Anjos 2017

Foundations and Trends
R
in Electric Energy Systems
Vol. 1, No. 4 (2017) 220–310

c 2017 M. F. Anjos and A. J. Conejo
DOI: 10.1561/3100000014
Unit Commitment in Electric Energy Systems
Miguel F. Anjos
Department of Mathematics and Industrial Engineering
GERAD & Polytechnique Montreal
anjos@stanfordalumni.org
Antonio J. Conejo
Department of Integrated Systems Engineering
Department of Electrical and Computer Engineering
The Ohio State University
conejonavarro.1@osu.edu
Contents
1 Introduction to the Unit Commitment Problem 221

1.1 Outline of this Book . . . . . . . . . . . . . . . . . . . . . 223
2 Deterministic Unit Commitment 224

2.1 Binary Decision Variables . . . . . . . . . . . . . . . . . . 227
2.2 Startup Cost . . . . . . . . . . . . . . . . . . . . . . . . . 227
2.3 Ramping Constraints . . . . . . . . . . . . . . . . . . . . 228
2.4 Uptime and Downtime Constraints . . . . . . . . . . . . . 229
2.5 Generation Limits . . . . . . . . . . . . . . . . . . . . . . 230
2.6 Recapitulation of the Basic UC Formulation . . . . . . . . 231
2.7 Illustrative Example . . . . . . . . . . . . . . . . . . . . . 231
3 Network-Constrained Deterministic Unit Commitment 237

3.1 DC Power Flow . . . . . . . . . . . . . . . . . . . . . . . 238
3.2 Integration of Network Constraints in UC . . . . . . . . . 238
3.3 Recapitulation of the Network-Constrained Deterministic
UC Formulation . . . . . . . . . . . . . . . . . . . . . . . 239
4 Security-Constrained Deterministic Unit Commitment 246

4.1 Pre- and Post-Contingency Operating Conditions . . . . . 247
ii
iii
4.2 Recapitulation of the Security-Constrained Deterministic

UC Formulation . . . . . . . . . . . . . . . . . . . . . . . 248
5 AC Network-Constrained Unit Commitment 256

5.1 AC Power Flow . . . . . . . . . . . . . . . . . . . . . . . 256
5.2 Transformation of the AC Power Flow Equations to DC . . 258
5.3 Integration of AC Network Constraints in UC . . . . . . . 259
5.4 Recapitulation of the AC Network-Constrained UC Formu-
lation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
6 Stochastic Unit Commitment 268

6.1 Two-Stage Stochastic Optimization . . . . . . . . . . . . . 269
6.2 Optimization Objective . . . . . . . . . . . . . . . . . . . 270
6.3 First-Stage Constraints . . . . . . . . . . . . . . . . . . . 271
6.4 Second-Stage Constraints . . . . . . . . . . . . . . . . . . 271
6.5 Recapitulation of the Stochastic UC Formulation . . . . . 272
7 Robust Unit Commitment 277

7.1 Three-Level Adaptive Robust Formulation . . . . . . . . . 278
7.2 Uncertainty Set . . . . . . . . . . . . . . . . . . . . . . . 278
7.3 Optimization Objective . . . . . . . . . . . . . . . . . . . 279
7.4 First-Level Constraints . . . . . . . . . . . . . . . . . . . . 280
7.5 Second-Level Constraints . . . . . . . . . . . . . . . . . . 280
7.6 Third-Level Constraints . . . . . . . . . . . . . . . . . . . 281
7.7 Recapitulation of the Robust UC Formulation . . . . . . . 281
8 Computational Aspects 285

8.1 Computational Efficiencies . . . . . . . . . . . . . . . . . 285
8.2 Solving the Formulations . . . . . . . . . . . . . . . . . . 289
9 Conclusions and Directions for Future Research 295

iv
Appendices 296
A Derivation of the Dual of the Third-Level Problem in Section

8.2.1 297
B Notation 303
B.1 Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
B.2 Constants . . . . . . . . . . . . . . . . . . . . . . . . . . 304
B.3 Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
References 308
Abstract
The unit commitment problem is a fundamental problem in the elec-

tric power industry. The objective of unit commitment is to determine
an optimal schedule for each generating unit so that the demand for
electricity is met at minimum cost for the system as a whole. This
tutorial presents the most relevant mathematical optimization models
for the unit commitment problem. It is intended as a starting point for
learning about this important problem, and thus only the key technical
details are included. Likewise, we point out selected references instead
of providing a comprehensive literature review of the area.
M. F. Anjos and A. J. Conejo. Unit Commitment in Electric Energy Systems.

Foundations and Trends R
in Electric Energy Systems, vol. 1, no. 4, pp. 220–310,
2017.
DOI: 10.1561/3100000014.
1
Introduction to the Unit Commitment Problem
The unit commitment (UC) problem addresses a fundamental decision

in the operation of a power system, namely determining the schedule
of power production for each generating unit in the system so that
the demand for electricity is met at minimum cost. The schedule must
also ensure that each unit operates within its technical limits; these
typically include ramping constraints and minimum uptime/downtime
constraints. Units that are scheduled to produce electricity during a
given time period are said to be committed for that period.
Various jurisdictions solve UC on a daily basis. In particular, it is
the standard tool for clearing spot markets, and particularly day-ahead
markets in the USA. In North American jurisdictions without markets,
the system operators use UC to determine the day-ahead commitments
and dispatches.
The UC problem can be formulated as a mixed-integer nonlinear
optimization problem, and it is generally large-scale and nonconvex.
It is NP-hard in general, but its practical importance has motivated a
tremendous amount of research dedicated to techniques for computing
global optimal solutions. This is both because of the significance of
the operational costs and because in competitive market environments,
221
222 Introduction to the Unit Commitment Problem
the nonconvexity of the UC problem allows the existence of multiple

local optimal solutions that may lead to considerably different pricing
and market settlement outcomes. Indeed, a mixed-integer linear (or
nonlinear but convex) optimization model of the UC problem is among
the few techniques that can provide provably global optimal solutions
for the commitment decisions and corresponding financial settlements.
At the same time, the time available to solve the problem is a hard
constraint in practice. Hence, UC is an optimization problem that is
both important and challenging.
Various important aspects of UC can be integrated in a mixed-
integer nonlinear optimization approach, but the time required to solve
the UC models is a hard practical limitation that restricts the size and
scope of UC formulations. For this reason there is no single formu-
lation of UC; instead it is a matter of designing a formulation that
incorporates the important aspects of the problem for a given context
while ensuring that the resulting optimization problem can be solved
to optimality, or near-optimality, in a reasonable time.
With the increasing penetration of stochastic sources of electric-
ity in modern power systems, most notably wind and solar gener-
ation, techniques for handling uncertainty are acquiring greater im-
portance in UC modeling. We focus our presentation on two well-
known techniques for modeling uncertainty in mathematical optimiza-
tion, namely stochastic optimization and robust optimization. These
are by no means the only mathematical optimization techniques for
handling uncertainty, but we believe that they are the most relevant in
the context of UC because power system operators will always prefer
approaches that enforce constraints, rather than satisfying them with
some probability, which is the basis of most other approaches.
All the formulations that we present here are mathematical op-
timization problems. The Introduction to Optimization of the NEOS
Guide provides information about the different classes of mathematical
optimization problems and the software available to solve them. Most of
the state-of-the-art solvers, whether commercial or open source, can be
accessed for free on the NEOS Server [Czyzyk et al., 1998, Dolan, 2001,
1.1. Outline of this Book 223
Gropp and Moré, 1997]. All the computations made in the preparation
of this book were carried out on the NEOS Solver.
1.1 Outline of this Book
We introduce in the next six chapters a selection of formulations of

UC that integrate different aspects of the problem. We discuss the mo-
tivation for and the detailed structure of each formulation and then
recapitulate the mathematical model. Each chapter concludes with a
small example, accompanied by a description of how the results illus-
trate the features of the corresponding formulation.
We begin in Chapter 2 with a basic formulation of UC that focuses
on the modeling of the generating units and ensuring that generation
meets demand (with spinning reserves). The next step is to integrate
the impact of the power network; this can be done using power flow
equations in either linear (DC) form (Chapter 3) or alternating current
(AC) form (Chapter 5). The security of the system is a common con-
cern. In Chapter 4 we integrate constraints to ensure that the system
can cope with the failure of one of its major components.
The subsequent two chapters are concerned with modeling uncer-
tainty in the data for UC. We consider two modeling approaches: Chap-
ter 6 introduces a stochastic optimization approach that is based on
the use of scenarios, and Chapter 7 presents a robust optimization ap-
proach that focuses on the worst-case operating conditions.
While we briefly comment in the presentation of each example on
how the computational results were obtained, a detailed discussion of
the computational aspects of solving each formulation is given in Chap-
ter 8. Chapter 9 provides concluding remarks and discusses future
research.
2
Deterministic Unit Commitment
This chapter presents a basic formulation for UC, based on the presen-
tation in Ostrowski et al. [2012]. This formulation focuses on modeling
the fundamental aspects of UC, namely the operational limits of gener-
ating units, and ensuring that generation meets demand with spinning
reserves (explained below).
Our starting point is the following simple yet insightful formulation
of UC:
XX
min cj (pj (t)) + cU
j (t) (2.1)
pj (t)
t∈T j∈J
s.t.
X
pj (t) = D(t), ∀t ∈ T (2.2)
j∈J
X
pj (t) ≥ D(t) + R(t), ∀t ∈ T (2.3)
j∈J
pj (t), pj (t) ∈ Πj (t), ∀j ∈ J, ∀t ∈ T, (2.4)
where each value of t ∈ T is a time period in the planning horizon,

pj (t) denotes the quantity of electricity generated by unit j at time
t, and cUj (t) denotes the cost of starting up unit j in period t; we
224
225
specify this cost in Section 2.2. Typically, each t represents one hour
and T represents a 24-hour horizon (e.g., for day-ahead planning). The
hourly load forecast D(t) and spinning reserve requirements R(t) are
determined a priori by the system operator, but the maximum available
power at time t from unit j, denoted pj (t), is computed as part of the
solution and satisfies pj (t) ≤ pj (t).
The difference pj (t)−pj (t), i.e., the difference at time t between the
actual planned output pj (t) of unit j and the maximum amount pj (t)
it could provide, ensures that the optimal solution provides enough ca-
pacity for spinning reserves. Spinning reserves consist of the generating
capacity that is available from units that are already producing elec-
tricity to satisfy the demand but can still increase their power output
in case of need, for example if there is a disruption at another unit.
Let us look at each of the lines in this formulation. The objective
function (2.1) shows that UC aims to minimize the total cost of power
generation, with cj (pj (t)) denoting the cost of producing pj (t) units
of electricity using unit j at time t. It is common practice to model
cj (pj (t)) as a convex piecewise linear, strictly monotonically increasing
function:
cj (pj (t)) ≥ αjs pj (t) + βjs , s = 1, . . . , Cj ,
where Cj is the number of linear pieces in the cost function of unit j,

and αjs and βjs are fixed coefficients.
The concavity and strictly monotonic increase of cj (pj (t)) reflect
the increasing marginal cost of generation, and the piecewise linear-
ity facilitates the algorithmic solution of the model. However, with the
progress in optimization algorithms and software, it is now reasonable
to express cj (pj (t)) using a convex quadratic function because the re-
sulting model is straightforward to solve using state-of-the-art conic
optimization algorithms; see, e.g., Anjos and Lasserre [2011].
Constraint (2.2) requires that the total amount of generation at
time t meets the demand for that time period, and constraint (2.3)
requires the total maximum available production capacity available at
time t to be sufficient to meet the demand plus the required spinning
reserve requirement. Note that while the latter is an inequality ensuring
that sufficient reserve is present for the chosen commitment, the former
226 Deterministic Unit Commitment
is an equality requiring that the amount of energy generated is precisely

equal to the demand. One of the consequences of constraint (2.2) being
an equality is that the marginal value obtained from the dual optimal
solution can be positive or negative. These marginal values can be
interpreted as energy prices, and in certain circumstances such as cases
of extremely tight ramping limits, it makes sense to have negative prices
for energy [O’Neill et al., 2005]. We will not analyze such cases in any
detail.
If one were to optimize (2.1) subject to constraints (2.2) and (2.3)
only, then the outcome would be the absolute minimum generation
cost for the prescribed demand and spinning reserves. However, this
outcome would be highly unrealistic because it would not take into
account myriad aspects of UC that are critical in practice. These as-
pects include the operational constraints of the generating units, the
limitations imposed by the transmission network, and the security re-
quirements to maintain the stability of the system. Furthermore, the
uncertainties in the load forecast are particularly significant in the pres-
ence of stochastic and/or distributed generation, price-responsive de-
mand, and contingencies involving a transmission line or a generating
unit. These aspects are addressed in later chapters of this tutorial.
It is in the level of detail of the modeling that the various formula-
tions of UC differ. Roughly speaking, a more detailed formulation yields
a solution of higher quality but at a higher computational cost. This
is the fundamental tradeoff in formulating UC, and we will illustrate it
throughout this tutorial.
A more subtle issue is that different formulations may provide
equivalent levels of detail but differ significantly in terms of computa-
tional performance. We provide some comments about computational
issues throughout the tutorial, but we defer a detailed discussion of
computational matters to Chapter 8.
We begin by considering in this chapter the limitations imposed
by the physical characteristics and operating conditions of the gen-
erating units. These vary according to the nature of each unit (e.g.,
coal, gas, nuclear, hydro). In particular, units have to operate within a
given range of generation limits, their power output cannot change too
2.1. Binary Decision Variables 227
rapidly, and when their on/off status changes, it cannot change again
before a minimum amount of time has passed.
All these operational constraints are encapsulated in constraint
(2.4), and the set Πj (t) is formed of the constraints discussed in the
sections that follow.
2.1 Binary Decision Variables
We introduce three binary variables per generating unit to model the

