∀ ik ∈ L, ∀ t ∈ T, (3.11)
PiDown ≤ PitG − Pi,t−1 G
≤ PiUp ∀ i ∈ G, ∀ t ∈ T, (3.12)
X X
wit PiG,max − PitG ≥ PReserve ∀ t ∈ T. (3.13)
i∈G i∈G

The objective function (3.2) includes terms for unit startup costs C SU and shutdown costs C SD in addition
to the generation costs in each time period. The power flow equations (3.3)–(3.4) are generalized to include
the effects of phase-shifting and tap-changing transformers (Section 5.5). (We abuse notation slightly by
using (Vt , δt , ϕt , Tt ) to represent the vector of all voltage magnitudes, voltage angles, and engineering control
settings for time period t.) The generation limits (3.5)–(3.6) are modified such that uncommitted units must
have zero real and reactive power generation. Box constraints (3.9)–(3.10) enforce engineering control limits
and (3.11) limits the current on each transmission line; these constraints and associated sets H and K are
described in detail in Section 5.5. Constraint (3.12) specifies positive and negative generator ramp limits
P Up and P Down , respectively; these are physical limitations of the generators. Constraint (3.13) requires
a spinning reserve margin of at least PReserve ; sometimes this constraint is written such that PReserve is a
fraction of the total load in each time period.
The SCUC formulation (3.2)–(3.13) is one of many possible formulations. Some formulations include
more precise ramp limits and startup and shutdown characteristics; others include constraints governing
generator minimum uptime and downtime. Because of the scale and presence of binary decision variables,
 AN INTRODUCTION TO OPTIMAL POWER FLOW 9

SCUC is one of the most difficult power systems optimization problems. Bai and Wei [9] and Zhu [62] provide
more discussion of SCUC, including detailed formulations.
More broadly, SCUC belongs to a class of problems known as multi-period or dynamic OPF. Dynamic
OPF problems span multiple time periods. Mathematically, each time period requires a single copy of
the underlying PF equations; the dimensionality therefore scales linearly with the time horizon. Coupling
constraints then link the different time periods. In the case of the SCUC, the ramping constraints (3.12)
provide this coupling. Grid-scale storage systems, such as batteries, also require coupling constraints which
create dynamic OPF problems [47]. Other examples include system design problems such as discussed in [22],
where multiple OPF problems are coupled by binary design variables.

3.3. Optimal Reactive Power Flow. Optimal reactive power flow (ORPF), also known as reactive
power dispatch or VAR control, seeks to optimize the system reactive power generation in order to minimize
the total system losses. In ORPF, the system real power generation is determined a priori, from the outcome
of, for example, a DC-OPF algorithm, UC, or another form of ED. A basic ORPF formulation is

min P1 , (3.14)
s.t. Pi (V, δ, ϕ, T ) = PiG − PiL ∀ i ∈ N, (3.15)
Qi (V, δ, ϕ, T ) = QGi − Qi
L
∀ i ∈ N, (3.16)
QG,min
i ≤ QG i ≤ Qi
G,max
∀ i ∈ G, (3.17)
Vimin ≤ Vi ≤ Vimax ∀ i ∈ N, (3.18)
δimin ≤ δi ≤ δimax ∀ i ∈ N, (3.19)
ϕmin
ik ≤ ϕik ≤ ϕik
max
∀ ik ∈ H, (3.20)
min max
Tik ≤ Tik ≤ Tik ∀ ik ∈ K, (3.21)
max
Iik (V, δ) ≤ Iik ∀ ik ∈ L. (3.22)

The power flow equations (3.15)–(3.16) are again generalized to include the effects of phase-shifting and
tap-changing transformers (Section 5.5). The vector of control variables is

u = P1 , QG


i:i∈G , ϕik:ik∈H , Tik:ik∈K ,

while the vector of state variables x = (δ, V ) is identical to the classical formulation. In ORPF, all real
power load and generation is fixed except for the real power at the slack bus, P1 . Minimizing P1 is therefore
equivalent to minimizing total system loss.
One motivation for using ORPF is the reduction of the variable space compared to fully coupled OPF [17];
another is the ability to reschedule reactive power to optimally respond to changes in the system load without
changing previously established real power setpoints. Many interior point algorithms for OPF have focused
specifically on ORPF [23]. Zhu [62, ch. 10] discusses several approximate ORPF formulations and their
solution methods.

3.4. Reactive Power Planning. Reactive power planning (RPP) extends the ORPF problem to the
optimal allocation of new reactive power sources—such as capacitor banks—within a power system in order
to minimize either system losses or total costs. RPP modifies ORPF to include a set of possible new
reactive power sources; the presence or absence of each new source is modeled with a binary variable. The
combinatorial nature of installing new reactive power sources has inspired many papers which apply heuristic
methods to RPP [24].
10 STEPHEN FRANK AND STEFFEN REBENNACK

A basic RPP formulation which minimizes total costs is

X
min C1 (P1 ) + wi CiInstall , (3.23)
i∈Q

s.t. Pi (V, δ, ϕ, T ) = PiG − PiL ∀ i ∈ N, (3.24)

Qi (V, δ, ϕ, T ) = QGi + Qi
New
− QLi ∀ i ∈ N, (3.25)
QG,min
i ≤ QG i ≤ Qi
G,max
∀ i ∈ G, (3.26)
wi QNew,min
i ≤ QNew
i ≤ wi QNew,max
i ∀ i ∈ Q, (3.27)
Vimin ≤ Vi ≤ Vimax ∀ i ∈ N, (3.28)
δimin ≤ δi ≤ δimax ∀ i ∈ N, (3.29)
ϕmin
ik ≤ ϕik ≤ ϕik
max
∀ ik ∈ H, (3.30)
min max
Tik ≤ Tik ≤ Tik ∀ ik ∈ K, (3.31)
max
Iik (V, δ) ≤ Iik ∀ ik ∈ L, (3.32)
where CiInstall represents the capital cost of each new reactive power source i ∈ Q; QNew i is the amount of
reactive power provided by each new reactive power source, subject to limits QNew,min i and QNew,max
i ; and
wi is a binary variable governing the decision to install each new reactive power source. The modified vector
of control variables is
u = P1 , QG New


i:i∈G , wi:i∈Q , Qi:i∈Q , ϕik:ik∈H , Tik:ik∈K .

Some variants of RPP also include real power dispatch in the decision variables or include multiple load
scenarios.
By necessity, RPP optimizes with respect to uncertain future conditions—typically reactive power re-
quirements for worst-case scenarios. This uncertainty, together with the problem complexity, make RPP
a very challenging optimization problem [59]. Zhang et al. [59, 60] review both formulations and solution
techniques for RPP.
4. The Admittance Matrix. The power flow equations are the defining constraints in OPF, and the
bus admittance matrix Ye in turn forms the core of the power flow equations. OPF data sources do not
typically provide Ye directly, and we therefore summarize the theory and mechanics of its construction here.
Readers uninterested in the electrical engineering details may wish to skip this section; such readers may
reference (4.10)–(4.13) as needed for the mathematical definition of the matrix elements.
The elements of Ye derive from the application of Ohm’s and Kirchoff’s laws (Appendix B) to a steady-
state AC electrical network. Recall from Section 2.2 that the set of buses N and set of branches L form
undirected graph (N, L) which describes the electrical network. Each branch (i, k) ∈ L has an associated
series admittance yeik which governs the voltage-current relationship between buses i and k. According to
Kirchoff’s voltage law (KVL) and Ohm’s law,
 
Ieik = Vei − Vek yeik , (4.1)

in which Ieik is the current flowing through branch ik from bus i to bus k. Branches may also have an
Sh
associated shunt admittance yeik , which represents leakage of current from within the branch to the reference
node. In physical terms, this leakage occurs all along the branch, but in the model the shunt admittance is
applied in equal parts to the buses at each end of the branch as illustrated in Figure 4.1. This representation
is called the Π branch model.
By Kirchoff’s current law (KCL), the current Iei injected into bus i must exactly equal the sum of all
currents flowing out of bus i via the various series and shunt admittances, including any shunt admittance
yeiS associated with the bus itself. Neglecting off-nominal branch turns ratios (described in Section 4.1), the
bus injection current is
 
1 X 1 X X  X 
Iei = Vei yeiS + Sh
yeik + Sh 
yeki + Vei − Vek yeik + Vei − Vek yeki . (4.2)
2 2
k:(i,k)∈L k:(k,i)∈L k:(i,k)∈L k:(k,i)∈L
 AN INTRODUCTION TO OPTIMAL POWER FLOW 11

Ieik yeik Ieki

Bus i Bus k

Sh
yeik Sh
yeik
Vei Vek
2 2

Fig. 4.1: Π branch model.

Ieik Ieik
0
Ieki
yeik
Bus i Bus k

Sh
yeik Sh
yeik
Vei Vei0 Vek
2 2

a:1

Fig. 4.2: Generalized Π branch model, including off-nominal turns ratio.

Matrix equation (2.1) is equivalent to (4.2) if the elements of Ye are defined as

Yeii = Sum over all admittances connected to bus i

(4.3)
−Yeik = Sum over all admittances connected between bus i and bus k

Since only a single branch (i, k) connects bus i to bus k, the off-diagonal elements become Yeik = Yeki = −e
yik .
If there is no connection between buses i and k, Yeik = 0. Typically, each bus connects to only a few branches,
such that L ∈ O(N ). Thus, Ye is sparse, having dimension N × N but only N + 2L nonzero entries.
4.1. Branch Models. The Π model of Figure 4.1 is sufficient for modeling the majority of power
systems branch elements, including transmission lines, cables, and transformers. Most power systems trans-
formers have nominal turns ratios, that is, the voltage ratio across the transformer exactly equals change in
system voltage base across the transformer (a 1:1 voltage ratio in per-unit, with no phase shift). Because
the per-unit system automatically accounts for nominal turns ratios (see Appendix C), no corrections to the
admittance matrix are necessary.
Some transformers, however, do not have exactly a 1:1 voltage ratio in per-unit. Such transformers
are labeled “off-nominal”; this category includes fixed-tap transformers with other than unity turns ratios,
tap-changing transformers, and phase-shifting transformers. In practical power system models, improperly
neglecting off-nominal turns ratios may severely distort computed power flow, such that any resulting OPF
solution may be unusable. Therefore, off-nominal transformers require modified entries in Ye to account for
the additional voltage magnitude or phase angle change relative to the nominal case.
The generalized Π branch model Figure 4.2 extends the Π model to accommodate both nominal and
Sh
off-nominal turns ratios. The model includes series admittance yeik , shunt admittance yeik , and an ideal
transformer. Bus i is the tap bus and bus k is the impedance bus, or Z bus. The transformer turns ratio in
per-unit is a : 1, where a is a complex exponential consisting of magnitude T and phase shift ϕ,

a = T ejϕ ,

such that in the figure Vei = a Vei0 and Ie = Ieik

0
/a∗ . (To simplify the mathematical presentation, we abuse
notation throughout this section by omitting the subscript ik on a, T , and ϕ.) Selecting T = 1 and ϕ = 0
yields the nominal turns ratio.
12 STEPHEN FRANK AND STEFFEN REBENNACK

In order to model the effects of the off-nominal turns ratio, we require a partial admittance matrix Ye 0
such that
! ! !
Ieik Yeii0 Yeik0 Vei
= e0 e0 .
Ieki Yki Ykk Vek

The elements of Ye 0 are derived from the application of KVL, KCL, and Ohm’s Law. From KCL at node i
and the definitions of Vei0 and Iei0 ,
 Sh 
1 e yeik 1
Iik = ∗ Vi
e + yeik − ∗ Vek yeik . (4.4)
aa 2 a
Similarly, at bus k,
Sh
 