following operational aspects:
• vj (t) equals 1 if unit j is on in time period t, and 0 if it is off;
• yj (t) equals 1 if unit j starts up at the beginning of time period
t, and 0 otherwise;
• zj (t) equals 1 if unit j shuts down at the beginning of time period
t, and 0 otherwise.
To ensure the logical coherence of the values assigned to these vari-

ables, the following constraints are included in the model:
vj (t − 1) − vj (t) + yj (t) − zj (t) = 0 ∀j ∈ J, ∀t ∈ T. (2.5)
They ensure that yj (t) and zj (t) take on appropriate values when a unit
starts up or shuts down. Note that constraints (2.5) require knowledge
of the values vj (0) that represent the on/off status of all generating
units in the time period just before the planning horizon starts (t = 0).
For example, suppose that vj (0) = 1 and vj (1) = 0, i.e., unit j is on at
time t = 0 but not at t = 1. Then constraints (2.5) will require that
yj (1) − zj (1) = −1, which holds only if yj (1) = 0 and zj (1) = 1, which
is consistent with the variable definitions.
2.2 Startup Cost
A startup cost cUj (t) is incurred when unit j starts up at the beginning
of time period t. This cost depends on how long the unit has been
inactive. For example, for thermal units this cost is maximum when
the boiler is completely cold.
While the startup cost function is nonlinear, it can be discretized

into hourly periods, giving a stepwise function:
τ
!
X
cU
j (t) ≥ Kjτ vj (t) − vj (t − n) , τ = 0, 1, . . . , τjcold ,
n=1
∀j ∈ J, ∀t ∈ T, (2.6)
where τjcold is the number of time periods unit j takes to cool down,
and Kjτ is the cost of starting up unit j after it has been shut down for
τ cold
τ time periods, where 0 = Kj0 ≤ Kj1 ≤ . . . ≤ Kj j due to Newton’s
law of cooling. Note that constraints (2.6) require knowledge of the
operational history of j in the most recent τjcold time periods before
t = 1, i.e., vj (1 − τjcold ), . . . , vj (0).
For simplicity throughout this tutorial, we assume that the startup
cost cU
j (t) is constant for each j ∈ J.
2.3 Ramping Constraints
The power output of a generating unit cannot change arbitrarily

quickly. For this reason, the production levels pj (t) are constrained
by ramping constraints as well as startup and shutdown rates.
Specifically, the ramping up of output from time period t to time
period t + 1 is limited as follows:
pj (t) − pj (t − 1) ≤ RjU vj (t − 1) + SjU yj (t), ∀j ∈ J, ∀t ∈ T, (2.7)
where RjU is the maximum ramp-up rate of unit j, and SjU is its maxi-
mum startup rate. Note that if the unit is on in time period t − 1, i.e.,
if vj (t − 1) = 1, then the increase in power output at time t cannot
be larger than RjU , while if the unit is turned on in the current time
period t, i.e., if yj (t) = 1, then it can output at most SjU during this pe-
riod. Moreover, constraints (2.7) require knowledge of the values vj (0)
and pj (0), respectively the on/off status and the generation level of all
generating units, just before the planning horizon starts (t = 0).
The constraints on ramping down from time period t to t + 1 are
deduced using similar arguments:
pj (t − 1) − pj (t) ≤ RjD vj (t) + SjD zj (t) ∀j ∈ J, ∀t ∈ T, (2.8)
2.4. Uptime and Downtime Constraints 229
where RjD is the maximum ramp-down rate of unit j, and SjD is its
maximum shutdown rate. We observe here that if unit j is shut down
in time period t+1, then its output at t cannot be more than SjD ; and if
it is not, its power output decrease at t cannot be larger than RjD . Note
also that constraints (2.8) require knowledge of pj (0), the generation
levels of all generating units at t = 0.
2.4 Uptime and Downtime Constraints
These constraints account for the fact that a unit cannot be turned on
or off arbitrarily. If unit j starts up in time period t then it has to run
for at least TjU time periods before it can be shut down. Similarly, if it
is shut down at t then it has to remain off for at least TjD periods.
We express the minimum uptime requirement as:
t
X
yj (k) ≤ vj (t) ∀t ∈ [ Lj + 1, . . . , |T | ] ∀j ∈ J, (2.9)
k=t−TjU +1, k≥1
where Lj = min{|T |, Uj } and Uj is the number of time periods that j

is required to be on at the start of the planning horizon. Constraints
(2.9) ensure that if unit j is off at t then it did not start up within the
previous TjU time periods. Note that these constraints require knowl-
edge of the operational history of j in the TjU − 1 time periods before
t = 1.
For example, suppose that TjU = 3. If unit j is off at t, then vj (t) =
0, and therefore yj (t) = 0, yj (t − 1) = 0, and yj (t − 2) = 0 must hold.
Conversely, if any one of yj (t), yj (t − 1), or yj (t − 2) equals 1, then
vj (t) = 1 must hold.
Constraints (2.9) also account for the fact that the unit may have
been started up not too long before the beginning of the planning
horizon.
Similarly, we express the minimum downtime requirement as:
t
X
vj (t) + zj (k) ≤ 1 ∀t ∈ [ Fj + 1, . . . , |T | ] ∀j ∈ J (2.10)
k=t−TjD +1, k≥1
where Fj = min{|T |, Dj } and Dj is the number of time periods that

unit j is required to remain off at the start of the planning horizon.
These constraints ensure that if unit j is on at t then it did not shut
down within the previous TjD time periods. Note that these constraints
require knowledge of the operational history of j in the TjD − 1 time
periods before t = 1.
For example, suppose that TjD = 4. If unit j is on at time t, then
vj (t) = 1, and therefore zj (t) = 0, zj (t − 1) = 0, zj (t − 2) = 0, and
zj (t−3) = 0 will hold. Conversely, if any one of zj (t), zj (t−1), zj (t−2),
or zj (t − 3) equals 1, then vj (t) = 0 must hold.
Again the fact that the unit may have been shut down not too long
before the beginning of the planning horizon is taken into account.
2.5 Generation Limits
If a unit is on in time period t then its power output must satisfy

appropriate lower and upper limits P j and P j . A straightforward way
to express this is as follows:
P j vj (t) ≤ pj (t) ≤ pj (t) ≤ P j vj (t) ∀j ∈ J, ∀t ∈ T (2.11)
Note that if unit j is off at t, then (2.11) forces pj (t) and pj (t) to be
zero.
It is also necessary to account for the limits on pj (t), the maximum
available power at time t from unit j. This quantity is used to meet the
reserve requirements, as shown by constraint (2.3), and its magnitude
is constrained by the ramp-up, startup, and shutdown limits.
We have already noted that if vj (t) = 0 then (2.11) sets pj (t) to zero.
If vj (t) = 1, then the limits on pj (t) are enforced using the following
two sets of constraints:
pj (t) ≤ pj (t − 1) + RjU vj (t − 1) + SjU yj (t), ∀j ∈ J, ∀t ∈ T (2.12)
With vj (t) = 1, constraint (2.12) limits pj (t) to be no greater than the
power output in the previous time period plus the maximum ramp-up
rate. The exception is if unit j was started at the beginning of period
t, in which case the constraint limits pj (t) to the startup rate.
pj (t) ≤ P j vj (t) − zj (t + 1) + zj (t + 1)SjD , ∀j ∈ J, ∀t ∈ T

(2.13)
2.6. Recapitulation of the Basic UC Formulation 231
Also with vj (t) = 1, constraint (2.13) ensures that if unit j is shut down
at the beginning of time period t + 1, then pj (t) is bounded above
by the maximum shutdown rate. Otherwise, the constraint becomes
redundant, imposing the same upper bound as in (2.11).
2.6 Recapitulation of the Basic UC Formulation

XX
j yj (t)
Ξ
t∈T j∈J
s.t.
X
pj (t) = D(t), ∀t ∈ T
j∈J
X
pj (t) ≥ D(t) + R(t), ∀t ∈ T
j∈J
cj (pj (t)) ≥ αjs pj (t) + βjs , s = 1, . . . , Cj , ∀j ∈ J

vj (t − 1) − vj (t) + yj (t) − zj (t) = 0, ∀j ∈ J, ∀t ∈ T
pj (t) − pj (t − 1) ≤ RjU vj (t − 1) + SjU yj (t), ∀j ∈ J, ∀t ∈ T
pj (t − 1) − pj (t) ≤ RjD vj (t) + SjD zj (t), ∀j ∈ J, ∀t ∈ T
t
X
yj (k) ≤ vj (t), ∀t ∈ [Lj + 1, . . . , |T |], ∀j ∈ J
k=t−TjU +1, k≥1
t
X
vj (t) + zj (k) ≤ 1, ∀t ∈ [Fj + 1, . . . , |T |], ∀j ∈ J
k=t−TjD +1, k≥1
P j vj (t) ≤ pj (t) ≤ pj (t) ≤ P j vj (t), ∀j ∈ J, ∀t ∈ T

pj (t) ≤ pj (t − 1) + RjU vj (t − 1) + SjU yj (t), ∀j ∈ J, ∀t ∈ T
1)SjD ,

pj (t) ≤ P j vj (t) − zj (t + 1) + zj (t + ∀j ∈ J, ∀t ∈ T
where the optimization variables in set Ξ are pj (t), pj (t), vj (t), yj (t),
and zj (t), ∀j ∈ J, ∀t ∈ T .
2.7 Illustrative Example
We consider a small example with three generators. For simplicity we

assume that the costs cj and cU
j are constant for every unit j.
The parameters for the generators are as follows:
Unit (j) cU
j cj Pj Pj SjU SjD RjU RjD TjU TjD
1 800 5 80 300 100 80 50 30 3 2
2 500 15 50 200 70 50 60 40 2 2
3 250 30 30 100 40 30 70 50 1 2
Unit 1 is a generator that can produce a large quantity of power at low
cost, but it incurs a high startup cost and ramps up relatively slowly.
Unit 2 is a generator that has lower production capacity than unit 1
and a slightly higher cost but costs less to startup. Unit 3 has the lowest
production capacity and is expensive, but it has the lowest startup cost
and ramps up more quickly. The intuition here is that the UC problem
will use unit 1 first, then unit 2, and finally unit 3 if necessary.
Further, suppose that we have the following operating conditions
at t = 0, i.e., just before our planning horizon starts:
Unit (j) vj pj Uj Dj
1 1 120 2 0
2 0 0 0 0
3 0 0 0 0
Thus, at t = 0, unit 1 is on and producing 120 MW, and units 2 and 3
are off. Furthermore, unit 1 is required to stay on for two time periods
at the start of the planning horizon.
The formulation in Section 2.6 is a mixed-integer linear optimiza-
tion problem. We solved it using BARON [Tawarmalani and Sahinidis,
2005], one of the optimization solvers on NEOS that can handle this
type of problem.
2.7.1 Initial Demand

Let us assume that we are planning for the next six time periods and
that the net demands and reserve requirements are:
Time period t 1 2 3 4 5 6
Demand D(t) 240 250 200 170 230 190
Reserve R(t) 10 10 10 10 10 10
2.7. Illustrative Example 233
For simplicity, we set the reserve to 10 units of demand for each time
period t.
Solving the formulation in Section 2.6, we obtain the following re-
sults:
p1 (t) 170 200 200 170 200 190
p2 (t) 70 50 0 0 0 0
p3 (t) 0 0 0 0 30 0
Dual of (2.2) 15 5 5 5 5 5
p1 (t) − p1 (t) 10 10 10 10 10 10
p2 (t) − p2 (t) 0 0 0 0 0 0
p3 (t) − p3 (t) 0 0 0 0 0 0
pj (t) − pj (t)
P
10 0 0 0 0 0
j
Excess over R(t) 0 0 0 0 0 0
Dual of (2.3) 0 0 0 0 0 0
The total cost is 9100. We observe that the marginal price of energy is
15 for the first period and 5 for every subsequent period. We note also
that the reserve requirements are met, and that the marginal price of
the reserve is zero for every period.
The on/off status of each of the generators during the planning
horizon is as follows:
Unit 1 on on on on on on
Unit 2 on on off off off off
Unit 3 off off off off on off
We see that unit 1 is committed throughout, unit 2 is turned on for the

first two periods, and unit 3 is turned on at t = 5 only. Unit 2 is turned
on at t = 1 because the level of demand in the first time period means
that unit 1, which is operating at a level of 120 at t = 0, cannot ramp
up fast enough to satisfy both demand and reserve. The fact that the
demand at t = 1 is high compared to the production levels at t = 0 is
reflected in the higher marginal price of energy at t = 1.
Once on, unit 2 must run for two periods before it can be turned
off (T2U = 2), even though unit 1 would be able to satisfy both demand
and reserve at t = 2.
There is a peak in demand at t = 5, but it lasts only one period
and unit 3 is now available. Therefore, the optimal solution is to turn
on unit 3 for only that period and to increase the output of unit 1 to
make up the difference.
If we check the constraints, we see that the ramp-up constraint
(2.7) for units 1 and 2 is binding at t = 1, and that the lower bound
constraints on the production, i.e., the leftmost inequality constraints
in (2.11), are binding for unit 2 from t = 2 onward, and for unit 3 in
every time period.
2.7.2 Impact of Operational Constraints