1 yeik
Ieki = − Vei yeik + Vek + yeik . (4.5)
a 2
Rearranging (4.4) and (4.5) to form a matrix equation yields Ye 0 ,
  Sh  
1 yeik 1
+ yeik − ∗ yeik  Ve
! ! !
Ieik  aa∗ 2 a i 0 V
ei
= Sh  Vek = Y Vek .
 e (4.6)
Ieki  1 yeik
− yeik + yeik
a 2
When constructing the full admittance matrix Ye , the corrected relationships of (4.6) must be preserved.
This is done via appropriate substitutions in (4.2); Section 4.2 gives formulas for the resulting admittance
matrix entries. A branch with off-nominal magnitude only (real valued a) leaves Ye a symmetric matrix, but
a phase-shifting transformer (complex a) does not.
4.1.1. Transmission Lines and Cables. The line characteristics for transmission lines and cables
are most often specified as a series impedance Rik + jXik and a branch shunt admittance jbSh ik , which is
sometimes given as “line charging” reactive power. The Π branch series admittance yeik for inclusion in Ye is
then
1 Rik Xik
yeik = = 2 2 −j 2 2 . (4.7)
Rik + jXik Rik + Xik Rik + Xik
As is typical for Π branch models, shunt susceptance bSh ik is divided into two equal parts which are applied
to the buses at each end of the branch. For short lines, branch shunt susceptance is negligible and therefore
is usually omitted. Transmission lines and cables have nominal turns ratios: T = 1.0 and ϕ = 0.
4.1.2. Transformers. Like cables, power systems transformers have a series impedance Rik + jXik
which models the electrical characteristics of the transformer windings. In power flow analysis, however,
the transformer series resistance is often neglected, yielding yeik = −j/Xik . Although transformer models
Sh
may also include a shunt admittance yeik which represents the losses and magnetizing characteristics of the
transformer core, this shunt admittance is nearly always neglected as well.
Off-nominal transformers have either T 6= 1 and/or ϕ 6= 0. In OPF, both T and ϕ may be control
variables: controllable T models an on-load tap changer, while controllable ϕ models a phase shifting trans-
former.
4.2. Construction Equations for Admittance Matrix. Any branch model may be distilled into
Sh
a generic series admittance yeik , shunt admittance yeik , and complex turns ratio (nominal or off-nominal)
aik = Tik ejϕik . Using (4.3) and incorporating the corrected off-nominal voltage-current relationships intro-
duced in Section 4.1, the entries of Ye become
  X  
S
X 1 1 Sh 1 Sh
Yii = yei +
e
2 yeik + yeik + yeki + yeki , (4.8)
|aik | 2 2
k:(i,k)∈L k:(k,i)∈L
X 1 X 1
Yeik = − y ik − yeki , i 6= k, (4.9)
a∗ik
e
aik
k:(i,k)∈L k:(k,i)∈L
 AN INTRODUCTION TO OPTIMAL POWER FLOW 13

where aik = 1 for any branch with a nominal turns ratio. Equations (4.8)–(4.9) may be separated into real
and imaginary parts using the definition Ye = G + jB and the identity aik = Tik (cos ϕik + j sin ϕik ),
   
X 1 1 Sh X 1 Sh
Gii = giS + 2 gik + gik + gki + gki , (4.10)
Tik 2 2
k:(i,k)∈L k:(k,i)∈L
X 1 X 1
Gik = − (gik cos ϕik − bik sin ϕik ) − (gki cos ϕki + bki sin ϕki ) , i 6= k (4.11)
Tik Tki
k:(i,k)∈L k:(k,i)∈L
  X  
S
X 1 1 Sh 1 Sh
Bii = bi + 2 bik + bik + bki + bki , (4.12)
Tik 2 2
k:(i,k)∈L k:(k,i)∈L
X 1 X 1
Bik = − (gik sin ϕik + bik cos ϕik ) − (−gki sin ϕki + bki cos ϕki ) , i 6= k. (4.13)
Tik Tki
k:(i,k)∈L k:(k,i)∈L

Appendix D describes how to obtain the required parameters for constructing the admittance matrix from
publicly available data sources.

5. The Power Flow Equations. The AC power flow equations transform the complex-domain ma-
trix function (2.2) into a set of real-valued simultaneous equations suitable for inclusion in mathematical
programming formulations. The most prevalent and compact form of the power flow equations is the bus
injection model. However, the less compact branch flow model also describes (2.2) and has several advantages
with respect to convex relaxation of OPF problems.

5.1. Bus Injection Model. The bus injection model is usually synonymous with the term “AC power
flow equations” and provides a compact representation of power system behavior in terms of the real and
reactive power injection at each system bus. Use of the bus injection model is ubiquitous in conventional
power flow and in the early OPF literature.
For a given Ye , equation (2.2) may be decomposed into a set of equations for the real and reactive power
injections by evaluating the real and imaginary parts of S, respectively. The resulting pair of equations may
be written in several equivalent forms depending on whether the voltages and admittance matrix elements
are expressed in polar or rectangular coordinates. In the literature, the most common forms of the AC power
flow equations are, in order,
1. Selection of polar coordinates for voltage, Vei = Vi ∠δi , and rectangular coordinates for admittance,
Yeik = Gik + jBik :

N
X

Pi (V, δ) = Vi Vk Gik cos (δi − δk ) + Bik sin (δi − δk ) ∀ i ∈ N, (5.1)
k=1
N
X

Qi (V, δ) = Vi Vk Gik sin (δi − δk ) − Bik cos (δi − δk ) ∀ i ∈ N. (5.2)
k=1

2. Selection of polar coordinates for voltage, Vei = Vi ∠δi , and polar coordinates for admittance,
Yeik = Yik ∠θik :

N
X
Pi (V, δ) = Vi Vk Yik cos (δi − δk − θik ) ∀ i ∈ N, (5.3)
k=1
N
X
Qi (V, δ) = Vi Vk Yik sin (δi − δk − θik ) ∀ i ∈ N. (5.4)
k=1

3. Selection of rectangular coordinates for voltage, Vei = Ei + jFi , and rectangular coordinates for
14 STEPHEN FRANK AND STEFFEN REBENNACK

admittance, Yeik = Gik + jBik :

N
X
Pi (E, F ) = Gik (Ei Ek + Fi Fk ) + Bik (Fi Ek − Ei Fk ) ∀ i ∈ N, (5.5)
k=1
N
X
Qi (E, F ) = Gik (Fi Ek − Ei Fk ) − Bik (Ei Ek + Fi Fk ) ∀ i ∈ N. (5.6)
k=1

Power systems texts [26,56,62] provide exact derivations of these three forms of the AC power flow equations.
(The fourth form—selection of rectangular coordinates for voltage and polar coordinates for admittance—is
theoretically possible but has no advantages for practical use.) Each form of the equations involves real-
valued quantities only. However, all forms are equivalent and give the exact solution to the power flow under
the analysis assumptions outlined in Appendix B.3.

The bus injection model of the AC power flow equations yields exactly 2N simultaneous equations in
4N variables. Of these, only N + M − 1 equations in are needed for conventional PF. Despite the equations’
nonlinearity and complexity, efficient solution methods exist for the conventional PF problem (see Section
6.1).

5.2. Branch Flow Model. The branch flow model is an alternative form of the power flow equations
that expresses the real and reactive power flow in each system branch, as derived from (4.1). Baran and
Wu [10, 11] originated this model as part of an optimal capacitor placement problem in radial distribution
systems, although earlier researchers had used similar, linear network flow models derived from DC power
flow (Section 5.3) [8]. The branch flow model has received considerable recent interest because it offers
advantages for convex relaxation of OPF problems [20, 21].

Low [33] summarizes the branch flow model using three sets of complex equations,

Vei − Vek = Z
eik Ieik ∀ (i, k) ∈ L, (5.7)
∗
Sik = Vei Ieik ∀ (i, k) ∈ L, (5.8)
X X  2


∗ 2
Si = Sik − Ski − Zeki Ieki + yeiS Vei ∀ i ∈ N. (5.9)
k|ik∈L k|ki∈L

in which Zeik = 1/e

yik is the complex series impedance of each system branch. (Note that (5.7)–(5.9) neglect
branch shunt admittances and off-nominal turns ratios.) The system variables are 2L real-valued components
of directed branch currents i → k, 2L directed branch real and reactive power flows i → k, N bus voltage
magnitudes, N bus voltage angles, and 2N bus real and reactive power injections, for a total of 4N + 4L
variables. Similarly, there are 2L directed branch current definitions i → k, 2L directed branch real and
reactive power definitions i → k, and 2N bus power balance equations, for a total of 2N + 4L simultaneous
equations. As with the bus injection model, the system is underdetermined by exactly 2N variables. However,
the dimensionality of the system of equations is considerably larger.

Frank and Rebennack [22] present an alternative form of the branch flow model which (i) uses rectangular
coordinates for the bus voltages and (ii) substitutes the real and reactive power flow at the receiving of each
branch as variables in place of the branch currents. The branch power definitions follow directly from (4.6)
 AN INTRODUCTION TO OPTIMAL POWER FLOW 15

and the definition of complex power,

S
 
1 gik
Ei2 + Fi2


Pik = 2 gik +
Tik 2
1
+ (−gik cos ϕik + bik sin ϕik ) (Ei Ek + Fi Fk )
Tik
1
+ (gik sin ϕik + bik cos ϕik ) (Ei Fk − Fi Ek ) ∀ (i, k) ∈ L, (5.10)
Tik
S
 
gik
Ek2 + Fk2


Pki = gik +
2
1
+ (−gik cos ϕik − bik sin ϕik ) (Ei Ek + Fi Fk )
Tik
1
+ (gik sin ϕik − bik cos ϕik ) (Ei Fk − Fi Ek ) ∀ (i, k) ∈ L, (5.11)
Tik
bSik
 
1
Ei2 + Fi2


Qik = − 2 bik +
Tik 2
1
+ (gik sin ϕik + bik cos ϕik ) (Ei Ek + Fi Fk )
Tik
1
+ (gik cos ϕik − bik sin ϕik ) (Ei Fk − Fi Ek ) ∀ (i, k) ∈ L, (5.12)
Tik
bSik
 
Ek2 + Fk2


Qki = − bik +
2
1
+ (−gik sin ϕik + bik cos ϕik ) (Ei Ek + Fi Fk )
Tik
1
+ (−gik cos ϕik − bik sin ϕik ) (Ei Fk − Fi Ek ) ∀ (i, k) ∈ L. (5.13)
Tik
Bus power balance equations (5.9) are still required, but instead take the form
X X
Pi = Pik + Pik ∀ i ∈ N, (5.14)
k|ik∈L k|ki∈L
X X
Qi = Qik + Qik ∀ i ∈ N. (5.15)
k|ik∈L k|ki∈L

Unlike (5.7)–(5.9), equations (5.10)–(5.15) do include the effects of branch shunt admittances and off-nominal
turns ratios. The system of equations still contains 4N + 4L variables (2N bus voltage components, 2N
bus real and reactive power injections, and 4L direct branch real and reactive power flows) and 2N + 4L
simultaneous equations.
5.3. DC Power Flow. The AC power flow equations are nonlinear. For conventional PF, this nonlin-
earity requires the use of an iterative numerical method; for OPF it implies both a nonlinear formulation and
non-convexity in the feasible region. In order to simplify the system representation, power systems engineers
have developed a linear approximation to the power flow equations. This approximation is called DC power
flow—so named because the equations resemble the power flow in a direct current (DC) network. However,
the DC power flow equations still model an AC power system.
The conventional development of the DC power flow equations requires several assumptions regarding
the power system [46, 62]:
1. All system branch resistances are approximately zero, that is, the transmission system is assumed
to be lossless. As a result, all θik = ±90° and all Gik = 0.
2. The differences between adjacent bus voltage angles are small, such that sin(δi − δk ) ≈ δi − δk and
cos(δi − δk ) ≈ 1.
3. The system bus voltages are approximately equal to 1.0. This assumption requires that there is
sufficient reactive power generation in the system to maintain a level voltage profile.
16 STEPHEN FRANK AND STEFFEN REBENNACK

4. Reactive power flow is neglected.

Applying these assumptions to (5.1) produces the DC power flow equation
N
X
Pi (δ) ≈ Bik (δi − δk ) (5.16)
k=1

Under normal operating conditions, DC power flow models real power transfer quite accurately. It
has been successfully used in many OPF applications that require rapid and robust solutions, including in
commercial software. However, the assumptions required for DC power flow can lead to significant errors
for stressed systems.
Neglecting off-nominal turns ratios, the exact equation for branch power transfer is

Pik = gik Vi2 − gik Vi Vk cos (δi − δk ) − bik Vi Vk sin (δi − δk ), (5.17)

cf. (5.1), while the DC power flow approximation is

Pik ≈ −bik (δi − δk ) . (5.18)

The bik term dominates the exact expression because Vi2 ≈ Vi Vk cos (δi − δk ) and therefore the first two
terms in (5.17) largely cancel. We observe that (5.18) overestimates the magnitude of the branch power
transfer (5.17) if
(i) The bus voltages at either end of the branch are depressed relative to the assumed value of 1.0
p.u., or
(ii) The angle difference between the buses is too large.
Observation (ii) follows from the relationship |sin (δi − δk )| ≤ |δi − δk |. Depressed voltages and larger than
normal angle differences are common in stressed power systems. In particular, large differences in voltage
in different areas of the system can lead to significant error [51]. Therefore, the DC power flow equations
should not be used for OPF under stressed system conditions unless they have been carefully evaluated for
accuracy in the system under test.
As an example, consider an arbitrary transmission line from bus i to bus k with series admittance
0.05 − j2.0. Let Vei = 0.95∠0° and Vek = 0.90∠−20°. (While these numbers do not represent normal
operation, they are plausible for a stressed power system. Operating voltages as low as 0.9 p.u. are allowable
in emergency conditions, and angle differences of up to ±30° can occur on long, heavily loaded transmission
lines.) The exact power transfer for this line is

Pik = 0.05 · 0.952 − 0.05 · 0.95 · 0.90 cos (0° + 20°) − 2.0 · 0.95 · 0.90 sin (0° + 20°) = 0.590 p.u.