Suppose now that the peak at t = 5 is even higher, equal to 260, while
everything else remains unchanged. The resulting generation schedule
is:
p1 (t) 170 180 150 120 170 140
p2 (t) 70 70 50 50 90 50
p3 (t) 0 0 0 0 0 0
Dual of (2.2) 15 15 -5 -5 15 5
p1 (t) − p1 (t) 10 10 10 10 0 10
p2 (t) − p2 (t) 0 0 0 0 10 0
p3 (t) − p3 (t) 0 0 0 0 0 0
pj (t) − pj (t)
P
10 10 10 10 10 10
j
Dual of (2.3) 0 0 0 0 0 0
The total cost is 10850. The on/off status of each of the generators
during the planning horizon is as follows:
Unit 3 off off off off off off
As would be expected, the higher load leads to a significantly higher

cost. A more significant change is that unit 2 operates in every time
period, and unit 3 is not used. The combination of the ramping limits
for unit 1 and the minimum up and down times of unit 2 require both
units to be on for all six periods.
The binding constraints are the ramp-up constraint on unit 1 at
t = 1 and t = 5, the ramp-down constraint on unit 1 at t = 3, the
lower bound constraint on the production of unit 2 at t = 3, t = 4, and
t = 6, and the lower bound constraint on the production of unit 3 in
every period.
An important observation here is the occurrence of negative prices
for t = 3 and t = 4. These arise because of the inability of the system,
under the given operational constraints, to respond quickly enough
to the fluctuation of the load, as evidenced by the various ramping
constraints that are binding. We can confirm this by increasing R1U
from 50 to 60 and R1D from 30 to 50 and re-running the optimization.
The resulting production schedule is:
p1 (t) 180 200 200 170 190 140
p2 (t) 60 50 0 0 70 50
p3 (t) 0 0 0 0 0 0
Dual of (2.2) 15 5 5 5 15 5
The total cost is 9850, and all the marginal energy prices are positive.
horizon is as follows:
Unit 2 on on off off on on
We see that with its increased ramping capabilities, unit 1 (which has
the cheapest cost per unit of energy) can serve almost all the load. Unit
2 contributes for just four time periods, at lower levels than before.
3
Network-Constrained Deterministic Unit
Commitment
The basic UC formulation in Section 2.6 is concerned only with allocat-

ing sufficient power generation to satisfy the demand plus the spinning
reserves. It does not account for the transmission system that carries
the electricity from the generating units to the loads. However, there
are limits on the power flow across the lines, and accounting for these
ensures that the allocated power can suitably flow within the system.
The impact of transmission is particularly important in markets where
accurate modeling of the congestion of the power lines and the con-
sequent differences in the locational marginal prices of electricity are
essential for proper market operation.
Power flows are typically modeled in either AC or DC form, and
both can be used in the UC context. The DC power flow equations are a
linearized version of the AC equations that are nonlinear (and noncon-
vex). The DC approach is more efficient computationally and provides
a reasonably good estimate of the system behavior under normal con-
ditions, but it can lead to misleading results in extreme conditions. In
this chapter we discuss the DC approach (based on the results in Motto
et al. [2002]). The AC approach will be presented in Chapter 5.
237
238 Network-Constrained Deterministic Unit Commitment
3.1 DC Power Flow
Let N be the set of nodes in the power system network, and let nm
be the line connecting nodes n and m. The DC approximation of the
power flow on a line postulates that it is proportional to the differ-
ence in voltage angles between the nodes at the endpoints of the line.
Specifically, the power flow across the line nm is given by
Bnm (θn (t) − θm (t)) ,
where Bnm is the negative of the series susceptance of the line (a pos-
itive constant), and θn (t) is the voltage angle at node n. In essence,
Bnm is a proportionality constant linking the differences in the voltage
angle (electrical “height”) and the power flow (see Section 5.1 for more
details on the susceptance and related quantities). Thus, the net power
flow (generation minus demand) at node n is given by
X
Bnm (θn (t) − θm (t)) ,
m∈ΛL
n
where ΛLn is the set of nodes directly connected to node n.
3.2 Integration of Network Constraints in UC
To incorporate the network effects, we change the constraints (2.2)

to enforce power balance at each node independently. Thus, (2.2) is
replaced by
X X X
pj (t) − Di (t) = Bnm (θn (t) − θm (t)) ∀n ∈ N, ∀t ∈ T,
j∈ΛG
n i∈ΛD
n m∈ΛL
n
(3.1)
where Di (t) is the load of demand i in period t, ΛG
is the set of gen-
n
D
erating units located at node n, and Λn is the set of demands located
at node n.
It is also necessary to enforce transmission capacity limits for each
transmission line:
−P nm ≤ Bnm (θn (t) − θm (t)) ≤ P nm , ∀n ∈ N, ∀m ∈ ΛLn , ∀t ∈ T,

(3.2)
3.2. Integration of Network Constraints in UC 239
where P nm is the transmission capacity of line nm.

In addition, we set a reference angle:
θn̂ (t) = 0 ∀t ∈ T, (3.3)
where n̂ is the node of the reference angle.

Finally, appropriate reserve levels must be enforced in each reserve
area, so constraints (2.3) are replaced by the following constraints:
X X
pj (t) ≥ Di (t) + Rr (t) ∀r ∈ R, ∀t ∈ T, (3.4)
j∈ΩG
r i∈ΩD
r
where R is the set of reserve areas, Rr (t) is the reserve required in

reserve area r in time period t, ΩG
r is the set of generating units in area
D
r, and Ωr is the set of demands in area r.
The required reserve in area r, Rr (t), should take into account the
interconnections of area r with neighboring areas, since these areas may
also provide reserve to area r (and vice versa).
We note that constraint (3.4) is a proxy for the need to enforce
contingency constraints relating to possible outages of each generator.
Provided that the reserve areas are chosen so that they reflect the
limitations arising from the transmission system, this proxy provides
a better representation than the system-wide reserve constraint (2.3).
Nevertheless, it does not completely represent the requirement for the
delivery of reserves under every possible generation outage. A more
accurate representation of this requirement could be achieved using
the security-constrained UC presented in Chapter 4.
3.3 Recapitulation of the Network-Constrained Determinis-

tic UC Formulation
We use boxes to highlight the constraints and variables that are specific
to the integration of the network. The remainder of the model is from
the basic UC formulation in Section 2.6.
XX
j yj (t)
Ξ
t∈T j∈J
s.t.
pj (t) − Bnm (θn (t) − θm (t)),

P P P
j∈ΛG
n i∈ΛD
n
Di (t) = m∈ΛL
n
∀n ∈ N, ∀t ∈ T
−P nm ≤ Bnm (θn (t) − θm (t)) ≤ P nm ,
∀n ∈ N, ∀m ∈ ΛLn , ∀t ∈ T
θn̂ (t) = 0, ∀t ∈ T
pj (t) ≥ ∀r ∈ R, ∀t ∈ T
P P
j∈ΩG
r i∈ΩD
r
Di (t) + Rr (t),
t
X
yj (k) ≤ vj (t) ∀t ∈ [Lj + 1, . . . , |T |], ∀j ∈ J
k=t−TjU +1, k≥1
t
X
vj (t) + zj (k) ≤ 1 ∀t ∈ [Fj + 1, . . . , |T |], ∀j ∈ J
k=t−TjD +1, k≥1

1) + zj (t + 1)SjD ,

pj (t) ≤ P j vj (t) − zj (t + ∀j ∈ J, ∀t ∈ T
zj (t), and θn (t) , ∀n ∈ N , ∀j ∈ J, ∀t ∈ T .
Let us consider a small power network with six buses. The three gen-
erators are the same as in Section 2.7, and we locate them in nodes
1, 2, and 3 respectively. Nodes 4, 5, and 6 are demand nodes. Thus,
the primary flow of power is from nodes 1, 2, and 3 to nodes 4, 5, and
6. For simplicity we further assume that there is a single reserve area,

corresponding to the whole network.
The formulation in Section 3.3 is also a mixed-integer linear opti-
mization problem, and we solved it using BARON.
3.4.1 Initial Case
First, we consider the same net demand and reserve requirement as in

Section 2.7.1. We divide the demand among the three demand nodes
as follows:
D4 (t) 100 100 80 140 100 80
D5 (t) 90 100 80 30 90 60
D6 (t) 50 50 40 0 40 50
Total demand 240 250 200 170 230 190
Reserve R(t) 10 10 10 10 10 10
Further, suppose that the network lines have the following trans-
mission capacities P nm :
1 2 3 4 5 6
1 - - - 100 100 100
2 - - - 100 100 100
3 - - - - 100 100
4 100 100 - - - -
5 100 100 100 - - -
6 100 100 100 - - -
Here a dash denotes the absence of a line between the two nodes, and
Bnm equals 0.8 for every line.
Solving the formulation in Section 3.3, we obtain the following gen-

eration schedule:
p1 (t) 170 200 200 170 200 190
p2 (t) 70 50 0 0 0 0
p3 (t) 0 0 0 0 30 0
Dual at node 1 15 5 5 5 5 5
Dual at node 2 15 5 5 5 5 5
Dual at node 3 15 5 5 5 5 5
Dual at node 4 15 5 5 5 5 5
Dual at node 5 15 5 5 5 5 5
Dual at node 6 15 5 5 5 5 5
p1 (t) − p1 (t) 10 10 10 10 10 10
p2 (t) − p2 (t) 0 0 0 0 0 0
p3 (t) − p3 (t) 0 0 0 0 0 0
pj (t) − pj (t)
P
10 10 10 10 10 10
j
Dual of (3.4) 0 0 0 0 0 0
The total cost is 9100. Note that this is the same total cost as in Section
2.7.1.
We observe that because power balance is now enforced at each
node of the network (constraints (3.1)), each of these constraints has
its own dual variable, and hence we obtain nodal energy prices. In this
example, they are equal to 15 at every node for t = 1 and equal to 5
at every node for the subsequent time periods.
We note also that the reserve requirements are met, and that the
marginal price of the reserve is zero in every period. Moreover, the
on/off status of each of the generators during the planning horizon is
also the same as in Section 2.7.1:
The binding constraints are the ramp-up constraint on unit 1 at

t = 1, and the lower bound constraints on the production of unit 2 at
every period except t = 1 and the production of unit 3 at every time
period.
In addition, this network-constrained formulation gives us informa-
tion about the flows of power in the system. For example, at t = 1 the
power flows are as follows:
1 2 3 4 5 6
1 - - - -66.67 -58.33 -45
2 - - - -33.33 -25 -11.67
3 - - - - -6.67 6.667
4 66.67 33.33 - - - -
5 58.33 25 6.67 - - -
6 45 11.67 -6.67 - - -
Observe that we have, as desired, a total flow of 100 into node 4. Two-
thirds of this power flows from 1 to 4, and one-third from 2 to 4.
We also have an inflow of 90 into node 5, with 58 1/3 coming from
node 1, 25 from node 2, and 6 2/3 from node 3. Note that the power
coming from node 3 is not generated there but flows into 3 from node 6.
Indeed, at node 6 we have as desired a net inflow of 50, but the actual
inflows are 45 from node 1 and 11 2/3 from node 2; these two inflows
add up to 56 2/3, of which 6 2/3 continues on to node 5.
3.4.2 Impact of Reduced Transmission Capacity

Changes in the transmission capacity can significantly affect the UC
solution. To illustrate this, consider again the case of Section 3.4.1 but
suppose that we reduce P 16 (and also P 61 ) from 100 to 40; the optimal
solution to the formulation in Section 3.3 becomes the following:
p1 (t) 160 157.5 147.5 120 162.5 137.5
p2 (t) 50 92.5 52.5 50 67.5 52.5
p3 (t) 30 0 0 0 0 0
Dual at node 1 5 5 5 5 5 5
Dual at node 2 5 25 5 5 15 15
Dual at node 3 5 30 5 5 17.5 17.5
Dual at node 4 5 15 5 5 10 10
Dual at node 5 5 20 5 5 12.5 12.5
Dual at node 6 5 40 5 5 22.5 22.5
p1 (t) − p1 (t) 10 10 10 10 7.5 10
p2 (t) − p2 (t) 0 0 0 0 2.5 0
p3 (t) − p3 (t) 0 0 0 0 0 0
pj (t) − pj (t)
P
10 10 10 10 10 10
j
Dual of (3.4) 0 0 0 0 0 0
The total cost is 11550, an increase of 27%.