The DC approximate power transfer is

Pik ≈ −2.0 sin (0° + 20°) ≈ 0.684 p.u.

The error in the approximate power transfer is 16%; most of this error is attributable to the voltage difference.
In systems with severely depressed voltages, large absolute errors in line power are typical, although the
relative error in the proportion of total power flowing in each line may be much smaller [51].
Even under normal operation, the approximation of a lossless transmission network can also lead to
significant errors in generator scheduling, branch power flow estimates, and marginal fuel cost estimates.
Power transfer errors for certain critically loaded branches can be much higher than the average branch
error. Therefore, in practical DC power flow models an estimate of the losses must be reintroduced using
approximate methods, especially if the network is large [51]. For further discussion regarding the advantages
and disadvantages of DC power flow, including loss approximation methods, we refer the interested reader
to Rau [45] and Stott et al. [51]. Throughout the rest of this paper, we use the AC power flow equations.
5.4. The Power Flow Equations as Constraints. In OPF, the power flow equations form the core
of the set of equality constraints. A large majority of OPF formulations use the bus injection model because
(i) it is more compact and (ii) it more closely resembles conventional PF, which facilitates the development
of tailored algorithms. The advantages of the branch flow model are the explicit modeling of branch power
 AN INTRODUCTION TO OPTIMAL POWER FLOW 17

flows, which facilitates inclusion of transmission capacity constraints, and that the form of the equations
enables certain convex relaxations.
Traditionally, the branch impedances and elements of Ye are considered constant: most algorithmic
development and analysis in the engineering literature assumes constant Ye . In newer OPF formulations
Ye may also contain control (decision) variables due to phase-shifting or tap-changing transformers, which
significantly complicates the problem by introducing bilinear, trilinear, or other nonlinear terms.
5.4.1. Considerations for the Bus Injection Model. The key consideration when using the bus
injection model is the choice of polar or rectangular coordinates for voltage. The advantage of voltage polar
coordinates is that constraints on the voltage magnitude can be enforced directly,

Vi ≥ Vimin ,
Vi ≤ Vimax .

In voltage rectangular coordinates, on the other hand, voltage magnitude limits require the functional in-
equality constraints
q
Ei2 + Fi2 ≥ Vimin ,
q
Ei2 + Fi2 ≤ Vimax .

Similarly, if the voltage magnitude is fixed (for instance at a PV bus; see Section 2.3), then in polar
coordinates Vi can be replaced with a constant value. In rectangular coordinates, however, a fixed voltage
magnitude requires the equality constraint
q
Ei2 + Fi2 = Vi .

Thus, for a fixed voltage magnitude, use of voltage polar coordinates leads to a reduction of variables while
use of voltage rectangular coordinates leads to an increase in (non-linear, non-convex) equality constraints.
For this reason, polar coordinates are preferred both for conventional PF and most OPF formulations.
There is, however, one compelling reason to use voltage rectangular coordinates: expressing voltage in
rectangular coordinates eliminates trigonometric functions from the power flow equations. The resulting
power flow equations (5.5)–(5.6) are quadratic, which presents several advantages [53]:
1. The elimination of trigonometric functions speeds evaluation of the equations.
2. The 2nd order Taylor series expansion of a quadratic function is exact; this yields an efficiency
advantage in higher-order interior-point algorithms for OPF.
3. The Hessian matrix for a quadratic function is constant and must be evaluated only once. This
simplifies the application of Newton’s method to the Karush-Kuhn-Tucker (KKT) conditions of the
OPF formulation.
4. If the side constraints are also quadratic, then the entire OPF problem becomes a Quadratically-
Constrained Program (QCP), although it remains nonconvex. This enables the use of specialized
solution algorithms.
In some cases, these computational advantages outweigh the disadvantages associated with increased dimen-
sionality and enforcing voltage magnitude constraints. Table 5.1 summarizes the differences between the two
voltage coordinate choices.
For the admittance matrix elements, rectangular coordinates (5.1)–(5.2) are typically favored because
they facilitate the use of certain approximations in fast-decoupled solution methods for conventional PF [62].
These approximations are also useful in the development of the DC power flow equations (Section 5.3).
Finally, rectangular coordinates facilitate the inclusion of transformer voltage ratios and phase angles as
decision variables. However, none of these advantages strongly affects the AC power flow equations as
implemented in many OPF formulations, and some models use the polar form of the equations (5.3)–(5.4).
5.4.2. Considerations for the Branch Flow Model. Because solving OPF problems is NP-hard,
convex relaxations are or particular interest, especially when conditions are known that cause these relax-
ations to be exact. There are several different approaches to convexify the OPF problem, dependent on the
formulation used [33]. Examples of such relaxations include semidefinite programming, second order cone
18 STEPHEN FRANK AND STEFFEN REBENNACK

Table 5.1: Comparison of the selection of voltage polar coordinates versus voltage rectangular coordinates
for the power flow equations. Bold entries indicate the more favorable characteristic.

Polar Coords. Rectangular Coords.

Number of variables in conventional PF N +M −1 2N − 2
Voltage magnitude limits Simple bounds Nonlinear functional
inequality constraint
Fixed voltage magnitude Variable elimination Nonlinear functional
by substitution equality constraint
Nature of PF equations Trigonometric Quadratic
1st derivative of PF equations Trigonometric Linear
2nd derivative of PF equations Trigonometric Constant

programming and, convex Lagrangian dual. Quite surprisingly, the exactness of some of these relaxations has
been proven for radial networks (under mild conditions). However, for mesh networks under the classical bus
injection model, examples with large duality gap have been published and, as a consequence, the relaxations
are no longer exact.
In contrast, the branch flow model allows for exact relaxations and convexifications for both radial and
mesh networks (under mild conditions, for instance, no upper bounds on loads) [20, 21]. Moreover, the
interpretation of the decision variables in a branch flow model (for instance, branch power and current flows)
are more intuitive compared to the semidefinite matrix and second order cone programming relaxations and
tend to be numerically more stable compared to second order cone programming relaxations for the bus
injection models [34].
If the rectangular coordinate branch flow model (5.10)–(5.13) is used, then a linear relaxation may be ob-
tained by introducing auxiliary variables for the quadratic terms and subsequently applying the McCormick
inequalities [22]. Although the relaxation is not exact, in practice it can be made relatively tight by intro-
ducing tailored cuts which constrain the branch power flows. The linear relaxation allows rapid generation
of lower bounds on the objective function value and facilitates mixed integer-linear relaxations for more
complex OPF variants.
5.5. Side Constraints. Besides the power flow equations, a typical OPF problem includes a number
of side constraints. The simplest of these are box constraints (2.8)–(2.11) which define the upper and lower
limits for the controllable generation, the bus voltage magnitudes, and the bus voltage angles.
If the system contains controllable phase-shifting or tap-changing transformers, then the corresponding
phase angles ϕ and tap ratios T are introduced into the set of control variables. The control variable vector
u becomes
G
, QG


u = Pi:i∈G i:i∈G , ϕik:ik∈H , Tik:ik∈K ,

where H and K are the sets of branches with controllable-phase shifting transformers and tap-changing
transformers, respectively. Since ϕ and T alter the elements of admittance matrix Ye , the left hand sides of
(2.6) and (2.7) become functions of ϕ and T : Pi (V, δ, ϕ, T ) and Qi (V, δ, ϕ, T ), respectively. The formulation
is also augmented with bound constraints on the phase angles

ϕmin max
ik ≤ ϕik ≤ ϕik ∀ ik ∈ H (5.19)

and the tap ratios

min max
Tik ≤ Tik ≤ Tik ∀ ik ∈ K. (5.20)

Adding controllable phase-shifting or tap-changing transformers significantly complicates the OPF problem,
rendering many common approximations and relaxations invalid. Some OPF research has therefore explored
alternative models for controllable transformers and other advanced power flow control devices, such as the
inclusion of auxiliary buses or idealization of such transformers as controlled power injections [3, 32, 58].
Many recent OPF formulations also include line limits, that is, upper bounds on the branch current
magnitudes. Since line current is voltage dependent, an exact bound requires a function inequality constraint.
 AN INTRODUCTION TO OPTIMAL POWER FLOW 19

By Ohm’s law, the current magnitude in branch ik is

Iik = Vei − Vek yik ,

in which yik is the magnitude of the branch admittance. Thus, one version of the exact bound is
max
Vei − Vek yik ≤ Iik ,
max
Iik
q
2 2
⇔ (Vi cos δi − Vk cos δk ) + (Vi sin δi − Vk sin δk ) ≤ ,
yik
max 2
2 2 (Iik )
⇔ (Vi cos δi − Vk cos δk ) + (Vi sin δi − Vk sin δk ) ≤ 2 ∀ ik ∈ L. (5.21)
yik
(For branches with off-nominal turns ratios, the tap bus voltage Vi and angle δi in (5.21) must be corrected
for the off-nominal turns ratio. In this case, Vi is replaced by Vi0 = Vi /Tik and δi is replaced by δi0 = δi − ϕik .)
Rather than bounding the square of the current as is given in (5.21), most formulations instead bound
the real and reactive power flow in the line,
2 2
Pik + Q2ik ≤ (Sik
max
) .
In almost all cases, this simplification is sufficiently accurate.
6. Solution Methods. In this section, we provide a concise discussion of OPF solution methods and
their relationship to conventional PF. Because many OPF algorithms rely on or incorporate aspects of
conventional PF algorithms, we summarize conventional PF solution methods first.
6.1. Solution Methods for Conventional PF. Recall from Section 2.3 that conventional PF seeks
a numerical solution to the exactly determined system of complex nonlinear equations
 ∗
S = Ve ◦ Ye Ve .

Newton’s method is typically used to solve this system and is fully described in power systems texts [26,56,62].
Here, we summarize the solution of the AC power flow equations with voltage polar coordinates. The solution
method for voltage rectangular coordinates is similar; Zhu [62] provides a good summary.
The 1st order Taylor series approximation of (5.1)–(5.2) about the current estimate of V and δ yields
 
∂P ∂P
   
∆P  ∂δ ∂V  ∆δ
≈ ∂Q ∂Q  ∆V ,

∆Q

∂δ ∂V
∆δ
≈J , (6.1)
∆V
in which J is the Jacobian matrix of system (5.1)–(5.2). At each iteration, the mismatches in the power flow
equations are
∆Pi = PiG − PiL − Pi (V, δ) ,


(6.2)
G L


∆Qi = Qi − Qi − Qi (V, δ) . (6.3)
Newton’s method consists of iteratively solving (6.1) for the ∆δ and ∆V required to correct the mismatch in
the power flow equations computed from (6.2)–(6.3). Newton’s method is locally quadratically convergent.
Therefore, given a sufficiently good starting point, the method reliably finds the correct solution to the
power flow equations. However, convergence is not guaranteed: the method can fail to converge if the
starting voltage estimate is poor, as can occur in a stressed power system [13]. To address the convergence
issue, a few researchers have applied globally convergent solution methods [29] to conventional PF and
OPF, including the application of homotopy [18], a backtracking line search method [41], and trust region
methods [41, 49]. Such methods are generally robust with respect to poor quality starting points but incur
a significant computational penalty.
20 STEPHEN FRANK AND STEFFEN REBENNACK

6.1.1. Decoupled Power Flow Algorithms. In practical power systems, real power injections are
strongly coupled to voltage angles and reactive power injections are strongly coupled to voltage magnitudes.
Conversely, real power injections are weakly coupled to voltage magnitudes and reactive power injections are
weakly coupled to voltage angles. This feature has led to the development of decoupled solution methods for
the power flow equations [26, 62]. The most basic decoupling method is to use a set of approximate Taylor
series expansions of the form

∂P
∆P ≈ ∆δ,
∂δ
∂Q
∆Q ≈ ∆V.
∂V
This allows the use of separate Newton updates for δ and V with correspondingly smaller matrices—a
significant computational advantage.
Although decoupled PF uses an approximate update method, it still computes exact real and reactive
power mismatches ∆P and ∆Q from (6.2)–(6.3) and updates both V and δ at each iteration. Decoupled PF
therefore is locally convergent to the exact solution to the power flow. However, because of the approximated
Jacobian matrix, more iterations are required for convergence [26]. Zhu [62] discusses several decoupled PF
variants in detail.