One observation is that the structure of the UC optimal solution
changes: unit 2 is turned on throughout, and unit 3 is on in the first
time period:
Unit 3 on off off off off off
The ramp-down constraint on unit 2 at t = 3 is binding, and so are

the lower bound constraints on the production of unit 2 at t = 1 and
t = 4 and on the production of unit 3 in every time period.
However, the key observation here is that the nodal energy prices are
no longer equal throughout the network in each time period. Specifically,
for t = 2, t = 5, and t = 6, the prices between nodes vary significantly.
This is due to the reduced capacity on the line between nodes 1 and
6; the capacity constraint on this line becomes binding in these time
periods.
At t = 2 the power flows are, with the binding flows shown in bold:
1 2 3 4 5 6
1 - - - -60.83 -56.67 -40.00
2 - - - -39.17 -35.00 -18.33
3 - - - - -8.33 8.33
4 60.83 39.17 - - - -
5 56.67 35.00 8.33 - - -
6 40.00 18.33 -8.33 - - -
At t = 5 they are:
1 2 3 4 5 6
1 - - - -65.83 -56.67 -40.00
2 - - - -34.17 -25.00 -8.33
3 - - - - -8.33 8.33
4 65.83 34.17 - - - -
5 56.67 25.00 8.33 - - -
6 40.00 8.33 -8.33 - - -
At t = 6:
1 2 3 4 5 6
1 - - - -54.17 -43.33 -40.00
2 - - - -25.83 -15.00 -11.67
3 - - - - -1.67 1.67
4 54.17 25.83 - - - -
5 43.33 15.00 1.67 - - -
6 40.00 11.67 -1.67 - - -
As a consequence of the congestion on the line between nodes 1 and

6, most prices are noticeably higher throughout the network during
these three time periods. The prices are highest at node 6, reaching 40
at t = 2, and 22.5 at t = 5 and t = 6. We conclude by emphasizing
that, in general, a single binding constraint may suffice to cause the
prices to be different at every node.
4
Security-Constrained Deterministic Unit
Commitment
The UC problem as considered so far does not take into account the
need to maintain the security of the power system when planning the
day-ahead hourly generation schedule. This is particularly important
in jurisdictions where the security issues may be neglected because of
market pressures and related financial concerns.
The security-constrained UC (SCUC) is a formulation of UC that
integrates key aspects of system security while scheduling generation.
The idea is that the system should be able to continue to meet all the
load demands in the event of failure in one of its major components;
such events are referred to as contingencies. Our concern here is with
contingencies whose impact cannot be handled by the spinning reserves.
For simplicity we consider contingencies pertaining solely to trans-
mission lines and identify each of them by the susceptances of the
transmission lines under the contingency, denoted Bnm c , where the su-
perscript c indicates contingency c. Contingencies pertaining to gener-

ating units can be incorporated similarly.
The formulation in this chapter, based on the work in Bouffard et al.
[2005], embodies a corrective view, i.e., we ensure that the system can
transition from the contingency state (where some transmission line
246
4.1. Pre- and Post-Contingency Operating Conditions 247
limits are violated) to a safe post-contingency state (where no trans-

mission line limit is violated). In other words, the contingency may
occur at any time within the planning horizon, and there is sufficient
generation capacity available to support the transition to a safe op-
erating state. We do not impose a (comparatively more expensive)
preventive view that guarantees that all post-contingency states are
safe. This is the most common criterion in practice.
4.1 Pre- and Post-Contingency Operating Conditions
We expand the formulation in Section 3.3 to integrate the security

constraints. Let C denote the set of possible contingencies. For each
contingency c ∈ C, we introduce variables pcj (t) to denote the post-
contingency c power produced at time t of unit j, and θnc (t) to denote
the post-contingency c voltage angle of node n. Using these variables,
it is straightforward to express the post-contingency version of the net-
work constraints (3.1), (3.2), and (3.3):
X X X
pcj (t) − Di (t) = c
Bnm (θnc (t) − θm
c
(t)) ,
j∈ΛG
n i∈ΛD
n m∈ΛL
n
∀n ∈ N, ∀t ∈ T, ∀c ∈ C (4.1)
c
− P nm ≤ Bnm (θnc (t) − c
θm (t)) ≤ P nm ,
∀n ∈ N, ∀m ∈ ΛLn , ∀t ∈ T, ∀c ∈ C (4.2)
θn̂c (t) = 0, ∀t ∈ T, ∀c ∈ C (4.3)
In a similar manner, the post-contingency versions of the ramping

constraints (2.7) and (2.8) are as follows:
pcj (t) − pcj (t − 1) ≤ RjU vj (t − 1) + SjU yj (t), ∀j ∈ J, ∀t ∈ T, ∀c ∈ C

(4.4)
pcj (t − 1) − pcj (t) ≤ RjD vj (t) + SjD zj (t), ∀j ∈ J, ∀t ∈ T, ∀c ∈ C (4.5)
The generation limits also apply to the post-contingency power out-

puts:
P j vj (t) ≤ pcj (t) ≤ P j vj (t), ∀j ∈ J, ∀t ∈ T, ∀c ∈ C (4.6)

248 Security-Constrained Deterministic Unit Commitment
Finally, it is necessary to link the pre-contingency and post-

contingency conditions by bounding their difference:
pcj (t) − pj (t) ≤ XjU , ∀j ∈ J, ∀t ∈ T, ∀c ∈ C (4.7)

pj (t) − pcj (t) ≤ XjD , ∀j ∈ J, ∀t ∈ T, ∀c ∈ C (4.8)
where XjU (respectively XjD ) is the maximum increase (resp. decrease)

in the power that can be provided by unit j to transition the system
to a safe post-contingency state.
4.2 Recapitulation of the Security-Constrained Deterministic

UC Formulation
We use boxes to highlight the constraints and variables that are specific
to the integration of the contingencies. The remainder of the formula-
tion is from the network-constrained UC model in Section 3.3.
For the sake of simplicity, we do not represent the minimum
up/down time constraints or the regional reserve constraints.
XX
j yj (t)
Ξ
t∈T j∈J
s.t.
X X X
pj (t) − Di (t) = Bnm (θn (t) − θm (t)) , ∀n ∈ N, ∀t ∈ T
j∈ΛG
n i∈ΛD
n m∈ΛL
n
− P nm ≤ Bnm (θn (t) − θm (t)) ≤ P nm , ∀n ∈ N, ∀m ∈ ΛLn , ∀t ∈ T

θn̂ (t) = 0, ∀t ∈ T
P j vj (t) ≤ pj (t) ≤ P j vj (t), ∀j ∈ J, ∀t ∈ T
pcj (t) − c (θ c (t) − θ c (t)) ,
P P P
j∈ΛG
n i∈ΛD
n
Di (t) = m∈ΛL
n
Bnm n m
∀n ∈ N, ∀t ∈ T, ∀c ∈ C
c (θ c (t) − θ c (t)) ≤ P
−P nm ≤ Bnm n m nm ,
∀n ∈ N, ∀m ∈ ΛLn , ∀t ∈ T, ∀c ∈ C
θn̂c (t) = 0, ∀t ∈ T , ∀c ∈ C
pcj (t) − pcj (t − 1) ≤ RjU vj (t − 1) + SjU yj (t), ∀j ∈ J, ∀t ∈ T, ∀c ∈ C
pcj (t − 1) − pcj (t) ≤ RjD vj (t) + SjD zj (t), ∀j ∈ J, ∀t ∈ T, ∀c ∈ C
P j vj (t) ≤ pcj (t) ≤ P j vj (t), ∀j ∈ J, ∀t ∈ T, ∀c ∈ C
pcj (t) − pj (t) ≤ XjU , ∀j ∈ J, ∀t ∈ T, ∀c ∈ C
pj (t) − pcj (t) ≤ XjD , ∀j ∈ J, ∀t ∈ T, ∀c ∈ C
where the optimization variables in set Ξ are pj (t), vj (t), yj (t), zj (t),
θn (t), pcj (t) , and θnc (t) , ∀n ∈ N , ∀j ∈ J, ∀t ∈ T, ∀c ∈ C.
Let us consider the same power network with six buses as in Section
3.4, with generators in nodes 1, 2, and 3, and demand at nodes 4, 5,
and 6. For simplicity, as in the formulation of Section 4.2, we do not

account for minimum up/down times or reserve requirements.
Like the previous formulations, the formulation in Section 4.2 is
a mixed-integer linear optimization problem, and we solved it using
BARON.
4.3.1 Initial Case (No Contingency)
We consider the following nodal net demands over the six time periods:
D4 (t) 100 100 100 100 100 100
D5 (t) 0 40 40 40 40 0
D6 (t) 0 0 20 0 30 0
Total demand 100 140 160 140 170 100
and the following network lines and transmission capacities P nm :
1 2 3 4 5 6
1 - - - 100 - 100
2 - - - 100 100 100
3 - - - - 100 100
4 100 100 - - - -
5 - 100 100 - - -
6 100 100 100 - - -
A dash indicates the absence of a line between the two nodes, and Bnm
equals 0.8 for every line.

eration schedule with a total cost of 4850:
p1 (t) 100 140 160 140 170 100
p2 (t) 0 0 0 0 0 0
p3 (t) 0 0 0 0 0 0
Dual at node 1 5 5 5 5 5 5
Dual at node 2 5 5 5 5 5 5
Dual at node 3 5 5 5 5 5 5
Dual at node 4 5 5 5 5 5 5
Dual at node 5 5 5 5 5 5 5
Dual at node 6 5 5 5 5 5 5
We observe that all the power is provided by the generator at node

1, which is the lowest-cost generator and is already running at t = 0
with p1 (0) = 120, as per the initial conditions in Section 2.7. We also
note that all nodal energy prices are equal to 5 at every node for every
time period.
The on/off status of the generators is:
and the highest power flows occur at t = 5:
1 2 3 4 5 6
1 - - - -97.33 - -72.67
2 - - - -2.67 -19.33 22.00
3 - - - - -20.67 20.67
4 97.33 2.67 - - - -
5 - 19.33 20.67 - - -
6 72.67 -22.00 -20.67 - - -
Observe that none of the transmission capacity constraints are binding.
4.3.2 Impact of Single Contingency
Let us now consider the possibility of a single contingency c correspond-

ing to the loss of the line between nodes 1 and 6.
We set XjU and XjD to half of the corresponding one-hour ramping
rate of unit j. Specifically, in accordance with the data in Section 2.7,
we have:
Unit (j) RjU RjD XjU XjD

1 50 30 25 15
2 60 40 30 20
3 70 50 35 25
Solving the formulation in Section 4.2 with the highlighted con-

straints corresponding to this contingency and the power change limits,
we obtain the following generation schedule with a (much higher) total
cost of 7550:
p1 (t) 100 90 110 90 90 100
p2 (t) 0 50 50 50 50 0
p3 (t) 0 0 0 0 30 0
Dual at node 1 5 5 5 5 5 5
Dual at node 2 5 5 5 5 5 5
Dual at node 3 5 5 5 5 5 5
Dual at node 4 5 5 5 5 5 5
Dual at node 5 5 5 5 5 5 5
Dual at node 6 5 5 5 5 5 5
Unit 2 off on on on on off
If the line between 1 and 6 stops operating, there is no longer a

feasible solution with all the generation taking place at node 1. This
is because if the contingency occurs, then there is no longer enough
transmission capacity out of node 1 to satisfy all the demand. It is
necessary to reduce the output of the unit at node 1 and turn on one
or both of the more expensive generators, resulting in a higher total
generation cost.
At t = 5, the period of highest demand, the power flows are:
1 2 3 4 5 6
1 - - - -64.00 - -26.00
2 - - - -36.00 -16.00 2.00
3 - - - - -24.00 -6.00
4 64.00 36.00 - - - -
5 - 16.00 24.00 - - -
6 26.00 -2.00 6.00 - - -
Comparing these flows with those in Section 3.4.1, we observe in partic-
ular the reduced flows on the two lines out of node 1, and the increased
flow on the line between nodes 2 and 4.
In the event of the contingency, the generation schedule would be:
pc1 (t) 100 90 95 90 90 100
pc2 (t) 0 50 65 50 50 0
pc3 (t) 0 0 0 0 30 0
where the power levels in bold are those where pcj (t) differs from pj (t),
and at t = 5, the power flows would be:
1 2 3 4 5 6
1 - - - -90.00 - -
2 - - - -10.00 -22.50 -17.50
3 - - - - -17.50 -12.50
4 90.00 10.00 - - - -
5 - 22.50 17.50 - - -
6 - 17.50 12.50 - - -
We observe that the only change in generation is at t = 3, requiring

unit 1 to decrease its output by 15 and unit 2 to increase it by the same
amount. Note that the changes are within the limits determined by in-
equalities (4.7) and (4.8). The power flows at t = 5 change significantly;
all the flows are now from the nodes with generation to those with de-
mand. Nevertheless, none of the transmission capacity constraints are
binding. In fact, none of the contingency-related inequality constraints
are binding.
4.3.3 Impact of Reduced Maximum Power Change