6.1.2. DC Power Flow Algorithms. In DC power flow, there is no distinction between PV and PQ
buses because all voltage magnitudes are considered to be 1.0 and reactive power is neglected. As with the
AC power flow, the slack bus angle is fixed. Because the DC power flow equations are linear, they may be
solved directly for the voltage angles using

δ = B −1 P.

(This is simply a solved matrix representation of (5.16).)

6.2. Solution Methods for Optimal Power Flow. The existing OPF literature describes an im-
mense variety of approaches and solution methods. As mentioned in the introduction, our intent in this
paper is to provide an operations researcher with sufficient information to formulate OPF problems, under-
stand the challenges and open problems in the field, and comprehend the electrical engineering dominated
literature on OPF solution methods. Therefore, we refrain from summarizing the vast literature on tailored
OPF solution methods and instead refer the reader to the two surveys [23, 24]. We do, however, highlight
several popular methods and their relationship to conventional PF.
Many classical nonlinear optimization techniques have been applied to the OPF problem, including
gradient descent methods, Newton’s method, sequential linear programming (SLP), and sequential quadratic
programming (SQP). Recall from Section 2.4 that many OPF algorithms partition the decision variables into
a set of control variables u and a set of state variables x. At each search step, the algorithm fixes u and
derives x by solving a conventional PF. When this approach is used, the Jacobian matrix J from conventional
PF plays several important roles:
1. It provides the linearization of the power flow equations required for SLP and SQP iterations,
2. It provides sensitivities in the power flow injections with respect to the state variables,
3. It provides a direct calculation of portions of the Hessian matrix of the Lagrangian function for OPF
algorithms based on Newton’s method (see [52]), and
4. It is therefore often used to improve computational efficiency in computing the KKT conditions for
the Lagrangian function.
More recently, nonlinear interior point methods (IPMs) and various convex relaxation approaches such
as semidefinite programming have become popular for OPF [23]. Many such methods rely heavily on specific
characteristics of the PF equations in order to establish convergence properties or evaluate the tightness of the
relaxation. Many meta-heuristic methods and hybrid methods combining meta-heuristics with deterministic
approaches have also been applied to OPF [24]. In the case of meta-heuristics, the typical approach is again
to select a trial vector for u and compute the corresponding state vector x and objective function value using
conventional PF.
 AN INTRODUCTION TO OPTIMAL POWER FLOW 21

6.3. Practical Considerations. In our experience, there are several practical and computational
aspects of OPF stemming from the power flow equations that can cause confusion or lead to solution errors.
Two of these is the use of the per-unit system and formats for the exchange of OPF data, discussed in
Appendices C and D, respectively. We note several others here.
6.3.1. Decoupled Solution Methods for OPF. Like decoupled PF, decoupled OPF algorithms take
advantage of the strong P -δ and Q-V relationships in the power flow equations by formulating separate real
and reactive OPF subproblems. These subproblems and their optima are assumed to be independent. Unlike
decoupled PF, however, decoupled OPF solves the subproblems sequentially rather than simultaneously: the
real subproblem solves for the optimal values of P and δ while holding Q and V constant, and the reactive
subproblem solves for the optimal values of Q and V while holding P and δ constant [17, 52]. To enable this
decomposition, decoupled OPF neglects the weak P -V and Q-δ relationships in the constraints, introducing
slight errors in the computed power flows. Decoupled OPF is therefore distinctly different from decoupled
PF in that the decoupled OPF solution is inexact.
The error in the decoupled OPF solution with respect to the true optimum is a function of the accuracy
of the decoupling assumptions and can be significant if the system includes advanced control devices [61].
These assumptions should therefore be evaluated for accuracy if a decoupled OPF approach is considered.
In the OPF literature, it is not always immediately apparent whether decoupled OPF is in use or whether a
decoupled PF procedure is used within the solution algorithm for a coupled OPF. Because of the implications
for the OPF solution quality, the careful reader should try to discern which is the case.
6.3.2. Degrees versus Radians. Power systems engineers usually report angles in degrees, including
in data files for OPF (see Appendix D). For computation, these angles must be converted to radians, for two
reasons:
1. Nearly all optimization software and algebraic modeling languages—including AMPL and GAMS—
implement trigonometric functions in radians, not degrees.
2. Even when using the DC power flow equations (which require no trigonometric function evaluations),
radians must be used due to the derivation of the equations. If degrees are used, the powers computed
from DC power flow will have a scaling error of 180/π.
Power flow software typically handles these conversions transparently, accepting input and giving output
in degrees. Thus, it can be difficult to remember that general-purpose optimization software requires an
explicit conversion.
6.3.3. System Initialization. In both conventional PF and OPF, the convergence of the power flow
equations depends strongly on the selection of a starting point. Given a starting point far from the correct
solution, the power flow equations may converge to a meaningless solution, or may not converge at all. In
the absence of a starting point, standard practice is to initialize all voltage magnitudes to 1.0 p.u. and all
voltage angles to zero; this is called a “cold start” or a “flat start”. The alternative is a “warm start”, in
which the voltages and angles are initialized to the solution of a pre-solved power flow. Warm starts are
often used in online OPF to minimize computation time and ensure that the search begins from the current
system operating condition. Warm starts also aid the algorithm in converging to the nearest local minimum,
which is often desirable during real-time operation of power systems.
If a good starting point is unavailable or the OPF algorithm encounters convergence problems, it is
also possible to switch to a more robust solution method. Trust region methods and their variants coupled
with IPMs are a popular and effective approach [41, 49, 55]. The primary disadvantage of such methods is a
significant computationally penalty compared to their less-robust counterparts.
7. Numerical Example. We now present a small numerical example to illustrate some key elements
of a typical OPF formulation. The network of Figure 7.1 provides the basis for the example OPF problem.
It has N = 5 buses and L = 6 branches, with corresponding sets
N = {1, 2, 3, 4, 5}
and
L = {(1, 2), (1, 3), (2, 4), (3, 4), (3, 5), (4, 5)} .
Using this network, we derive the admittance matrix (Section 7.1), define the power flow equations (Section
7.2), and develop the classical OPF problem and its optimal solution (Section 7.3).
22 STEPHEN FRANK AND STEFFEN REBENNACK

Bus 2 Bus 4
Branch (2,4)

2)
(1,
n ch
a
Bus 1 Br )
,4
(3

Branch (4,5)
ch
r an
Br B
an
ch
(1,
3)

Branch (3,5)

Bus 3 Bus 5

Fig. 7.1: Bus and branch indices in an example 5-bus electrical network.

Table 7.1: Branch impedance data for the network of Figure 7.1. All quantities except phase angles are
given in per-unit. Dots indicate nominal voltage ratios and phase angles.

From Bus To Bus Series Resistance Series Reactance Shunt Susceptance Voltage Ratio Phase Angle
i k Rik Xik bSh
ik Tik ϕik
1 2 0.000 0.300 0.000 · ·
1 3 0.023 0.145 0.040 · ·
2 4 0.006 0.032 0.010 · ·
3 4 0.020 0.260 0.000 · −3.0°
3 5 0.000 0.320 0.000 0.98 ·
4 5 0.000 0.500 0.000 · ·

7.1. Admittance Matrix. Table 7.1 provides a set of branch data for the example system. Note that
branch (3, 4) is a phase-shifting transformer and branch (4, 5) has an off-nominal voltage ratio. In addition
to the branch data, bus 2 has a shunt susceptance of j0.30 pu and bus 3 has a shunt conductance of 0.05 pu.
To compute the admittance matrix for this system, we first compute the series admittance yeik of each
branch using (4.7). For example, the series admittance of branch (1, 3) is

0.023 0.145
ye13 = 2 2
−j ≈ 1.067 − j6.727.
0.023 + 0.145 0.0232 + 0.1452
The remaining branches have series admittances

ye12 ≈ 0.000 − j3.333,

ye24 ≈ 5.660 − j30.189,
ye34 ≈ 0.294 − j3.824,
ye35 ≈ 0.000 − j3.125,

and

ye45 ≈ 0.000 − j2.000.

Next, we construct Ye using (4.8)–(4.9). For example, diagonal element Ye33 consists of summing the
admittances of branches (1, 3), (3, 4), and (3, 5), plus the contributions of the shunt conductance at bus 3
and the shunt susceptance of branch (1, 3). Branch admittances ye34 and ye35 have off-nominal turns ratios

a34 = 1.0e−j3.0° ≈ 0.999 − j0.052

 AN INTRODUCTION TO OPTIMAL POWER FLOW 23

and

a35 = 0.98e−j0.0° = 0.980,

respectively. (Note that a34 a∗34 = 1.0 if rounding errors are neglected.) Therefore, matrix element Ye33 is

0.04 0.294 − j3.824 j3.125

Ye33 ≈ (1.067 − j6.727) + j + − + 0.05 ≈ 1.41 − j13.78.
2 (0.999 − j0.052) (0.999 + j0.052) 0.9802

An example off-diagonal element is Ye34 , which from (4.9) is

0.294 − j3.824
Ye34 ≈ − ≈ −0.09 + j3.83.
(0.999 + j0.052)

The full admittance matrix is

 
1.07 − j10.04 0.00 + j3.33 −1.07 + j6.73 0 0
 0.00 + j3.33
 5.66 − j33.22 0 −5.66 + j30.19 0 

Ye ≈ 
−1.07 + j6.73 0 1.41 − j13.78 −0.09 + j3.83 .
0.00 + j3.19
 0 −5.66 + j30.19 −0.49 + j3.80 5.95 − j36.01 0.00 + j2.00
0 0 0.00 + j3.19 0.00 + j2.00 0.00 − j5.13

7.2. Power Flow Equations. Using the previously developed admittance matrix, we can write the
real and reactive power flow equations for any bus in the 5-bus example system. For example, from (5.1),
the real power injection at bus 1 is
5
X

P1 (V, δ) = V1 Vk G1k cos (δ1 − δk ) + B1k sin (δ1 − δk ) ,
k=1
≈ 1.07V12 cos (δ1 − δ1 ) − 1.07V1 V3 cos (δ1 − δ3 ) − 10.04V12 sin (δ1 − δ1 )
+ 3.33V1 V2 sin (δ1 − δ2 ) + 6.73V1 V3 sin (δ1 − δ3 ),
≈ 1.07V12 − 1.07V1 V3 cos (δ1 − δ3 ) + 3.33V1 V2 sin (δ1 − δ2 ) + 6.73V1 V3 sin (δ1 − δ3 ).

Similarly, from (5.2), the reactive power injection at bus 1 is

5
X

Q1 (V, δ) = V1 Vk G1k sin (δ1 − δk ) − B1k cos (δ1 − δk ) ,
k=1
≈ 1.07V12 sin (δ1 − δ1 ) − 1.07V1 V3 sin (δ1 − δ3 ) + 10.04V12 cos (δ1 − δ1 )
− 3.33V1 V2 cos (δ1 − δ2 ) − 6.73V1 V3 cos (δ1 − δ3 ),
≈ −1.07V1 V3 sin (δ1 − δ3 ) + 10.04V12 − 3.33V1 V2 cos (δ1 − δ2 ) − 6.73V1 V3 cos (δ1 − δ3 ).

7.3. Optimal Power Flow Formulation and Solution. We now develop the classical OPF for-
mulation for the 5-bus example system presented in Figure 7.1, with the addition of two advanced control
elements. The branch impedance data are as given in Table 7.1, except that we assign ϕ34 and T35 to be
decision variables representing a phase shifting transformer and an on-load tap changer, respectively. ϕ34
and T35 have limits

−30.0° ≤ ϕ34 ≤ 30.0°

and

0.95 ≤ T35 ≤ 1.05.

Tables 7.2 and 7.3 give the bus data (voltage limits, load, and generation). The system power base is 100
MW.
24 STEPHEN FRANK AND STEFFEN REBENNACK

Table 7.2: Bus data for the network of Figure 7.1. All quantities are given in per-unit. Dots indicates zero
values.