Suppose that we consider the same contingency as in Section 4.3.2 but
reduce X1D , the maximum decrease in the power output of unit 1, from
15 to 5. The optimal generation schedule is now:
p1 (t) 100 90 105 90 90 100
p2 (t) 0 50 55 50 50 0
p3 (t) 0 0 0 0 30 0
Dual at node 1 5 5 15 5 5 5
Dual at node 2 5 5 15 5 5 5
Dual at node 3 5 5 15 5 5 5
Dual at node 4 5 5 15 5 5 5
Dual at node 5 5 5 15 5 5 5
Dual at node 6 5 5 15 5 5 5
with a higher cost of 7600. The power levels in bold are those where
pj (t) differs from pj (t) in Section 4.3.2. The on/off status of the gener-
ators is unchanged from that of Section 4.3.2.
In the event of the contingency, the generation schedule would be:
pc1 (t) 100 90 100 90 90 100
pc2 (t) 0 50 60 50 50 0
pc3 (t) 0 0 0 0 30 0
where the power levels in bold are those where pj (t) differs from pcj (t).
As was the case in Section 4.3.2, we see that the changes in gen-
eration take place at t = 3 for generators 1 and 2. However, there are
additional impacts that are different from those in Section 4.3.2. One
of them is in the nodal energy prices. First, note that the nodal prices
at t = 3 have increased from 5 to 15 in the absence of the contingency
event.
Second, the nodal prices in the event of the contingency are no
longer all equal to zero: the energy price at node 1 for t = 3 is now
-10, indicating that in the event of the contingency, a lower level of
generation for unit 1 at t = 3 would be desirable. Indeed, we saw from
the results in Section 4.3.2 that in the event of the contingency and
with a greater ability to decrease the power output at node 1, the level
of generation for unit 1 at t = 3 (equal to 95) would be lower than here.
This impact of the tighter limit on the maximum decrease in output
from unit 1 in the event of a contingency is also reflected in the dual
variable of the constraint on maximum power output decrease for unit
1 at t = 3: this marginal price is now equal to 10 (whereas it was 0
before).
Another dual quantity that is no longer zero is the marginal price
of transmission capacity from node 1 to node 4 at t = 3: this is now
equal to 10, reflecting the fact that the constraint on the power flow
between nodes 1 and 4 at t = 3 is binding.
5
AC Network-Constrained Unit Commitment
The impact of the power network was integrated in Chapter 3 using

the DC power flow equations. While the DC linearization has many
advantages, one drawback is that it provides no information about
several key quantities, among them the changes in voltage magnitude
and the reactive power flows. One of the consequences is that losses in
transmission cannot be accurately computed, and this can significantly
affect the conclusions drawn from the models.
In this chapter, we instead use the AC power flow equations. The
AC approach has the advantage that it is more accurate, but the result-
ing UC formulation is a mixed-integer nonlinear programming problem
that is computationally challenging. The formulation presented here is
based on the approach in Fu et al. [2005]. The computational aspects
are discussed in Chapter 8.
5.1 AC Power Flow
An electric current produces a magnetic field around it. When the

current is alternating, this magnetic field is constantly changing as a
result of the oscillations of the current. This change in the magnetic field
256
5.1. AC Power Flow 257
induces another electric current to flow in the same wire, in a direction

opposite to the flow of the original current. This phenomenon is called
reactance, and unless the voltage and current are perfectly in phase, it
limits the power that can be effectively transferred. The resistance of
the line limits the power flow as well, and with the reactance forms the
impedance of the line. For each line nm, we have
Znm = Rnm + jXnm
where Rnm is the resistance of the line, Xnm is its reactance, and j =
√
−1.
The admittance (the inverse of the impedance) is a measure of how
easily the current is allowed to flow. The admittance of the line is
denoted Ynm , and we have
1 Rnm Xnm
Ynm = = Gnm − jBnm = 2 2
−j 2 2 )
. (5.1)
Znm (Rnm + Xnm ) (Rnm + Xnm
Note that we already introduced Bnm in Section 3.1; susceptance is

another name for the imaginary part of admittance.
Reactive power accounts for the fact that power is not completely
transferred as active (or real) power when voltage and current are not
in phase. The resulting apparent power is equal to the magnitude of
the vector sum of active and reactive power.
In principle, generating units can provide both active and reactive
power, and loads can have demand for reactive power. We therefore let
qj (t) denote the reactive power output of generating unit j at period
t, and Qi (t) denote the reactive power load of demand i at period t.
Further, let Vn (t) be the voltage magnitude of node n, and S nm the
apparent power capacity of line nm. Finally, bshunt
nm is half of the shunt
susceptance of line nm (shunt means “in parallel”).
Using the quantities defined above, and denoting the difference be-
tween the voltage angles at the endpoints of line nm by
θnm (t) = θn (t) − θm (t),

258 AC Network-Constrained Unit Commitment
the AC power flow equations are:

X X
pj (t) − Di (t) =
j∈ΛG
n i∈ΛD
n
X h i
Vn (t)Vm (t) Gnm cos θnm (t) + Bnm sin θnm (t) − Gnm Vn2 (t)
m∈ΛL
n
(5.2)
X X
qj (t) − Qi (t) =
j∈ΛG
n i∈ΛD
n
X h i
Vn (t)Vm (t) Gnm sin θnm (t) − Bnm cos θnm (t) + Vn2 (t)(Bnm − bshunt
nm )
m∈ΛL
n
(5.3)
∀n, ∀t, and
2
Vn (t)Vm (t) (Gnm cos θnm (t) + Bnm sin θnm (t)) − Gnm Vn2 (t) +
h i 2
Vn (t)Vm (t) Gnm sin θnm (t) − Bnm cos θnm (t) + Vn2 (t)(Bnm − bshunt
nm )
2
≤ S nm , ∀n, ∀m ∈ ΛLn , ∀t (5.4)
Constraints (5.2) enforce active power balance for each node, and
constraints (5.3) do the same for reactive power balance. Constraints
(5.4) enforce transmission capacity limits (in terms of apparent power)
for each line.
5.2 Transformation of the AC Power Flow Equations to DC
The DC linearization of (5.2)–(5.4) is obtained by assuming that the

conductance is negligible, and that in all time periods, the variations
in voltage angle and voltage magnitude between the two nodes of a line
are small. Under these assumptions, we can make the approximations:
Gnm ≈ 0, cos θnm (t) ≈ 1, sin θnm (t) ≈ θnm (t),
and
|Vn (t)| = |Vm (t)| = 1 (in normalized units).
5.3. Integration of AC Network Constraints in UC 259
Substituting these approximations into (5.2), we obtain:

X X X
pj (t) − Di (t) = Bnm θnm (t),
j∈ΛG
n i∈ΛD
n m∈ΛL
n
which is precisely 3.1.

Note that because of the assumption that the voltage magnitudes
are all unity, the linearization cannot track reactive power flows. There-
fore, the constraint (5.3) is dropped. Finally, with the same approxima-
tions, constraints (5.4) are equivalent to (3.2) in the DC formulation.
5.3 Integration of AC Network Constraints in UC
Similarly to the development in Section 3.2, we integrate the AC net-

work into formulation (2.6) by replacing the constraints (2.2) with con-
straints (5.2), (5.3), and (5.4).
Moreover, constraints (2.3) need to be replaced by the constraints
below:
X X
pj (t) ≥ Di (t) + Rr (t) ∀r, ∀t (5.5)
j∈ΩG
r i∈ΩD
r
These constraints, which are identical to (3.4), enforce appropriate re-

serve levels in each reserve area. Similar constraints can be enforced for
reactive power, but for simplicity they are not stated in this tutorial.
Additional constraints include:
Vnmin ≤ Vn (t) ≤ Vnmax ∀n, ∀t, (5.6)

qjmin ≤ qj (t) ≤ qjmax ∀j, ∀t, (5.7)
where Vnmin and Vnmax are respectively the minimum and maximum
voltage magnitude at node n, and qjmin < 0 and qjmax > 0 are the
minimum and maximum reactive power output of generating unit j.
Finally, as in Section 3.1, we set a reference angle:
θn̂ (t) = 0 ∀t ∈ T, (5.8)
where n̂ is the node of the reference angle.

5.4 Recapitulation of the AC Network-Constrained UC For-

mulation
We use boxes to highlight the constraints and variables that are spe-
cific to the integration of the AC representation. The remainder of the
formulation is from the basic UC model in Section 2.6. For the sake of
simplicity, we do not account for minimum up/down times.
5.4. Recapitulation of the AC Network-Constrained UC Formulation 261
XX
j
Ξ
t∈T j∈J
s.t.
θn̂ (t) = 0, ∀t ∈ T
pj (t) ≥ ∀r ∈ R, ∀t ∈ T
P P
j∈ΩG
r i∈ΩD
r
Di (t) + Rr (t),
Vnmin ≤ Vn (t) ≤ Vnmax , ∀n ∈ N, ∀t ∈ T
qjmin vj (t) ≤ qj (t) ≤ qjmax vj (t) ∀j ∈ J, ∀t ∈ T
pj (t) −
P P
j∈ΛG
n i∈ΛD
n
Di (t) =
P h i
m∈ΛL
n
Vn (t)Vm (t) Gnm cos θnm (t) + Bnm sin θnm (t)
−Gnm Vn2 (t), ∀n ∈ N, ∀t ∈ T

−
P P
j∈ΛG
n
qj (t) i∈ΛD
n
Qi (t) =
h i
Vn (t)Vm (t) Gnm sin θnm (t) − Bnm cos θnm (t)
P
m∈ΛL
n
+Vn2 (t)(Bnm − bshunt

nm ), ∀n ∈ N, ∀t ∈ T
2
Vn (t)Vm (t) (Gnm cos θnm (t) + Bnm sin θnm (t)) − Gnm Vn2 (t)
h i
+ Vn (t)Vm (t) Gnm sin θnm (t) − Bnm cos θnm (t)
2
2
+Vn2 (t)(Bnm − bshunt
nm ) ≤ S nm , ∀n ∈ N, ∀m ∈ ΛLn , ∀t ∈ T
1) + zj (t + 1)SjD ,

pj (t) ≤ P j vj (t) − zj (t + ∀j ∈ J, ∀t ∈ T
zj (t), θn (t), qj (t) , and Vn (t) , ∀n ∈ N , ∀j ∈ J, ∀t ∈ T .
We consider the same power network with six buses as in Section 3.4,
with the same simplifying assumption that there is a single reserve
area, corresponding to the whole network.
Unlike the previous formulations, the formulation in Section 5.4 is
a mixed-integer nonlinear programming problem. We solved our exam-
ple using the KNITRO solver [Byrd et al., 2006] that is available on
NEOS and can handle this formulation for moderately sized instances
of UC. Large-scale instances of the AC network-constrained UC remain
computationally challenging; see Chapter 8 for more details.
5.5.1 Initial Case
We start with the same nodal demand and reserve requirements as in

Section 3.4.1:
D4 (t) 100 100 80 140 100 80
D5 (t) 90 100 80 30 90 60
D6 (t) 50 50 40 0 40 50
Total demand 240 250 200 170 230 190
Reserve R(t) 10 10 10 10 10 10
To allow for the increased power flow in apparent power due to ac-
counting for reactive power, we increase the line transmission capacity
limits S nm by 20% compared to the limits in Section 3.4.1:
1 2 3 4 5 6
1 - - - 120 120 120
2 - - - 120 120 120
3 - - - - 120 120
4 120 120 - - - -
5 120 120 120 - - -
6 120 120 120 - - -
We maintain the value of the susceptance Bnm = 0.8 for every line
and choose the following values for the additional parameters required
to describe the AC network:
• For every line, we set the conductance Gnm = 0.08 and the shunt
susceptance to zero.
• At each node, we set the minimum and maximum voltage mag-

nitudes (in per-unit) to Vnmin = 0.95 and Vnmax = 1.05.
• We set the nodal reactive power demands according to a typical

power factor of 0.75:
Q4 (t) 75.0 75.0 60.0 105.0 75.0 60.0
Q5 (t) 67.5 75.0 60.0 22.5 67.5 45.0
Q6 (t) 37.5 37.5 30.0 0.0 30.0 37.5
• We set the minimum and maximum reactive power capacity (in

absolute value) of the three generating units to 50% of their max-
imum active power output:
Unit qjmin qjmax

1 -150 150
2 -100 100
3 -50 50

eration schedule for active and reactive power:
p1 (t) 170.00 179.81 149.81 119.81 169.81 139.81
p2 (t) 69.73 69.90 50.00 50.00 59.92 50.00
p3 (t) 0 0 0 0 0 0
q1 (t) 150.00 150.00 135.05 65.86 150.00 145.95
q2 (t) 32.71 40.40 16.84 63.53 25.18 -1.55
q3 (t) 0 0 0 0 0 0
The total cost is 10388.5, up 14% from the cost of 9100 in Section
3.4.1. The voltage magnitudes are:
Node 1 0.9566 0.9563 0.9941 1.0498 0.9567 0.9838
Node 2 0.9520 0.9522 0.9896 1.0500 0.9518 0.9780
Node 3 0.9513 0.9511 0.9894 1.0493 0.9515 0.9786
Node 4 0.9500 0.9500 0.9886 1.0445 0.9500 0.9776
Node 5 0.9507 0.9504 0.9889 1.0489 0.9508 0.9785
Node 6 0.9519 0.9518 0.9900 1.0497 0.9522 0.9787

horizon is:
We see that the commitment of the units is greater than in Section

3.4.1; specifically unit 1 remains on for all time periods, and unit 2 is
now also on for all time periods (instead of only the first two periods).
Because of the increased commitment of unit 2, unit 3 is never called
upon (whereas before it came on at t = 5).
This global increase in the commitment of the units with respect
to the DC case is due to the need to provide voltage support, i.e., to
maintain the voltage magnitudes across the system within the desired
range of 0.95 to 1.05.
Let us now look at the power flows. For instance, at t = 1 the active
power flows are as follows:
1 - - - –66.76 -58.42 -45.06

2 - - - -33.24 -24.92 -11.60
3 - - - - -6.66 6.66
4 66.65 33.22 - - - -
5 58.33 24.91 6.66 - - -
6 45.01 11.61 -6.66 - - -
We see that the flows of active power are nearly identical to those in
Section 3.4.1. The small differences are negligible; the largest discrep-
ancy is observed on the line between 1 and 4 where the flow was 66.67
with the DC representation, and here the power injected at node 1 is
66.65, while the power that arrives at node 4 is 66.76. These discrep-
ancies are all less than 0.2% and essentially insignificant.
Because we are using the AC description for the network, we obtain
information about the reactive power flows as well. For t = 1 they are
as follows:
1 - - - -56.67 -50.44 -40.53

2 - - - -18.33 -12.04 -2.02
3 - - - - -5.02 5.04
4 57.72 18.53 - - - -
5 51.25 12.14 5.03 - - -
6 41.03 2.034 -5.03 - - -
Note that, unlike for the active flows, the reactive flows are not neces-
sarily symmetric, i.e., the flow of reactive power from node i to node j
is not necessarily the same as that from node j to node i.
Consider for instance the line between nodes 1 and 4: there is a
injection of 57.72 units at node 1, but only 56.67 units, or 1.8% less,
arrive at node 4. Similarly, 51.25 units are injected at node 1 into the
line connecting it to node 5, but only 50.44 units arrive at node 5,
nearly 1.6% less.
These differences in the reactive flows between two nodes reflect the
transmission losses on the respective lines.
5.5.2 Impact of Line Characteristics