Bus Load Real Power Load Reactive Power Min. Bus Voltage Max. Bus Voltage
i PiL QLi Vimin Vimax
1 · · 1.00 1.00
2 · · 0.95 1.05
3 · · 0.95 1.05
4 0.900 0.400 0.95 1.05
5 0.239 0.129 0.95 1.05

Table 7.3: Generator data for the network of Figure 7.1. All quantities are given in per-unit.

Min. Generator Max. Generator Min. Generator Max. Generator

Bus Real Power Real Power Reactive Power Reactive Power
i PiG,min PiG,max QG,min
i QG,max
i
1 −∞ ∞ −∞ ∞
3 0.10 0.40 −0.20 0.30
4 0.05 0.40 −0.20 0.20

Given this data, the sets defining the formulation are:

N = {1, 2, 3, 4, 5} ,
G = {1, 3, 4} ,
L = {(1, 2), (1, 3), (2, 4), (3, 4), (3, 5), (4, 5)} ,
H = {(3, 4)} ,

and

K = {(3, 5)} .

Let the three generator cost functions, in thousands of dollars, be

C1 P1G = 0.35P1G ,



2
C3 P3G = 0.20P3G + 0.40 P3G ,



2
C4 P4G = 0.30P4G + 0.50 P4G ,

in which the PiG are expressed in per-unit.

To develop the full formulation, it is first necessary to re-write Ye as derived in Section 7.1 to explicitly
include ϕ34 and T35 . Let

a34 = cos ϕ34 + j sin ϕ34

and

a35 = T35 .

(Note that a34 a∗34 = 1.0, 1/a34 = cos ϕ34 − j sin ϕ34 = a∗34 , and 1/a∗34 = cos ϕ34 + j sin ϕ34 = a34 .) Then,
 AN INTRODUCTION TO OPTIMAL POWER FLOW 25

using (4.8)–(4.9) and simplifying,

1
Ye33 = 1.41 − j10.53 − j 2 · 3.13,
T35
Ye34 = −0.29 cos ϕ34 − 3.82 sin ϕ34 + j (3.82 cos ϕ34 − 0.29 sin ϕ34 ) ,
Ye43 = −0.29 cos ϕ34 + 3.82 sin ϕ34 + j (3.82 cos ϕ34 + 0.29 sin ϕ34 ) ,
1
Ye35 = j · 3.13,
T35
and
1
Ye53 = j · 3.13.
T35

Ye44 , Ye55 , and the other remaining matrix elements are unchanged.
Bus 1 is the system slack bus, and therefore Ve1 is fixed to 1.0∠0.0°. To construct the formulation, we
round all numerical values to two decimal places. (This rounding does not affect model feasibility because
sufficient degrees of freedom exist in the state variables.) Following (2.5)–(2.11), (5.19), and (5.20), the full
formulation is

2
2
min 0.35P1G + 0.20P3G + 0.40 P3G + 0.30P4G + 0.50 P4G ,
s.t. P1G = 1.07 − 1.07V3 cos (−δ3 ) + 3.33V2 sin (−δ2 ) + 6.73V3 sin (−δ3 ) ,
0 = 5.66V22 − 5.66V2 V4 cos (δ2 − δ4 )
+ 3.33V2 sin (δ2 ) + 30.19V2 V4 sin (δ2 − δ4 ) ,
P3G = 1.41V32 − 1.07V3 cos (δ3 )
+ (−0.29 cos ϕ34 − 3.82 sin ϕ34 ) V3 V4 cos (δ3 − δ4 )
+ 6.73V3 sin (δ3 ) + (3.82 cos ϕ34 − 0.29 sin ϕ34 ) V3 V4 sin (δ3 − δ4 )
3.13
+ V3 V5 sin (δ3 − δ5 ) ,
T35
P4G − 0.900 = 5.95V42 − 5.66V4 V2 cos (δ4 − δ2 )
+ (−0.29 cos ϕ34 + 3.82 sin ϕ34 ) V4 V3 cos (δ4 − δ3 )
+ 30.19V4 V2 sin (δ4 − δ2 )
+ (3.82 cos ϕ34 + 0.29 sin ϕ34 ) V4 V3 sin (δ4 − δ3 ) + 2.00V4 V5 sin (δ4 − δ5 ) ,
3.13
−0.239 = V5 V3 sin (δ5 − δ3 ) + 2.00V5 V4 sin (δ5 − δ4 ) ,
T35
QG
1 = 10.04 − 1.07V3 sin (−δ3 ) − 3.33V2 cos (−δ2 ) − 6.73V3 cos (−δ3 ) ,
0 = −5.66V2 V4 sin (δ2 − δ4 ) + 33.22V22
− 3.33V2 cos (δ2 ) − 30.19V2 V4 cos (δ2 − δ4 ) ,
QG
3 = −1.07V3 sin (δ3 ) + (−0.29 cos ϕ34 − 3.82 sin ϕ34 ) V3 V4 sin (δ3 − δ4 )
 
3.13
+ 10.53 + 2 V32 − 6.73V3 cos (δ3 )
T35
3.13
− (3.82 cos ϕ34 − 0.29 sin ϕ34 ) V3 V4 cos (δ3 − δ4 ) − V3 V5 cos (δ3 − δ5 ) ,
T35
QG
4 − 0.400 = −5.66V4 V2 sin (δ4 − δ2 )
+ (−0.29 cos ϕ34 + 3.82 sin ϕ34 ) V4 V3 sin (δ4 − δ3 )
+ 36.01V42 − 30.19V4 V2 cos (δ4 − δ2 )
− (3.82 cos ϕ34 + 0.29 sin ϕ34 ) V4 V3 cos (δ4 − δ3 ) − 2.00V4 V5 cos (δ4 − δ5 ) ,
3.13
−0.129 = 5.13V52 − V5 V3 cos (δ5 − δ3 ) − 2.00V5 V4 cos (δ5 − δ4 ) ,
T35
26 STEPHEN FRANK AND STEFFEN REBENNACK

0.10 ≤ P3G ≤ 0.40,

0.05 ≤ P4G ≤ 0.40,
−0.20 ≤ QG
3 ≤ 0.30,

−0.20 ≤ QG4 ≤ 0.20,

−30.0° ≤ ϕ34 ≤ 30.0°,
0.95 ≤ T35 ≤ 1.05,
0.95 ≤ Vi ≤ 1.05, i ∈ {2, 3, 4, 5} ,
−180.0° ≤ δi ≤ 180.0°, i ∈ {2, 3, 4, 5} ,
P1G , QG
1 unrestricted.

Voltage angles δ1 , δ2 , δ3 , and δ4 are restricted to one full sweep of the unit circle. Slack bus generator powers
P1G and QG1 are unrestricted, and branch current limits are neglected.
For this formulation, the vector of control variables is

u = P1G , P3G , P4G , QG G G

1 , Q3 , Q4 , ϕ34 , T35

and the vector of state variables is

x = (δ2 , δ3 , δ4 , δ5 , V2 , V3 , V4 , V5 ) .

The optimal solution for this formulation is

V2 ≈ 0.981, V3 ≈ 0.957, V4 ≈ 0.968, V5 ≈ 0.959,

δ2 ≈ −12.59°, δ3 ≈ −1.67°, δ4 ≈ −13.86°, δ5 ≈ −9.13°,
P1G ≈ 0.947, P3G ≈ 0.192, P4G ≈ 0.053,
QG1 ≈ 0.387, QG3 ≈ −0.127, QG4 ≈ 0.200,
ϕ34 ≈ 12.38°, T35 ≈ 0.95,

with objective function value 0.4016596. If the controllable phase-shifting and tap-changing transformers
are instead fixed to ϕ34 = −3.0° and T35 = 0.98, as originally specified, the optimal solution becomes

V2 ≈ 0.983, V3 ≈ 0.964, V4 ≈ 0.970, V5 ≈ 0.950,

δ2 ≈ −7.50°, δ3 ≈ −4.22°, δ4 ≈ −8.20°, δ5 ≈ −8.64°,
P1G ≈ 0.946, P3G ≈ 0.195, P4G ≈ 0.058,
QG1 ≈ 0.249, QG3 ≈ −0.072, QG4 ≈ 0.200,

with objective function value 0.4041438, a cost increase of approximately 0.6%.

As an illustration, we implemented three versions of the classical formulation (2.5)–(2.11), augmented
with constraints (5.19) and (5.20), in the GAMS modeling language. The three versions each use a different
form of the power flow equations: (i) polar voltage coordinates with rectangular admittance coordinates
(5.1)–(5.2), (ii) polar voltage coordinates with polar admittance coordinates (5.3)–(5.4), and (iii) rectangular
voltage coordinates with rectangular admittance coordinates (5.5)–(5.6). The model is publicly available
in the publisher’s online edition of IIE Transaction. For the example, the model yielded identical optimal
solutions using three local nonlinear solvers, SNOPT, MINOS, and CONOPT, and verified as globally optimal
using the global solver LINDOGlobal.
8. Conclusion. In this paper, we have addressed the modeling of power systems, the OPF formulation
and common variants, and both theoretical and practical aspects of OPF problems. For the reader interested
in learning more, particularly regarding optimization algorithms that have been used for OPF, we recommend
any or all of the following:
1. Read the classical papers on OPF, for instance [7, 19, 50, 52]. These papers provide a detailed
discussion of the foundations of OPF and provide context for more recent work.
 AN INTRODUCTION TO OPTIMAL POWER FLOW 27

2. Review textbooks which describe the OPF problem [56, 62]. These textbooks provide clear, detailed
formulations and also provide lists of relevant references, although they often omit foundational
material on power system modeling.
3. Review the survey papers on OPF from the past several decades, for instance [23, 24, 30, 35, 38].
Reading the older surveys prior to the more recent ones provides insight into how OPF has developed
over time.
4. Experiment with the GAMS OPF formulations provided to accompany Section 7, which are available
in the online appendix of IIE Transactions. Alternatively, install and experiment with the OPF
capabilities available in MATPOWER [64]. Experimentation with either software will provide insight
into the practical challenges of OPF.
The material presented in this paper should provide a sufficient foundation for understanding the content of
the references cited above and elsewhere in the paper.
In recent years, OPF has become one of the most widely researched topics in electric power systems
engineering. We hope that this paper encourages a similar level of engagement within the operations research
community, particularly in the development of new, efficient OPF algorithms.

Acknowledgment. We thank Vitaliy Krasko, Timo Lohmann, and Greg Steeger of Colorado School of
Mines; Kathryn Schumacher of the University of Michigan; and the anonymous reviewers for their valuable
comments and suggestions during the drafting of this paper. This material is based upon work supported
by the National Science Foundation Graduate Research Fellowship under Grant No. DGE-1057607.

REFERENCES

[1] Power Systems Test Case Archive. Internet site. Available: http://www.ee.washington.edu/research/pstca/.
[2] E. Acha, H. Ambriz-Perez, and C.R. Fuerte-Esquivel, Advanced transformer control modeling in an optimal power
flow using newton’s method, IEEE Transactions on Power Systems, 15 (2000), pp. 290–298.
[3] M.M. Adibi, R.A. Polyak, I.A. Griva, L. Mili, and S. Ammari, Optimal transformer tap selection using modified
barrier-augmented lagrangian method, IEEE Transactions on Power Systems, 18 (2003), pp. 251–257.
[4] K.C. Almeida and F.D. Galiana, Critical cases in the optimal power flow, IEEE Transactions on Power Systems, 11
(1996), pp. 1509–1518.
[5] O. Alsac, J. Bright, M. Praise, and B. Stott, Further developments in LP-based optimal power flow, IEEE Transac-
tions on Power Systems, 5 (1990), pp. 697–711.
[6] O. Alsac and B. Stott, Optimal load flow with steady-state security, IEEE Transactions on Power Apparatus and
Systems, PAS-93 (1974), pp. 745–751.
[7] A.T. Azevedo, A.R.L. Oliveira, M.J. Rider, and S. Soares, How to efficiently incorporate facts devices in optimal
active power flow model, Journal of Industrial and Management Optimization, 6 (2010), pp. 315–331.
[8] X. Bai and H. Wei, Semi-definite programming-based method for security-constrained unit commitment with operational
and optimal power flow constraints, IET Generation, Transmission & Distribution, 3 (2009), pp. 182–197.
[9] M.E. Baran and F.F. Wu, Optimal capacitor placement on radial distribution systems, IEEE Transactions on Power
Delivery, 4 (1989), pp. 725–734.
[10] , Optimal sizing of capacitors placed on a radial distribution system, IEEE Transactions on Power Delivery, 4 (1989),
pp. 735–743.
[11] D. Bertsimas, D.B. Brown, and C. Caramanis, Theory and applications of robust optimization, SIAM Review, 53
(2011), pp. 464–501.
[12] D. Bienstock, Progress on solving power flow problems, OPTIMA, 93 (2013), pp. 1–7.
[13] R.C. Burchett, H.H. Happ, and K.A. Wirgau, Large scale optimal power flow, IEEE Transactions on Power Apparatus
and Systems, 101 (1982), pp. 3722–3732.
[14] F. Capitanescu and L. Wehenkel, Sensitivity-based approaches for handling discrete variables in optimal power flow
computations, IEEE Transactions on Power Systems, 25 (2010), pp. 1780–1789.
[15] J. Carpentier, Contribution to the economic dispatch problem, Bulletin de la Société Française des Electriciens, 8 (1962),
pp. 431–447.
[16] G.C. Contaxis, C. Delkis, and G. Korres, Decoupled optimal load flow using linear or quadratic programming, IEEE
Transactions on Power Systems, PWRS-I (1986), pp. 1–7.
[17] S. Cvijić, P. Feldmann, and M. Ilić, Applications of homotopy for solving ac power flow and ac optimal power flow,
in 2012 IEEE Power and Energy Society General Meeting, July 2012, pp. 1–8.
[18] H.W. Dommel and W.F. Tinney, Optimal power flow solutions, IEEE Transactions on Power Apparatus and Systems,
87 (1968), pp. 1866–1876.
[19] M. Farivar and S.H. Low, Branch flow model: Relaxations and convexification—part i, IEEE Transactions on Power
Systems, 28 (2013), pp. 2554–2564.
[20] , Branch flow model: Relaxations and convexification—part II, IEEE Transactions on Power Systems, 28 (2013),
pp. 2565–2572.
[21] S.M. Frank and S. Rebennack, Optimal Design of Mixed AC-DC Distribution Systems for Commercial Buildings: A
28 STEPHEN FRANK AND STEFFEN REBENNACK