Changes in the physical properties of the lines can significantly impact
the commitment outcomes. As an example, suppose that the value of
the line parameter bshunt
nm is greater than zero. In physical terms, this
means that the lines are longer and behave more like capacitors.
For illustration, setting bshunt
nm = 0.0002 for all the lines in our ex-
ample and solving the model, we obtain:
p1 (t) 170.00 199.68 199.74 169.74 199.68 189.77
p2 (t) 69.73 50.00 0 0 0 0
p3 (t) 0 0 0 0 30.00 0
q1 (t) 150.00 150.00 123.66 97.47 150.00 115.86
q2 (t) 3.73 11.64 0 0 0 0
q3 (t) 0 0 0 0 -3.33 0
The total cost is 9088.9, a decrease of 12.5% compared to the cost in
Section 5.5.1.
The voltage magnitudes are:
Node 1 0.9570 0.9567 0.9554 1.0129 0.9568 0.9553
Node 2 0.9511 0.9514 0.9510 1.0096 0.9513 0.9511
Node 3 0.9518 0.9515 0.9513 1.0114 0.9514 0.9515
Node 4 0.9500 0.9500 0.9500 1.0059 0.9500 0.9500
Node 5 0.9510 0.9506 0.9505 1.0107 0.9508 0.9512
Node 6 0.9521 0.9520 0.9517 1.0115 0.9522 0.9514
We see that the commitment of the units changes noticeably from that
in Section 5.5.1; while the commitment of unit 1 remains unchanged,
unit 2 is now committed for only the first two time periods, and unit
3 is brought on for t = 5 only.
If we look at the reactive power flows for t = 1, we have:
1 - - - -61.48 -53.45 -43.54

2 - - - -13.52 -5.40 4.626
3 - - - - -8.64 1.42
4 58.94 10.08 - - - -
5 50.64 1.87 5.04 - - -
6 40.41 -8.22 -5.04 - - -
We see that the differences have noticeably increased with the in-
crease in the shunt susceptance. If we again consider the line between
nodes 1 and 4, we see 58.94 units injected at node 1, and 61.48 units
arriving at node 4, an increase of 4.3%.
6
Stochastic Unit Commitment
A stochastic optimization formulation is relevant if the UC is affected

by important uncertainty in the data. Handling stochasticity is cur-
rently of great importance because of the uncertainty arising from the
variability in generation from stochastic production facilities such as
wind- and solar-based generating units. These types of generating units
often benefit from priority in dispatch by virtue of their low marginal
cost or regulatory policies. They are therefore not scheduled per se
but rather their production is subtracted from the demand, and other
units are then scheduled to meet the resulting net demand, i.e., the
actual demand minus the stochastic production. One consequence is
that the net demand curve fluctuations are increasing with the greater
penetration of such stochastic generation.
Demand fluctuations have traditionally been handled by ensuring a
sufficient level of reserve generation. This approach can be economically
inefficient, and the growing provision of power from stochastic sources
is increasing the cost of this inefficiency. With multiple jurisdictions
around the world experiencing significant increases in the proportion
of electricity generated by stochastic sources, the importance of models
explicitly incorporating the uncertainty in demand is increasing.
268
6.1. Two-Stage Stochastic Optimization 269
We present two approaches to handling the uncertainty in math-

ematical optimization models. The approach in this chapter follows
Morales et al. [2009] and uses stochastic optimization; in Chapter 7 we
present a robust optimization model.
Other mathematical optimization approaches to handle uncertainty
could also be used to formulate the UC problem, but they normally
involve the possibility that some constraints may not hold at the op-
timal solution. For instance, chance-constrained optimization is an es-
tablished technique based on the idea that with some (usually small)
probability, some or all of the constraints may not be satisfied at the
optimal solution. Chance-constrained optimization is useful in a variety
of contexts, but we believe that approaches that are not guaranteed to
enforce system constraints are unlikely to be adopted by power system
operators.
While there are other sources of uncertainty in UC, for the sake of
simplicity we consider only the uncertainty in net demand.
6.1 Two-Stage Stochastic Optimization
When formulating a stochastic UC, we consider two stages:
• The first stage pertains to the optimal scheduling of generation

capacity, i.e., the decisions about which units to commit in ad-
vance of the actual operation.
• The second stage constitutes a representation (prognosis) of a

number of plausible operating conditions that may arise in the
future as a result of the uncertainty realization. These possible
operating conditions are called scenarios, and for each scenario,
an optimal dispatch can be computed based on the commitment
decisions made in the first stage.
Reserves are scheduled in the first stage so that the system will be
able to accommodate any uncertainty realization, i.e., any operating
scenario.
The philosophy of this two-stage formulation is that in the first
stage, scheduling decisions are made using only the information that is
270 Stochastic Unit Commitment
available hours or days in advance of real-time operations. The uncer-

tainty is then realized in the second stage, and the dispatch adjusts the
amount scheduled in the first stage up or down, as required according
to the scenario.
The scenarios take into account the possible net demand realiza-
tions over the planning period. Each scenario is assigned a probability,
and the optimization objective is to minimize the sum of the deter-
ministic cost of the first-stage decisions and the expected cost of the
second-stage decisions.
Among the limitations of the stochastic optimization approach is
the fact that the quality of the solutions obtained critically depends on
the choice of the scenarios, in the sense that having a broader range of
scenarios usually leads to a more accurate model. However, increasing
the number of scenarios increases the computational cost of the opti-
mization (see Chapter 8 for additional discussion of the computational
aspects). Another issue is that this approach assumes explicit knowl-
edge of the probability distribution of the (uncertain) net demand. In
practice, this distribution is estimated empirically based on past data
and experience and/or using simulation models, and the limitations
of the probability estimation may impact the quality of the results.
The robust optimization approach presented in Chapter 7 avoids these
limitations.
6.2 Optimization Objective
Let Ω denote the set of scenarios, ω be the index for the scenarios,
and πω be the probability of occurrence of scenario ω. We assume that
P
ω∈Ω πω = 1.
Since the second-stage information is scenario-dependent, the opti-
mization model minimizes the deterministic cost of the first-stage de-
cisions (including the cost of the reserves scheduled) plus the expected
cost of the second-stage decisions (including the cost of the reserve
deployment actually called upon according to each scenario). For sim-
plicity, we consider the production and startup costs of all generating
units to be linear and constant, respectively, and the cost of deploying
6.3. First-Stage Constraints 271
reserves to be equal to the production cost of the generating unit that

deploys the reserve.
Under these assumptions, the objective function is given by:
XX X XX
cj pj (t) + cU
j yj (t) +
U
πω cj rjω D
(t) − rjω (t) (6.1)
t∈T j∈J ω∈Ω t∈T j∈J
where rjωU (t) and r D (t) are respectively the up-reserve and down-
jω
reserve deployed by generating unit j during time period t under sce-
nario ω.
6.3 First-Stage Constraints
The first-stage constraints are a subset of those from the basic formu-
lation in Section 2.6. We include the logical constraints (2.5) and the
generation limits (2.11), where in the latter we omit the reserve-related
variable pj (t). We also include constraint (2.2) to ensure that the total
amount of generation in time period t meets the expected net demand
Dexp (t) for that period:
X
pj (t) = Dexp (t), ∀t ∈ T. (6.2)
j∈J
There is no need to include the constraints related to the costs,

and for additional simplicity, we do not account explicitly for reserve
or for uptime and downtime requirements. Also for simplicity, we do
not include the ramping constraints because these are enforced in the
second stage, which is the actual operation stage. In some electricity
markets, the market rules require that ramping constraints be imposed,
in which case they must be included in the optimization problem.
6.4 Second-Stage Constraints
For the second stage, the first set of constraints states that the scenario-
dependent amount produced by unit j at time t under scenario ω
is equal to the amount scheduled in the first stage adjusted by the
amounts of up-reserve and down-reserve provided by unit j under sce-
nario ω:
U D
pjω (t) = pj (t) + rjω (t) − rjω (t), ∀j ∈ J, ∀t ∈ T, ∀ω ∈ Ω (6.3)
where pjω (t) is the actual output of unit j at time t under scenario ω.
The next constraints are versions of constraints (2.2), (2.7), (2.8),
and (2.11) (without the variable pj (t)), all of them indexed by ω:
X
pjω (t) = Dω (t), ∀t ∈ T, ∀ω ∈ Ω (6.4)
j∈J
pjω (t) − pjω (t − 1) ≤ RjU vj (t − 1) + SjU yj (t), ∀j ∈ J, ∀t ∈ T, ∀ω ∈ Ω

(6.5)
pjω (t − 1) − pjω (t) ≤ RjD vj (t) + SjD zj (t), ∀j ∈ J, ∀t ∈ T, ∀ω ∈ Ω
(6.6)
P j vj (t) ≤ pjω (t) ≤ P j vj (t), ∀j ∈ J, ∀t ∈ T, ∀ω ∈ Ω (6.7)
The final constraints link the commitment decisions and the various
operating conditions:
U
0 ≤ rjω (t) ≤ MjU vj (t), ∀j ∈ J, ∀t ∈ T, ∀ω ∈ Ω (6.8)
D
0≤ rjω (t) ≤ MjD vj (t), ∀j ∈ J, ∀t ∈ T, ∀ω ∈ Ω (6.9)
where MjU and MjD are respectively the maximum up-reserve and the
maximum down-reserve provided by generating unit j.
6.5 Recapitulation of the Stochastic UC Formulation
As mentioned above, for simplicity, we do not include minimum uptime

and downtime constraints, ramping constraints (in the first stage), or
network constraints. Also, no explicit (and thus arbitrary) reserve con-
straint is required in this formulation since the two-stage framework
explicitly takes the reserves into account. In addition to the constraints
from Sections 6.3 and 6.4, we include the logical coherence constraints
(2.5).
We use boxes to highlight the sets, constraints, and variables that
are specific to the integration of stochasticity. The stochastic UC for-
mulation is as follows:
XX
min cj pj (t) + cU
j yj (t)+
Ξ
t∈T j∈J

P P P U (t) − r D (t)
πω cj rjω
ω∈Ω t∈T j∈J jω
s.t.
Dexp (t), ∀t ∈ T
P
j∈J pj (t) =
U (t) − r D (t)
pjω (t) = pj (t) + rjω ∀j ∈ J, ∀t ∈ T, ∀ω ∈ Ω
jω
∀t ∈ T, ∀ω ∈ Ω
P
j∈J pjω (t) = Dω (t),
pjω (t) − pjω (t − 1) ≤ RjU vj (t − 1) + SjU yj (t), ∀j ∈ J,∀t ∈ T, ∀ω ∈ Ω
pjω (t − 1) − pjω (t) ≤ RjD vj (t) + SjD zj (t), ∀j ∈ J, ∀t ∈ T, ∀ω ∈ Ω
P j vj (t) ≤ pjω (t) ≤ P j vj (t), ∀j ∈ J, ∀t ∈ T, ∀ω ∈ Ω
U (t) ≤ M U v (t)
0 ≤ rjω ∀j ∈ J, ∀t ∈ T, ∀ω ∈ Ω
j j
D (t) ≤ M D v (t)
0 ≤ rjω ∀j ∈ J, ∀t ∈ T, ∀ω ∈ Ω
j j
where the optimization variables in set Ξ are pj (t), vj (t), yj (t), zj (t),
D (t) , r U (t) , ∀j ∈ J, ∀t ∈ T , ∀ω ∈ Ω .
pjω (t) , rjω jω
The formulation in Section 6.5 is a mixed-integer linear optimization

problem, and we solved it using BARON. The computational cost of
solving the stochastic formulation depends on the number of scenarios
|Ω|. While this was not an issue for this small example, techniques for
mitigating the limitations arising from a large number of scenarios are
discussed in Section 8.1.3.
6.6.1 Initial Case

We consider a system with three generators as described at the begin-
ning of Section 2.7. We additionally set the maximum up-reserve and
down-reserve provided by each generating unit j as follows:
Unit (j) MjU MjD

1 25 15
2 30 20
3 35 25
We set the expected net demand to the following levels:
Dexp (t) 220 250 200 170 230 190
Recall that no explicit reserves are required here.

We consider three scenarios defined as follows: scenario 1 corre-
sponds to demands 10% lower than the expected demand in every time
period; scenario 2 corresponds to demands equal to the expected de-
mand; and scenario 3 corresponds to demands 5% higher than the ex-
pected demand. Let us assign the three scenarios probabilities of 15%,
60%, and 25% respectively. Solving the model in Section 6.5, we obtain
the following results:
First-stage
p1 (t) 160 175 145 118 166 125
p2 (t) 60 75 55 52 64 65
p3 (t) 0 0 0 0 0 0
Scenario 1
p11 (t) 148 160 130 103 151 121
p21 (t) 50 65 50 50 56 50
p31 (t) 0 0 0 0 0 0
Scenario 2
p12 (t) 170 180 150 120 170 140
p22 (t) 50 70 50 50 60 50
p32 (t) 0 0 0 0 0 0
Scenario 3
p13 (t) 170 188 158 128 178 150
p23 (t) 61 75 52 50 65 50
p33 (t) 0 0 0 0 0 0
The total cost is 10130.75, and the on/off status of the generators
is:
6.6.2 Impact of Changing the Probabilities

The probabilities assigned to each of the scenarios impact the cost of
the optimal solution. To illustrate this, suppose that we change these
probabilities to 40%, 50%, and 10% respectively, so that scenario 1 is
more likely, and scenario 3 is less likely. The result is that, while the
commitment decisions are unchanged, the total cost decreases by 2.6%
to 9867.50, with no changes in the dispatch for the scenarios and small
changes to the output schedule in the first stage:
First-stage
p1 (t) 163 170 140 103 166 136
p2 (t) 57 80 60 67 64 54
p3 (t) 0 0 0 0 0 0
6.6.3 Impact of Changing the Scenarios

The choice of scenarios also impacts the cost of the optimal solution.
Suppose that we change scenario 3 to represent demands 15% higher
than the expected demand in every time period, and that the other
parameters are the same as in the initial case in Section 6.6.1. The
commitment decisions are again unchanged, but the total cost increases
by 10.6% to 11200.75.
There are no changes in the production schedule for the unchanged
scenarios, and the new output schedules for the first stage and for
scenario 3 are:
First-stage
p1 (t) 163 175 145 118 166 136
p2 (t) 57 75 55 52 64 54
p3 (t) 0 0 0 0 0 0
Scenario 3
p13 (t) 188 200 170 143 191 161
p23 (t) 65 88 60 53 74 58
p33 (t) 0 0 0 0 0 0
7
Robust Unit Commitment
Adaptive robust optimization is a technique for modeling uncertainty

that overcomes some of the drawbacks of stochastic optimization at the
cost of comparatively more restrictive modeling assumptions. Robust
optimization avoids the need to define scenarios and make assumptions
about their probabilities. Instead, a deterministic uncertainty set is de-
fined using limited information about the uncertain quantities: their
expected value and some estimate of their variance (or a range of pos-
sible deviations from the expected value). If additional information is
available, it can often be incorporated into this framework to poten-
tially improve the quality of the robust model.
In the context of UC, and under the same assumptions as those
pertaining to the stochastic programming formulation, the uncertainty
set is defined in terms of possible realizations of net demand. Once the
uncertainty set is defined, the model computes an optimal solution that
protects the system against every possible realization in the set, and
in particular the worst case. In this sense, the robust approach is more
conservative than the stochastic one.
277
278 Robust Unit Commitment
7.1 Three-Level Adaptive Robust Formulation
The robust formulation considered here was proposed in Bertsimas

et al. [2013]. It has three levels, according to the following logic:
First level The operator schedules the production, i.e., commits the
units, so as to minimize the generation cost. This is similar to the
first stage in the stochastic optimization model.
Second level For every schedule feasible for the first level, the net
demand realizes in the worst possible manner within the uncer-
tainty set, i.e., the production cost is maximized.
Third level Given the schedule and the worst-case demand realiza-
tion, the operator dispatches the committed units so as to mini-
mize the production cost.
Note that the robust UC takes into account all possible future de-
mands represented in the uncertainty set. The optimal robust UC so-
lution therefore will be feasible for any realization of the uncertainty
in the second level. This is in contrast with the basic formulation in
Chapter 2 that guarantees feasibility only for a single set of demands,
and the stochastic optimization formulation in Chapter 6 that considers
only a finite set of preselected scenarios.
7.2 Uncertainty Set
To define the uncertainty set, we first introduce the minimum and

maximum levels for the realization of demand at time t, denoted D(t)
and D(t) respectively. It is between these two levels that the second-
level problem can realize d(t), the net demand at time t:
D(t) ≤ d(t) ≤ D(t), ∀t ∈ T (7.1)
We further introduce the budget of uncertainty Γ. This is a parameter

that bounds the total deviation allowed, summed over all t ∈ T , for the
realized demand d(t) from D(t).
7.3. Optimization Objective 279
Given Γ, the uncertainty set is a deterministic set defined by the

following constraints:
X max{0, d(t) − D exp (t)}
≤ Γ, (7.2)
t∈T
D(t) − Dexp (t)
where, as in Section 6.3, Dexp (t) is the expected net demand. The
rationale behind this definition follows from the fact that our main
concern is to schedule enough capacity to meet demand, so we want to
be protected against unexpected increases in demand. Therefore, the
numerator in (7.2) measures the deviation from expected demand only
if d(t) lies in the interval [Dexp (t), D(t)]. Except for the presence of the
max function, (7.2) is a linear constraint on the variable d(t), and if
required, the max function can be linearized using binary variables.
The concept of a budget of uncertainty has an elegant practical
interpretation. The value of Γ can be chosen anywhere between 0 and
|T |. The value 0 corresponds to the realized demand d(t) being equal to
Dexp (t), i.e., no deviation is allowed (and hence there is no uncertainty).
At the other extreme, the value |T | corresponds to requiring protection
for the maximum possible deviation of d(t), i.e., d(t) is equal to D(t).
The higher the value of Γ, the higher the robust protection.
In other words, for smaller values of Γ, the realized demand d(t)
cannot deviate too much from Dexp , and hence there is less uncertainty
in d(t), As Γ increases, the range of values allowed for d(t) increases,
and so does the uncertainty.
7.3 Optimization Objective
The overall objective is to minimize the total operation cost of the

system. The optimization objective is in three parts, each corresponding
to one of the levels of the robust formulation:
 
XX XX
min  cU
j yj (t) + max min cj pj (t)  (7.3)
Ξ1 Ξ2 Ξ3
t∈T j∈J t∈T j∈J
where the variables for each optimization are:
• Ξ1 = {vj (t), yj (t), zj (t), ∀j ∈ J, ∀t ∈ T }

• Ξ2 = {d(t), ∀t ∈ T }
• Ξ3 = {pj (t), ∀j ∈ J, ∀t ∈ T }
The structure of this objective is aligned with the outline of the robust
formulation in Section 7.1: the commitment of the units is determined
using the variables in Ξ1 with the objective that the total cost is mini-
mized. The total cost is in two parts. The first part is the startup cost
(which we assume to be constant for simplicity), and the second part
is the cost of production.
The cost of production, given a feasible commitment, is maximized
over the demand variables in Ξ2 . This is where the demand is realized
in the worst possible manner while being restricted to lie in the uncer-
tainty set defined in Section 7.2. However, for each realization of the
demand, the cost of dispatch is minimized by the operator who assigns
the amount of power produced by each unit via the variables in Ξ3 ; the
constraints ensure that only units scheduled to be on can be used by
the operator.
7.4 First-Level Constraints
The first-level constraints are the logical constraints (2.5) that ensure
consistency of the values of the binary variables with respect to the
decisions to start up or shut down units.
7.5 Second-Level Constraints
As second-level constraints, we include (7.1) and (7.2) to define the

uncertainty set as described in Section 7.2.
We also enforce ramping constraints on the net demand to represent
the fact that the demand cannot change too rapidly:
d(t) − d(t − 1) ≤ RdU , ∀t ∈ T (7.4)

d(t − 1) − d(t) ≤ RdD , ∀t ∈ T (7.5)
7.6. Third-Level Constraints 281
7.6 Third-Level Constraints
For the third level, the constraints are the fundamental operational
requirements: balance of supply and demand (2.2), limits on ramping
generation up (2.7) or down (2.8), and generation limits as in (2.11) but
omitting the variable pj (t). For simplicity, we do not account explicitly
for dispatching reserves or for uptime and downtime requirements.
7.7 Recapitulation of the Robust UC Formulation
We use boxes to highlight the parts of the formulation that are spe-
cific to the robust approach. As noted earlier, reserve requirements, as
well as uptime and downtime limitations, are omitted for the sake of
simplicity.

P P U P P
min t∈T j∈J cj yj (t) + max min t∈T j∈J cj pj (t)
Ξ1 Ξ2 Ξ3
s.t.
s.t.
D(t) ≤ d(t) ≤ D(t), ∀t ∈ T
max{0,d(t)−Dexp (t)}
≤ Γ,
P
t∈T D(t)−Dexp (t)
d(t) − d(t − 1) ≤ RdU , ∀t ∈ T

d(t − 1) − d(t) ≤ RdD , ∀t ∈ T
s.t.
X
pj (t) = d(t), ∀t ∈ T
j∈J

where the optimization variable sets for the three levels are respectively:
• Ξ1 = {vj (t), yj (t), zj (t), ∀j ∈ J, ∀t ∈ T }

• Ξ2 = {d(t), ∀t ∈ T }
• Ξ3 = {pj (t), ∀j ∈ J, ∀t ∈ T }
The formulation in Section 7.7 is a three-level optimization problem,

and even small instances are not straightforward to solve. Indeed, even
two-level optimization (usually known as bilevel optimization) is known
to be challenging.
In view of this, we use a smaller example than in the previous
chapters. Specifically, we consider an example with two generators and
three time periods, and we apply the solution methodology described
in Section 8.2.
7.8.1 Initial Case
We consider the generating units 1 and 2, with the respective param-

eters and operating conditions at t = 0 as given in Section 2.7. In
particular, at t = 0, unit 1 is on and producing 120 MW, and unit 2 is
off.
We assume that the characteristics of the demand for the three time
periods are as follows:
Time period t 1 2 3
D(t) 130 190 150
Dexp (t) 170 230 190
D(t) 210 270 230
We further assume that d(0) = 200, and that the ramping limits on
the net demand are RdU = 50 and RdD = 50.
For this initial case, we set Γ = 2.0 and solve the model in Section
7.7. To follow the approach described in Section 8.2, we need to enu-
merate all the combinations of the commitment decisions v, y, and z.
Because the first-level constraints (2.5) completely determine the val-
ues of yj (t) and zj (t) given the values of the vj (t) variables, it suffices
to enumerate the possible values of vj (t) for j = 1, 2 and t = 1, 2, 3.
Further, note that because p1 (0) = 120 and S1D = 80, it follows
that v1 (1) = 1 must hold. This leaves us with 25 = 32 combinations
to enumerate. When we solve the optimization problem (8.2) for each
of these combinations, we find that most of them have no feasible so-
lution, i.e., the binary variables are set to combinations for which it is
impossible to satisfy all the constraints. In many of these cases, there is
simply not enough generation capacity committed to satisfy demand.
The following table gives the optimal value (maximum cost) for the
nine feasible combinations:
v1 (2) v1 (3) v2 (1) v2 (2) v2 (3) Cost

1 1 0 0 0 3100
1 1 0 1 0 4050
1 1 1 0 0 4250
1 1 1 1 0 5050
1 0 1 1 1 7300
1 1 0 0 1 4150
1 1 0 1 1 5450
1 1 1 0 1 5350
1 1 1 1 1 7350
We deduce that the optimal solution to the robust model has a to-
tal cost of 3100 and corresponds to the following on/off status of the
generators:
Time period t 1 2 3
Unit 1 on on on
Unit 2 off off off
The corresponding generation schedule is:
Time period t 1 2 3
p1 (t) 170 220 230
p2 (t) 0 0 0
The realized demand is d(t) = p1 (t), t=1,2,3 at optimality.

7.8.2 Impact of the Choice of Γ

As mentioned in Section 7.2, the parameter Γ represents a budget of
uncertainty in the realized demand within which the second-level op-
timization problem can maximize the cost of production (for a set of
commitment decisions fixed by the first-level problem). Let us examine
the sensitivity of the optimal value of the second-level problem for the
commitment corresponding to the optimal solution in Section 7.8.1,
namely
v1 (1) = 1, v1 (2) = 1, v1 (3) = 1, v2 (1) = 0, v2 (2) = 0, and v2 (3) = 0.
We already know that for Γ = 2.0, the optimal value is 3100. In

principle, Γ can take any value between 0 and 3 (= |T |). Values of Γ
equal to, or close to, zero correspond to fixing, or nearly fixing, d(t) =
Dexp (t) for all t ∈ T .
Solving the optimization problem (8.2) for six additional values of
Γ, we obtain the following:
Γ Optimal
(maximum) cost
0 2900
0.5 3000
1.0 3100
1.5 3100
2.0 3100
2.5 3100
3.0 3100
We observe that reducing the budget of uncertainty reduces the optimal

cost, as would be expected.
8
Computational Aspects
As stated in Chapter 1, the UC problem is in general hard to solve.