Nonconvex Generalized Benders Decomposition Approach, European Journal of Operational Research, 242 (2015),
pp. 710–729.
[22] S. Frank, I. Steponavice, and S. Rebennack, Optimal power flow: A bibliographic survey I, formulations and deter-
ministic methods, Energy Systems, 3 (2012), pp. 221–258.
[23] , Optimal power flow: A bibliographic survey II, non-deterministic and hybrid methods, Energy Systems, 3 (2012),
pp. 259–289.
[24] E. Freitag and R. Busam, Complex analysis, Springer, 2 ed., 2009.
[25] J.D. Glover, M.S. Sarma, and T.J. Overbye, Power system analysis and design, Cengage Learning, Stamford, CT,
4 ed., 2008.
[26] S. Granville, Optimal reactive dispatch through interior point method, IEEE Transactions on Power Systems, 9 (1994),
pp. 136–146.
[27] M. Greenberg, Advanced engineering mathematics, Prentice Hall, Upper Saddle River, NJ, 2 ed., 1998.
[28] S.P. Han, A globally convergent method for nonlinear programming, Journal of Optimization Theory and Applications,
22 (1977), pp. 297–309.
[29] M. Huneault and F.D. Galiana, A survey of the optimal power flow literature, IEEE Transactions on Power Systems,
6 (1991), pp. 762–770.
[30] J. Lavaei and S.H. Low, Zero duality gap in optimal power flow problem, IEEE Transactions on Power Systems, 27
(2012), pp. 92–107.
[31] C. Lehmköster, Security constrained optimal power flow for an economical operation of FACTS-devices in liberalized
energy markets, IEEE Transactions on Power Delivery, 17 (2002), pp. 603–608.
[32] S.H. Low, Convex relaxation of optimal power flow: A tutorial, in Bulk Power System Dynamics and Control - IX
Optimization, Security and Control of the Emerging Power Grid (IREP), Rethymnon, Greece, Aug. 2013, pp. 1–15.
[33] , Convex Relaxation of Optimal Power Flow Part I: Formulations and Equivalence, IEEE Transactions on Control
of Network Systems, 1 (2014), pp. 15–27.
[34] J.A. Momoh, M.E. El-Hawary, and R. Adapa, A review of selected optimal power flow literature to 1993 part II: Newton,
linear programming and interior point methods, IEEE Transactions on Power Systems, 14 (1999), pp. 105–111.
[35] J.A Momoh, R.J. Koessler, M. S. Bond, D. Sun, A. Papalexopoulos, and P. Ristanovic, Challenges to Optimal
Power Flow, IEEE Transactions on Power Systems, 12 (1997), pp. 444–447.
[36] J.A. Momoh, J.Z. Zhu, G.D. Boswell, and S. Hoffman, Power system security enhancement by opf with phase shifter,
IEEE Transactions on Power Systems, 16 (2001), pp. 287–293.
[37] J. A. Momoh, M.E. El-Hawary, and R. Adapa, A review of selected optimal power flow literature to 1993 part I:
Nonlinear and quadratic programming approaches, IEEE Transactions on Power Systems, 14 (1999), pp. 105–111.
[38] J. Nocedal and S. Wright, Numerical optimization, Springer, New York, NY, 2nd ed., 2006.
[39] J. O’Malley, Schaum’s outline of basic circuit analysis, Schaum’s Outline Series, McGraw-Hill, New York, NY, 2 ed.,
2011.
[40] S. Pajic and K.A. Clements, Power system state estimation via globally convergent methods, IEEE Transactions on
Power Systems, 20 (2005), pp. 1683–1689.
[41] J. Peschon, D.W. Bree, and L.P. Hajdu, Optimal power-flow solutions for power system planning, Proceedings of the
IEEE, 6 (1972), pp. 64–70.
[42] W. Qiu, A.J. Flueck, and F. Tu, A new parallel algorithm for security constrained optimal power flow with a nonlinear
interior point method, in IEEE Power Engineering Society General Meeting, 2005, pp. 2422–2428.
[43] R.L. Rardin, Optimization in operations research, Prentice Hall, Upper Saddle River, NJ, 1 ed., 1997.
[44] N.S. Rau, Issues in the path toward an RTO and standard markets, IEEE Transactions on Power Systems, 18 (2003),
pp. 435–443.
[45] , Optimization principles: Practical applications to the operation and markets of the electric power industry, Wiley-
IEEE Press, Hoboken, NJ, 2003.
[46] L. Snyder, Multi-period optimal power flow problems, OPTIMA, 93 (2013), pp. 8–9.
[47] E.M. Soler, V.A. de Sousa, and G.R.M. da Costa, A modified Primal-Dual Logarithmic-Barrier Method for solving
the Optimal Power Flow problem with discrete and continuous control variables, European Journal of Operational
Research, 222 (2012), pp. 616–622.
[48] A.A. Sousa, G.L. Torres, and C.A. Ca nizares, Robust optimal power flow solution using trust region and interior-point
methods, IEEE Transactions on Power Systems, 26 (2011), pp. 487–499.
[49] B. Stott and E. Hobson, Power system security control calculation using linear programming parts I and II, IEEE
Transactions on Power Apparatus and Systems, PAS-97 (1978), pp. 1713–1731.
[50] B. Stott, J. Jardim, and O. Alsac, DC power flow revisited, IEEE Transactions on Power Systems, 24 (2009), pp. 1290–
1300.
[51] D.I. Sun, B. Ashley, B. Brewer, A. Hughes, and W.F. Tinney, Optimal power flow by Newton approach, IEEE
Transactions on Power Apparatus and Systems, 103 (1984), pp. 2864–2880.
[52] G.L. Torres and V.H. Quintana, An interior-point method for nonlinear optimal power flow using voltage rectangular
coordinates, IEEE Transactions on Power Systems, 13 (1998), pp. 1211–1218.
[53] L.S. Vargas, V.H. Quintana, and A. Vannelli, A tutorial description of an interior point method and its applications
to security-constrained economic dispatch, IEEE Transactions on Power Systems, 8 (1993), pp. 1315–1323.
[54] H. Wang, C.E. Murillo-Sánchez, R.D. Zimmerman, and R.J. Thomas, On computational issues of market-based
optimal power flow, IEEE Transactions on Power Systems, 22 (2007), pp. 1185–1193.
[55] A.J. Wood and B.F. Wollenberg, Power generation, operation, and control, John Wiley & Sons, Inc., New York, NY,
2 ed., 1996.
[56] Working Group on a Common Format for Exchange of Solved Load Flow Data, Common format for exchange
of solved load flow data, IEEE Transactions on Power Apparatus and Systems, PAS-92 (1973), pp. 1916–1925.
 AN INTRODUCTION TO OPTIMAL POWER FLOW 29

[57] Y. Xiao, Y. H Song, and Y.-Z. Sun, Power flow control approach to power systems with embedded facts devices, IEEE
Transactions on Power Systems, 17 (2002), pp. 943–950.
[58] W. Zhang, F. Li, and L.M. Tolbert, Review of reactive power planning: Objectives, constraints, and algorithms, IEEE
Transactions on Power Systems, 22 (2007), pp. 2177–2186.
[59] W. Zhang and L.M. Tolbert, Survey of reactive power planning methods, in IEEE Power Engineering Society General
Meeting, vol. 2, June 2-16 2005, pp. 1430–1440.
[60] X.-P. Zhang, Fundamentals of electric power systems, John Wiley & Sons, Inc., 2010, ch. 1, pp. 1–52.
[61] J. Zhu, Optimization of power system operation, Wiley-IEEE, Piscataway, NJ, 2009.
[62] R.D. Zimmerman and C.E. Murillo-Sánchez, MATPOWER 4.1 user’s manual, Power Systems Engineering Research
Center (PSERC), 2011.
[63] R.D. Zimmerman, C.E. Murillo-Sánchez, and R.J. Thomas, MATPOWER: Steady-state operations, planning, and
analysis tools for power systems research and education, IEEE Transactions on Power Systems, 26 (2011), pp. 12–19.

Appendix A. Notation.
This appendix documents the definitions for the mathematical symbols used throughout the paper. See
also Section 2.1 for general comments on notation.
A.1. Sets, Indices, and Dimensions. The following dimensions and indices are used in the OPF
formulations within this paper:
N total number of system buses (nodes)
L total number of system branches (arcs)
M number of system load (PQ) buses
i, k indices corresponding to system buses and branches
c contingency case index
t time period index
System branches are indexed as arcs between buses. For example, the branch between buses i and k is
denoted by (i, k) or ik.
There is no standard set notation within the OPF literature. (Many authors do not use sets in their
formulations.) For convenience, however, we adopt the following sets in this article:
N set of system buses (nodes)
L set of system branches (arcs)
M set of load (PQ) buses
G set of controllable generation buses
H set of branches with controllable phase-shifting transformers
K set of branches with controllable tap-changing transformers
Q set of planned locations (buses) for new reactive power sources
C set of power system contingencies for contingency analysis
T set of time-periods for multi-period OPF
Remarks:
1. We use L to indicate the number of system branches because B is reserved for the bus susceptance
matrix.
2. In the optimization community, c often refers to a vector of objective function coefficients. In this
article, however, we use upper case C for objective function coefficients and reserve lower case c for
the contingency case index of security-constrained economic dispatch described in Section 3.1.
3. For clarity, we use H and K to represent sets of controllable phase-shifting and tap-changing trans-
formers rather than S (often used to designate sources or scenarios) and T (often used to designate
time periods). The letters H and K otherwise have no special association with phase-shifting and
tap-changing transformers.
A.2. Units. The following electrical units are used in this article:
V Volt (unit of electrical voltage)
A Ampere (unit of electrical current)
W Watt (unit of real electrical power)
VA Volt-Ampere (unit of apparent electrical power)
VAR Volt-Ampere Reactive (unit of reactive electrical power)
Watts, Volt-Amperes, and Volt-Amperes Reactive have the same SI (Système International, that is, met-
ric) unit representation (one Volt times one Ampere), but differ in physical interpretation as described in
Appendix B.4.
30 STEPHEN FRANK AND STEFFEN REBENNACK

A.3. Electrical Quantities. In power systems analysis, electrical quantities are represented in the
frequency domain as phasor quantities (complex numbers). Complex numbers may be represented as a
single complex variable, as two real-valued variables in rectangular form a+jb, or as two real-valued variables
in polar form c∠γ; all of these notations are found in the OPF literature. (Complex number notation is
explained in more detail in Section B.3.) Here, we document the usual symbols and relationships used for
the electrical quantities; some notational exceptions exist in the literature.
A.3.1. Impedance and Admittance.
Zeik complex impedance of branch ik
Rik resistance of branch ik (real part of Zeik )
Xik reactance of branch ik (imaginary part of Zeik )
Zeik = Rik + jXik
yeik complex series admittance of branch ik
gik series conductance of branch ik (real part of yeik )
bik series susceptance of branch ik (imaginary part of yeik )
yeik = 1/Zeik = gik + jbik
S
yei complex shunt admittance at bus i
giS shunt conductance of bus i (real part of yeiS )
bSi shunt susceptance of bus i (imaginary part of yeiS )
yeiS = giS + jbSi
Sh
yeik complex shunt admittance of branch ik
Sh Sh
gik shunt conductance of branch ik (real part of yeik )
Sh Sh
bik shunt susceptance of branch ik (imaginary part of yeik )
Sh Sh Sh
yeik = gik + jbik
Yeik complex ik th element of the bus admittance matrix
Yik magnitude of ik th element of the bus admittance matrix
θik angle of ik th element of the bus admittance matrix
Gik conductance of ik th element of the bus admittance matrix (real part of Yeik )
Bik susceptance of ik th element of the bus admittance matrix (imaginary part of Yeik )
Yeik = Yik ∠θik = Gik + jBik
Note the distinction between lowercase y, g, and b and uppercase Y , G, and B: the former represents the
values corresponding to individual system branch elements, while the latter refers to the admittance matrix
which models the interaction of all system branches.
A.3.2. Voltage.
Vei complex (phasor) voltage at bus i
Vi voltage magnitude at bus i
δi voltage angle at bus i
Ei real part of complex voltage at bus i
Fi imaginary part of complex voltage at bus i
Vei = Vi ∠δi = Ei + jFi
A.3.3. Current.
Iei complex (phasor) current injected at bus i
Ii magnitude of current injected at bus i
Ieik complex (phasor) current in branch ik, directed from bus i to bus k
Iik magnitude of current in branch ik
A.3.4. Power.
PiL load (demand) real power at bus i
QLi load (demand) reactive power at bus i
SiL load (demand) complex power at bus i
SiL = PiL + jQLi
PiG generator (supply) real power at bus i
QGi generator (supply) reactive power at bus i
 AN INTRODUCTION TO OPTIMAL POWER FLOW 31

i +
+
Vij i

−
Ii`
`
Vik j j
+ Ik` I`j
−
k − Vjk k

(+ Clockwise) (+ In)
Vij + Vjk − Vik = 0 Ii` + Ik` − I`j = 0

(a) Kirchoff’s Voltage Law (b) Kirchoff’s Current Law

Fig. B.1: Illustration of Kirchoff’s laws.

SiG generator (supply) complex power at bus i

SiG = PiG + jQGi
Pi net real power injection at bus i (Pi = PiG − PiL )
Qi net reactive power injection at bus i (Qi = QG L
i − Qi )
Si net complex power injection at bus i
Si = SiG − SiL = Pi + jQi
A.3.5. Other.
ϕik phase shift of phase-shifting transformer in branch ik
Tik tap ratio of tap-changing transformer in branch ik
A.3.6. Comments. The phasor indicator e is omitted for complex power S, as S is always understood
to be a complex quantity. In the literature, the indicator e is also often omitted for V , I, y, and Y , but we
include it here to disambiguate the complex quantities from their associated (real-valued) magnitudes.
Appendix B. Fundamental Concepts. This appendix summarizes several fundamental concepts of
electrical circuit theory and power systems analysis. For further background, we refer interested readers to
the excellent textbooks by O’Malley [40] and Glover, Sarma, and Overbye [26].
The fundamental electrical quantities are voltage and current. Voltage describes the potential energy
per unit of charge between two nodes in an electric circuit, while current describes the flow rate of electric
charge between circuit nodes. Voltage and current are analogous to pressure and fluid flow in a hydraulic
system. Voltage multiplied by current yields power, that is, energy transfer per unit time.
B.1. Ohm’s Law. Ohm’s law describes the relationship between the voltage V across and the current
I flowing through an electrical circuit element, such as a transmission line. Ohm’s law states

V = IR,

in which resistance R provides the constant of proportionality between voltage and current. Ohm’s law as
written here applies to direct current (DC) circuits, but is readily extended to the steady-state analysis of
alternating current (AC) circuits (see Appendix B.3).
B.2. Kirchoff ’s Laws. Electrical circuits consist of nodes (physical points of interconnection) con-
nected by circuit elements, which provide paths for electrical current. Kirchoff’s laws govern the physical
behavior of electric circuits. Kirchoff’s voltage law (KVL) states that the sum of voltages around a closed
loop in an electrical circuit is equal to zero (Figure B.1a). To apply KVL properly, the sign convention
for each voltage must be considered: if the positive node for the defined voltage is encountered first while
traversing the loop, then the voltage is added; if the negative node is encountered first, then the voltage is
subtracted (that is, its inverse is added).
Kirchoff’s current law (KCL) states that the sum of currents at an electrical node is equal to zero
(Figure B.1b). As with KVL, sign convention is important. Typically, currents defined as entering the node
32 STEPHEN FRANK AND STEFFEN REBENNACK

are added, while those defined as leaving the node are subtracted. Therefore, KCL may be stated equivalently
as follows: the sum of currents entering an electrical node is exactly equal to the sum of currents leaving
that node. KCL is the circuit analysis equivalent of flow balance in network theory.
KVL and KCL are complementary; both must be satisfied in any valid solution to a power flow problem.
In OPF, KVL and KCL are not enforced directly, but the construction of the admittance matrix ensures
that they remain satisfied (see Section 4).
B.3. AC Circuit Analysis. For power flow analysis, AC electric power systems are analyzed under
the assumption of sinusoidal steady-state operation. At sinusoidal steady state, all system voltages and
currents are sinusoids with fixed magnitude, frequency, and phase shift. Under these conditions, the time-
domain differential equations governing system voltage and current may be transformed into a set of complex
algebraic equations in the frequency domain. This phasor representation of the system is much easier to
solve.
A phasor represents a sinusoidal time-domain signal as a complex exponential in the frequency domain
using the relationship

c sin (2πf t + γ) cejγ

⇔
Time Domain Frequency Domain

The frequency f of the signal is fixed and therefore omitted from the phasor notation. The phasor cejγ may
be written as c∠γ in polar coordinates or as a + jb in rectangular coordinates, in which, according to Euler’s
formula,

a = c cos γ,
b = c sin γ,
p
c = a2 + b2 ,
b
γ = arctan .
a
For voltages and currents, c and γ represent magnitude and phase angle, written as Ve = V ∠δ for voltages
and Ie = I∠θ for currents. In electrical engineering, voltage and current phasors are expressed as root-
mean-square (RMS) quantities rather than peak quantities. This is done so that frequency domain power
calculations
√ yield the correct values without the need for a scaling factor. For a sinusoid, the RMS magnitude
is 1/ 2 times the peak magnitude. Thus, the time-domain voltage waveform

VPk sin (2πf t + δ)

has the frequency domain phasor

VPk jδ
√ e .
2
Ohm’s law, KVL, and KCL remain valid for AC circuit analysis. For AC quantities, Ohm’s law takes
the form

Ve = IeZ,
e

in which Ze = R + jX is the complex impedance that describes the relationship between sinusoidal voltage
V and sinusoidal current Ie for a particular circuit element. Resistance R models power consumption, while
e
reactance X represents the effect of electrical storage (capacitors and inductors). At steady state, storage
elements absorb and release energy during different portions of the AC cycle, producing a phase shift between
voltage and current. It is also common to write Ohm’s law in the form

Ye Ve = I,
e

in which Ye = 1/Ze is the admittance, the reciprocal of impedance. In rectangular coordinates, Ye = G + jB,
in which G is the conductance and B is the susceptance.
 AN INTRODUCTION TO OPTIMAL POWER FLOW 33

v (t )

Total Current
i (t ) = i d(t ) + i q(t )
t
i d(t )
Quadrature i q(t )

Direct

Instantaneous
Power Real v (t )i (t )
t
Power v (t )i d(t )
v (t )i q(t )

Reactive
Power

Fig. B.2: Conceptual illustration of real and reactive power using time domain waveforms. In the figure,
current i(t) lags voltage v(t) by 30°.

B.4. Complex Power. Power is the rate of energy transfer, that is, the derivative of energy with
respect to time. In the time domain, electrical power is the product of voltage and current. In AC power
systems, however, instantaneous electrical power fluctuates as voltage and current magnitude and polarity
change over time. For frequency domain analysis, power systems engineers use the concept of complex power
to characterize these time domain power fluctuations.
Complex power is a phasor quantity consisting of real power and reactive power. Real power represents
real work, that is, a net transfer of energy from source to load over time. Reactive power, on the other
hand, represents circulating energy—a cyclic exchange of energy that averages zero net energy transfer over
time. Reactive power is sometimes called imaginary power, both because it does not perform real work and
because it is the imaginary component of the phasor.
Real power transfer occurs when voltage and current are in phase, while reactive power transfer occurs
when voltage and current are 90° out of phase (that is, orthogonal). Given a reference voltage v(t), an
arbitrary (time domain) AC current i(t) can be represented by the sum of a direct component id (t) (in
phase with the voltage) and quadrature component iq (t) (orthogonal to the voltage). The direct component
corresponds to real power and the quadrature component reactive power, as illustrated in Figure B.2.
By convention, reactive power is considered positive when current lags voltage. Therefore, complex
power S is given by

S = Ve Ie∗ = P + jQ

and consists of orthogonal components P (real power) and Q (reactive power). (Here and elsewhere in
this paper, the symbol ∗ denotes complex conjugation rather than an optimal value; this use is typical in
34 STEPHEN FRANK AND STEFFEN REBENNACK

electrical engineering.) The magnitude of complex power, |S|, is called the apparent power and is often used
to specify power systems equipment and transmission line ratings. Complex and apparent power have units
of Volt-Amperes (VA), real power has units of Watts (W), and reactive power has units of Volt-Amperes
Reactive (VAR).
Appendix C. The Per-Unit System. In power systems analysis, electrical quantities are usually
expressed as a ratio of the actual SI quantity to a reference, or base, quantity; this transformation of
variables is called the per-unit system. Base quantities have SI units (Volts, Amperes, Ohms, etc.), while
per-unit quantities are dimensionless and are labeled using a designation, if any, of “p.u.”. The per-unit
value of an SI quantity x on a given base xBase is
x
xpu = .
xBase
Correct interpretation of the SI value of a per-unit quantity therefore requires knowledge of the base quantity.
For example, a power of 0.15 p.u. on a 10 MVA base equals 1.5 MW, but 0.15 p.u. on a 1000 MVA base
equals 150 MW. (In per-unit, real power, reactive power, and apparent power share a common base with
units of VA.)
In power systems analysis, base quantities exist for voltage, current, power, impedance, and admittance.
Once any two system bases are specified, the others are fixed exactly. In three-phase power flow analysis,
convention is to specify the voltage and power bases,

VBase = Line-to-line Voltage,

SBase = Three-phase Power,

and calculate the remaining system bases according to

SBase
IBase = √ ,
3 VBase
2
VBase VBase
ZBase = √ = ,
3I SBase
√ Base
3 IBase SBase 1
YBase = = 2 = .
VBase VBase ZBase
For mathematical convenience, power systems engineers typically set SBase to one of 10, 100, or 1000 MVA
and select VBase as the nominal line-to-line voltage at each bus. Thus, SBase is constant throughout the
power system, but the other bases are distinct at each system bus. When VBase is selected in this way, the
voltage ratio of most system transformers becomes 1:1 in per-unit, simplifying the development of the system
admittance matrix (Section 4).
With a proper selection of system bases, the per-unit system has several advantages over the SI system
of measurement, most notably
1. Per-unit quantities have consistent magnitudes on the order of 1.0, which improves the numerical
stability of power flow calculations;
2. The use of per-unit eliminates the need to distinguish between single-phase and three-phase electrical
quantities;
3. The use of per-unit eliminates the need to apply voltage and current scaling factors at the majority
of system transformers; and
4. The per-unit system is easier to interpret at a glance. (For example, per-unit voltage should always
lie within the approximate range 0.95–1.05 p.u., regardless of the SI voltage.)
An in-depth discussion of these advantages is beyond the scope of this paper; we instead refer the interested
reader to Glover et al. [26].
All calculations that can be performed in the SI system can also be performed in per-unit. However,
(i) per-unit and SI quantities cannot be mixed in calculations, and (ii) all per-unit calculations must be
performed on a consistent set of bases. Power systems texts frequently mix per-unit and SI units, for
instance, reporting power in MW but voltage in per-unit. In OPF, convention is to specify source and load
power in SI units, indicate the system power base, and specify all other quantities directly in per-unit (see
 AN INTRODUCTION TO OPTIMAL POWER FLOW 35

Table D.1: Field specification for IEEE Common Data Format bus data. The sixth column maps the field
to an index, parameter, or variable used in the classical OPF formulation given in Section 2.4. (Some fields
are used indirectly via inclusion in Ye .)

Field Columns Field Name Data Type Units Quantity in OPF

1 1–4 Bus Number Integer i (bus index)
2 6–17 Bus Name Text
3 19–20 Bus Area Integer
4 21–23 Loss Zone Numbera Integer
5 26 Bus Type Integer Specialb
6 28–33 Voltage Magnitude Numeric p.u. Vi
7 34–40 Voltage Angle Numeric deg. δi
8 41–49 Load Real Power Numeric MW PiL
9 50–58 Load Reactive Power Numeric MVAR QL i
10 59–67 Gen. Real Power Numeric MW PiG
11 68–75 Gen. Reactive Power Numeric MVAR QG i
12 77–83 Base Voltagea Numeric kV
13 85–90 Desired Voltage Numeric p.u. Vi (Specialc )
14 91–98 Max. Reactive Power Numeric MVAR QG,max
i
or
Max. Voltage Magnituded Numeric p.u. Vimax
15 99–106 Min. Reactive Power Numeric MVAR QG,min
i
or
Min. Voltage Magnituded Numeric p.u. Vimin
16 107–114 Bus Shunt Conductance Numeric p.u. giS
17 115–122 Bus Shunt Susceptance Numeric p.u. bS
i
18 124–127 Remote Bus Number Integer
a Optional field
b 0=PQ, 1=PQ (within voltage limits), 2=PV (within VAR limits), 3=Swing
c Indicates target voltage magnitude for voltage-controlled (PV) buses
d Gives reactive power limits if bus type is 2, voltage limits if bus type is 1

Appendix D). The powers must be converted to per-unit prior to evaluating the power flow equations, but
no other conversions are usually necessary.
Appendix D. Data Exchange. Two common formats for the academic exchange of power flow and
OPF case data are the IEEE Common Data Format [57] and the MATPOWER Case Format [63]. Most power
systems data are proprietary. However, a few publicly available test cases for OPF are distributed in one or
both of these formats [2, 63], and many OPF researchers use these test cases to benchmark algorithms. This
section summarizes the structure of these formats and their relationship to the classical OPF formulation;
the goal is to assist the reader in interpreting and applying the limited available published data.
D.1. The IEEE Common Data Format. The IEEE Common Data Format (CDF) was first de-
veloped in order to standardize the exchange of PF case data among utility companies [57]. It has since
been used to archive and exchange power systems test case data for the purpose of testing conventional PF
and OPF algorithms. The format includes sections, or “cards”, for title data, bus data, branch data, loss
zone data, and interchange data. (Originally, utilities exchanged CDF data by mail on paper card media.)
Only the title, bus, and branch data are relevant for classical OPF as described in this article. The full
specification for the IEEE CDF can be found in [57] and an abbreviated description is available at [2].
Each IEEE CDF data card consists of plain text with fields delimited by character column. The title
data card is a single line which includes summary information for the case, including the power base SBase in
MVA. The bus and branch data cards follow, beginning with the characters BUS DATA FOLLOWS and BRANCH
DATA FOLLOWS, respectively, and ending with the flag characters -999. Each line within the card gives the
data for a single bus or branch.
Tables D.1 and D.2 list the IEEE CDF field specifications for bus and branch data, respectively. The
36 STEPHEN FRANK AND STEFFEN REBENNACK

Table D.2: Field specification for IEEE Common Data Format branch data. The sixth column maps the
field to an index, parameter, or variable used in the classical OPF formulation given in Section 2.4. (Some
fields are used indirectly via inclusion in Ye .)

Field Columns Field Name Data Type Units Quantity in OPF

1 1–4 Tap Bus Number Integer i (from bus index)
2 6–9 Z Bus Number Integer k (to bus index)
3 11–12 Line Areaa Integer
4 13–15 Loss Zone Numbera Integer
5 17 Circuit Number Integer
6 19 Branch Type Integer Specialb
7 20–29 Branch Resistance Numeric p.u. Rik
8 30–39 Branch Reactance Numeric p.u. Xik
9 41–49 Branch Shunt Susceptance Numeric p.u. bSh
ik
10 51–55 Line Rating 1a,c Numeric MVA max
Iik
11 57–61 Line Rating 2a Numeric MVA
12 63–67 Line Rating 3a Numeric MVA
13 69–72 Control Bus Number Integer
14 74 Side Integer
15 77–82 Voltage Ratio Numeric p.u. Tik
16 84–90 Phase Angle Numeric deg. ϕik
17 91–97 Min. Voltage Tap Numeric p.u. min
Tik
or
Min. Phase Angled Numeric deg. ϕmin
ik
18 98–104 Max. Voltage Tap Numeric p.u. max
Tik
or
Max. Phase Angled Numeric deg. ϕmax
ik
19 105–111 Tap Step Size Numeric p.u.
or
Phase Angle Step Sized Numeric deg.
20 113–119 Min. Voltage Numeric p.u.
or
Min. MVar Transfer Numeric MVar
or
Min. MW Transfere Numeric MW
21 120–126 Max. Voltage Numeric p.u.
or
Max. MVar Transfer Numeric MVar
or
Max. MW Transfere Numeric MW
a Optional field
b 0=Transmission line, 1=Fixed T and ϕ, 2=Controllable T and fixed ϕ (voltage control),

3=Controllable T and fixed ϕ (MVAR control), 4=Fixed T and controllable ϕ

c Conversion to per-unit current (using rated branch voltage) is required
d Gives voltage tap limits or step if branch type is 2 or 3,

phase angle limits or step if branch type is 4

e Gives voltage limits if branch type is 2, MVAR limits if branch type is 3,

MW limits if branch type is 4

fields include a mixture of SI and per-unit quantities. Conversion of all quantities to per-unit is required
prior to use in an OPF formulation. Nominal-valued and unused fields in the data have zero entries. This
quirk of the specification requires some caution in processing the data; for example, a value of 0.0 in the
branch voltage ratio field should be interpreted as a nominal tap ratio (T = 1.0).
The IEEE CDF format is adapted to the compact exchange of system control data rather than OPF
data. The field structure therefore has several limitations:
 AN INTRODUCTION TO OPTIMAL POWER FLOW 37

1. Some IEEE CDF fields specify final variable values (for instance, voltages Vi ) for conventional power
flow. For OPF, these fields should be understood as a feasible or near feasible starting point rather
than an optimal solution. (Due to rounding, the reported solution may not be strictly feasible.)
2. The fields bus type (bus field 5) and branch type (branch field 6) specify system control methods,
and are therefore of limited use in OPF. However, the bus and branch types govern the interpretation
of certain other fields in the IEEE CDF, as described in the table footnotes. For example, for PQ
buses, bus fields 14 and 15 give voltage limits Vimax and Vimin , respectively. For PV buses, these
same fields instead give reactive power generation limits QG,max
i and QG,min
i , respectively.
3. For IEEE CDF fields which depend on the bus and branch types, the data are sufficient for conven-
tional PF but incomplete for OPF. For example, the IEEE CDF lacks voltage limits for PV buses
and reactive power generation limits at PQ buses; the field structure prevents these data from being
available. The user must supply (or assume) values for the incomplete data.
4. The IEEE CDF lacks other data required for OPF, including generator real power limits and cost
data.
Given these limitations, publicly archived IEEE CDF case data is most useful for obtaining the network
structure and associated bus and branch admittance data.
D.2. MATPOWER Case Format. MATPOWER [64] is an open-source software package for MAT-
LAB that includes functions for both conventional PF and OPF. The MATPOWER case format is a set of
standard matrix structures used to store power systems case data and closely resembles the IEEE CDF. MAT-
POWER case data consists of a MATLAB structure with fields baseMVA, bus, branch, gen, and gencost.
baseMVA is a scalar giving the system power base SBase in MVA. The remaining fields are matrices. Like
the IEEE CDF, the MATPOWER case structure uses a mixture of SI and per-unit quantities and specifies
nominal-valued branch tap ratios as 0 instead of 1.0. The format is described in detail in the MATPOWER
manual [63].
Tables D.3, D.4, and D.5 describe the bus, branch, and gen matrices. The gencost matrix has the same
number of rows as the gen matrix, but the column structure provides a flexible description of the generator
cost function. Column 1 specifies the type of cost model: 1 for piecewise linear or 2 for polynomial. Columns
2 and 3 give the generator startup and shutdown costs. The interpretation of column numbers 4 and greater
depends on the type of cost model:
• For a piecewise linear cost model, column 4 specifies the number of coordinate pairs n of the form
(P, C) that generate the piecewise linear cost function. The next 2n columns, beginning with column
5, give the coordinate pairs (P0 , C0 ), . . . , (Pn−1 , Cn−1 ), in ascending order. The units of C are $/hr
and the units of P are MW.
• For a polynomial cost model, column 4 specifies the number n of polynomial cost coefficients. The
next n columns, beginning with column 5, give the cost coefficients Cn−1 , . . . , C0 in descending order.
The corresponding polynomial cost model is Cn−1 P n−1 + . . . + C1 P + C0 . The units are such that
the cost evaluates to dollars $/hr for power given in MW.
If gencost is included, then MATPOWER case data contains nearly all the information necessary to
formulate the classical OPF problem as described in Section 2.4. However, MATPOWER makes no provision
for including transformer tap ratios or phase shifting transformer angles in the set of decision variables;
therefore, limits on these variables are not present in the data structure. The user must supply limits for
these controls if they exist in the formulation.
38 STEPHEN FRANK AND STEFFEN REBENNACK

Table D.3: Field specification for bus data matrix in MATPOWER case data (input fields only). The fifth
column maps the field to an index, parameter, or variable used in the classical OPF formulation given in
Section 2.4. (Some fields are used indirectly via inclusion in Ye .)

Column Field Description Data Type Units Quantity in OPF

1 Bus Number Integer i (bus index)
2 Bus Type Integer Speciala
3 Load Real Power Numeric MW PiL
4 Load Reactive Power Numeric MVAR QLi
5 Bus Area Integer
6 Bus Shunt Conductance Numeric MWb giS
7 Bus Shunt Susceptance Numeric MVARb bS
i
8 Voltage Magnitude Numeric p.u. Vi
9 Voltage Angle Numeric deg. δi
10 Base Voltage Numeric kV
11 Loss Zone Integer
12 Max. Voltage Magnitude Numeric p.u. Vimax
13 Min. Voltage Magnitude Numeric p.u. Vimin
a 1=PV, 2=PQ, 3=Swing, 4=Isolated
b Specified as a MW or MVAR demand for V = 1.0 p.u.

Table D.4: Field specification for branch data matrix in MATPOWER case data (input fields 1–11 only).
The fifth column maps the field to an index, parameter, or variable used in the classical OPF formulation
given in Section 2.4. (Some fields are used indirectly via inclusion in Ye .)

Column Field Description Data Type Units Quantity in OPF

1 Tap Bus Number Integer i (from bus index)
2 Z Bus Number Integer k (to bus index)
3 Branch Resistance Numeric p.u. Rik
4 Branch Reactance Numeric p.u. Xik
5 Branch Shunt Susceptance Numeric p.u. bSh
ik
6 Line Rating (Long-term) Numeric MVA max c
Iik
7 Line Rating (Short-term) Numeric MVA
8 Line Rating (Emergency) Numeric MVA
9 Voltage Ratio Numeric p.u. Tik
10 Phase Angle Numeric deg. ϕik
11 Branch Status Binary
a Conversion to per-unit current (using rated branch voltage) is required
 AN INTRODUCTION TO OPTIMAL POWER FLOW 39

Table D.5: Field specification for generator data matrix in MATPOWER case data (input fields 1–10 only).
The fifth column maps the field to an index, parameter, or variable used in the classical OPF formulation
given in Section 2.4.

Column Field Description Data Type Units Quantity in OPF

1 Bus Number Integer i (generator index)
2 Gen. Real Power Numeric MW PiG
3 Gen. Reactive Power Numeric MVAR QGi
4 Max. Reactive Power Numeric MVAR QG,max
i
5 Min. Reactive Power Numeric MVAR QG,min
i
6 Voltage Setpoint Numeric p.u.
7 Gen. MVA Basea Numeric MVA
8 Generator Statusb Binary
9 Max. Real Power Numeric MW PiG,max
10 Min. Real Power Numeric MW PiG,min
a Defaults to system power base S
Base
b 0 indicates generator out of service (remove from OPF formulation)