All the formulations presented in this tutorial are in practice large-
scale nonconvex problems, and thus they require state-of-the-art op-
timization techniques, and sometimes heuristics, to yield solutions for
practical use. In Section 8.1 we sketch selected research developments
that have improved, or potentially will improve, the ability to solve UC
formulations.
Throughout this tutorial, we have reported computational results
for small examples to illustrate UC solutions under different assump-
tions. We describe in Section 8.2 how these results were obtained.
8.1 Computational Efficiencies
8.1.1 Choice of Binary Variables
The reader may have already observed that the values of y(t) and z(t)
can be determined from the values of v(t). This observation can be
used to reduce the number of binary variables. This, however, is not
computationally effective, as shown in Ostrowski et al. [2012].
285
286 Computational Aspects
An alternative approach to reducing the number of variables was

proposed in Morales-España et al. [2013], where a distinction is made
between the ramping during startup/shutdown and the operational
ramping that occurs when the unit is operating above the minimum
production level P .
One of the general principles underlying integer optimization is that
of divide-and-conquer. This strategy takes the form of an algorithm
called branch-and-bound (or implicit enumeration) that solves a version
of the problem with the binary constraints relaxed to box constraints,
then recursively partitions the feasible region of the optimization prob-
lem into smaller, more manageable subproblems. There are a number
of ways to form the partitions, but the basic approach is to fix one of
the binary variables to each of the two possible values in turn, thus
generating two subproblems. This process can be represented using a
branch-and-bound tree.
Most (mixed) integer optimization solvers use some form of branch-
and-bound algorithm, and therefore the number of binary variables
matters in the sense that a smaller number of binary variables typi-
cally decreases the computational time. The general idea is that when
fewer variables are relaxed from binary to continuous, the algorithm
will require fewer subproblems, and thus the branch-and-bound tree
will be smaller. This, however, does not apply to many MILP prob-
lems, as shown for the UC in Ostrowski et al. [2012].
Some specific and general observations can be made in this regard.
In the specific case of papers about the UC problem, one observation
is that the conclusions that can be drawn from a direct comparison of
results is limited by the fact that different authors use different solvers,
or different versions of the same solver, and frequently different data
sets as well.
In general, it is important to realize that there is a difference be-
tween the computational time required to find a near-optimal solution
and the time required to prove its distance from optimality. The latter
often requires significantly more time than the former, and the times
reported again vary across the literature.
8.1. Computational Efficiencies 287
Furthermore, it is possible that formulations with more binary vari-

ables are easier to solve; see, e.g., Bertsimas and Weismantel [2005].
One reason is that it is not always true that when fewer variables are
relaxed from binary to continuous, fewer variables will be branched
on, and thus the branch-and-bound tree will be smaller. It may hap-
pen that branching on a y(t) or z(t) variable (if they are present) is
better than choosing any of the v(t) variables. Although making the
“best” branching decision is difficult in general (even for sophisticated
solvers), a larger pool of branching candidates is likely to result in a
smaller enumeration tree when a state-of-the-art solver is used.
Another reason is that in an integer (linear or nonlinear) optimiza-
tion problem, it often pays off to have more variables if the resulting
relaxation is tighter, i.e., closer to the convex hull of the feasible inte-
ger points. This is especially true for problems with binary variables. In
general, a formulation with more binary variables will be solved more
efficiently by branch-and-bound than a formulation with fewer binary
variables but for which the convex hull of the feasible points is not so
closely described. This is the fundamental idea underlying the concept
of lift-and-project in integer programming; see, e.g., Balas and Fischetti
[2017].
8.1.2 Approximations of the AC Power Flow
The integration of an AC network description as in Chapter 5 makes

the UC problem much more challenging computationally. This is be-
cause the AC power flow equations (5.2), (5.3), and (5.4) are nonlinear
and nonconvex, resulting in a significantly more difficult problem. One
of the key differences is that in mixed-integer linear optimization, if it is
possible to compute the solution, it is guaranteed that it is the best pos-
sible (or global optimal) solution. In contrast, mixed-integer nonlinear
optimization problems typically have multiple local optimal solutions,
making global optimality much more difficult to attain. There is thus
a preference for linear models whenever possible.
The UC formulation with AC network constraints in Section 5.4
is generally solved using decomposition techniques that separate the
binary variables and the nonlinearities. The basic idea is to solve a
simplified (typically linear or linearized) version of the problem and

then check compliance of the solution with the AC power flow equa-
tions. Violations of transmission flows and/or bus voltage requirements
lead to additional constraints that are added to the simplified problem,
and it is solved again. This technique is referred to as Benders’ decom-
position [Geoffrion, 1972] and was implemented in Fu et al. [2005] for
a version of SCUC with an AC network description.
A novel approach to linearize the AC power flows was proposed in
Coffrin and Van Hentenryck [2014] using linear approximations based
on accounting for the voltage phase angles and the voltage magnitudes,
and using a piecewise linear approximation of the cosine term in the
power flow equations. Specifically, three linearizations are proposed:
hot-start, warm-start, and cold-start. The linearizations differ according
to the voltage information available before the power flow is calculated.
The hot-start approximation assumes that a base-point AC solution is
known, and hence the model can use additional information such as
initial voltage magnitudes. The cold-start approximation assumes no
voltage information, and the warm-start approximation assumes that
target voltages are given (although an actual solution may not exist for
the target values).
8.1.3 Scenario Generation and Reduction
One drawback of two-stage stochastic optimization is that it assumes

explicit knowledge of the probability distribution of the (uncertain) net
demand. In practice, this distribution is estimated empirically based
on past data and experience and/or using simulation models, so the
limitations of the probability estimation may impact the quality of the
results.
Another drawback is that as the number of scenarios increases, the
optimization problems become larger and more challenging. This is sig-
nificant because, as mentioned in Chapter 6, the quality of the solutions
depends on the choice of the scenario set Ω. Generally speaking, increas-
ing the number of scenarios will increase the accuracy of the model, as
long as the choice of scenarios is sensible. However, a larger set of sce-
narios also means a greater number of variables and constraints, and
8.2. Solving the Formulations 289
thus an increase in the number and size of optimization problems that

must be solved.
The increase in computational cost can be mitigated by using sce-
nario reduction [Dupačová et al., 2003, Heitsch and Römisch, 2003].
Scenario reduction algorithms select a subset of the initial scenario set
and compute updated probabilities for the scenarios in that subset. All
other scenarios are effectively assigned a probability of zero, and the
optimization is then carried out using only the subset of scenarios. The
scenario reduction is based on a well-chosen definition of distance be-
tween the original and the reduced sets of probabilities. The idea is to
trade off scenario probabilities and distances between scenarios, so that
the selection process reduces the number of scenarios that have small
probabilities or are close to the remaining ones (in the probabilistic
sense).
The robust optimization approach presented in Chapter 7 bypasses
scenario generation and the associated computational issues. On the
other hand, the quality of its results greatly depends on the construc-
tion of the uncertainty set (Section 7.2); this process must be guided by
the characteristics of the power system under consideration. Moreover,
the robust formulation is challenging from a computational perspective;
this is discussed in the next section.
8.2 Solving the Formulations
The formulations in Sections 2.6, 3.3, and 4.2 are mixed-integer linear
optimization problems that can be solved using state-of-the-art soft-
ware. This is the technology that is commonly used nowadays to solve
UC in various jurisdictions on a day-ahead basis.
The formulation in Section 5.4 is also mixed-integer but with non-
linear constraints. As mentioned in the introduction to Chapter 5 and in
Section 8.1.2, mixed-integer nonlinear optimization is in general very
challenging. For our small example, a state-of-the-art mixed-integer
nonlinear optimization solver was adequate.
The formulation in Section 6.5 is a mixed-integer linear optimiza-
tion problem, but its solution depends on the number of scenarios |Ω|.
By keeping the number of scenarios small, we were able to use mixed-

integer linear optimization solvers here as well.
Finally, the formulation in Section 7.7 is challenging, even for small
instances, because it is a three-level optimization problem that requires
sophisticated optimization technology [Bertsimas et al., 2013]. To solve
our example, we proceeded as follows:
1. We start with the third-level problem:

XX
min cj pj (t)
pj (t)
t∈T j∈J
s.t.
X
d(t) − pj (t) = 0, ∀t ∈ T
j∈J
pj (t) − pj (t − 1) − [RjU vj (t − 1) + SjU yj (t)] ≤ 0, ∀j ∈ J, ∀t ∈ T

pj (t − 1) − pj (t) − [RjD vj (t) + SjD zj (t)] ≤ 0, ∀j ∈ J, ∀t ∈ T
P j vj (t) − pj (t) ≤ 0, ∀j ∈ J, ∀t ∈ T
pj (t) − P j vj (t) ≤ 0, ∀j ∈ J, ∀t ∈ T
We derive its dual problem, and hence its optimality conditions

as a set of (nonlinear) equations.
2. These conditions are then added as constraints in the second-level

problem, leading to an augmented second-level problem:
XX
max cj pj (t)
d(t)
t∈T j∈J
s.t.
D(t) ≤ d(t) ≤ D(t), ∀t ∈ T
≤ Γ, ∀t ∈ T
t∈T
D(t) − Dexp (t)
d(t) − d(t − 1) ≤ RdU , ∀t ∈ T
d(t − 1) − d(t) ≤ RdU , ∀t ∈ T
pj (t) is optimal for third-level problem with demand d(t)

3. Finally, we solve the first-level problem:

XX
max cU
j yj (t)
vj (t),yj (t),zj (t)
t∈T j∈J
s.t.
pj (t) and d(t) are optimal for augmented second-level problem
We carry out full enumeration of the possible combinations of
the binary variables, and for each combination we solve the aug-
mented second-level problem as above.
8.2.1 Dual Problem and Optimality Conditions for the Third-Level

Problem
The third-level problem can be expressed in the following form, with
dual variables λi associated with each constraint:
XX
min cj pj (t) (8.1)
pj (t)
t∈T j∈J
s.t.
X
d(t) − pj (t) = 0, ∀t ∈ T ← β(t)
j∈J
pj (t) − pj (t − 1) − K U (j, t) ≤ 0, ∀j ∈ J, ∀t ∈ T ← λ2 (j, t)

D
pj (t − 1) − pj (t) − K (j, t) ≤ 0, ∀j ∈ J, ∀t ∈ T ← λ3 (j, t)
LB
K (j, t) − pj (t) ≤ 0, ∀j ∈ J, ∀t ∈ T ← λ4 (j, t)
UB
pj (t) − K (j, t) ≤ 0, ∀j ∈ J, ∀t ∈ T ← λ5 (j, t)
where the following quantities are constant for fixed j ∈ J and t ∈ T :
K U (j, t) = RjU vj (t − 1) + SjU yj (t),

K D (j, t) = RjD vj (t) + SjD zj (t),
K LB (j, t) = P j vj (t), and
UB
K (j, t) = P j vj (t).
Note that the dual variable β(t) can be positive or negative, while the
variables λi must be non-negative.
Because (8.1) is a linear optimization problem, it has a dual that is

also a linear optimization problem, and the optimal values of these two
problems are equal at optimality; see, e.g., Vanderbei [2001], Theorem
5.2. The dual can be obtained using the primal–dual conversion rules;
see, e.g., Vanderbei [2001], Chapter 5, Section 8. Alternatively, it can
be derived from first principles as shown in Appendix A.
The dual is:
X X
max pj (0) − λ2 (j, 1) + λ3 (j, 1) + β(t)d(t)
λi (j,t),β(t)
j∈J t∈T
XXh
+ −λ2 (j, t)K U (j, t) − λ3 (j, t)K D (j, t)
t∈T j∈J
i
+ λ4 (j, t)K LB (j, t) − λ5 (j, t)K UB (j, t)
s.t.
cj − β(t) + λ2 (j, t) − λ3 (j, t) − λ4 (j, t) + λ5 (j, t)
− λ2 (j, t + 1) + λ3 (j, t + 1) = 0, ∀j ∈ J, t = 1, . . . , T − 1,
cj − β(T ) + λ2 (j, T ) − λ3 (j, T ) − λ4 (j, T ) + λ5 (j, T ) = 0, ∀j ∈ J
λi (j, t) ≥ 0, i = 2, 3, 4, 5,
β(t) free
Using the dual, there are two ways to obtain optimality conditions
for (8.1):
1. The third-level problem can be replaced by its optimality condi-

tions, which consist of:
• primal constraints (obtained from the primal)

• dual constraints (obtained from the dual)
• no duality gap (objective function of the primal = objective
function of the dual)
2. Because the second-level problem is a maximization problem and

the dual of the third-level problem is also a maximization prob-
lem, we can merge them into a single-level maximization.
Option 2 is generally more efficient than option 1, but for illustra-

tive purposes we choose option 1. We emphasize that both options are
equally valid.
By the duality theory of linear optimization, the conditions for si-
multaneous optimality of the primal and dual are
Primal feasibility:
X
d(t) − pj (t) ≤ 0, ∀t ∈ T
j∈J
pj (t) − pj (t − 1) − K U (j, t) ≤ 0, ∀j ∈ J, ∀t ∈ T
D
pj (t − 1) − pj (t) − K (j, t) ≤ 0, ∀j ∈ J, ∀t ∈ T
LB
K (j, t) − pj (t) ≤ 0, ∀j ∈ J, ∀t ∈ T
UB
pj (t) − K (j, t) ≤ 0, ∀j ∈ J, ∀t ∈ T
Dual feasibility:
cj − β(t) + λ2 (j, t) − λ3 (j, t) − λ4 (j, t) + λ5 (j, t) − λ2 (j, t + 1)

+λ3 (j, t + 1) = 0, ∀j ∈ J, t = 1, . . . , T − 1
λi (j, t) ≥ 0, i = 2, 3, 4, 5
No duality gap:
XX X X
cj pj (t) = pj (0)(−λ2 (j, 1) + λ3 (j, 1)) + β(t)d(t)
t∈T j∈J j∈J t∈T
XXh
+ −λ2 (j, t)K U (j, t) − λ3 (j, t)K D (j, t) + λ4 (j, t)K LB (j, t)
t∈T j∈J
i
−λ5 (j, t)K UB (j, t)
These optimality conditions can now be added as constraints to the

second-level problem. (Note that the set of variables increases as a
consequence.)
The resulting formulation of the second-level problem is:

XX
max cj pj (t)
Ξ̃2 t∈T j∈J
s.t.
D(t) ≤ d(t) ≤ D(t), ∀t ∈ T
≤ Γ, ∀t ∈ T
t∈T
D(t) − Dexp (t)
d(t) − d(t − 1) ≤ RdU , ∀t ∈ T
d(t − 1) − d(t) ≤ RdU , ∀t ∈ T
X
d(t) − pj (t) ≤ 0, ∀t ∈ T
j∈J
pj (t) − pj (t − 1) − K U (j, t) ≤ 0, ∀j ∈ J, ∀t ∈ T
pj (t − 1) − pj (t) − K D (j, t) ≤ 0, ∀j ∈ J, ∀t ∈ T
K LB (j, t) − pj (t) ≤ 0, ∀j ∈ J, ∀t ∈ T
pj (t) − K UB (j, t) ≤ 0, ∀j ∈ J, ∀t ∈ T
cj − β(t) + λ2 (j, t) − λ3 (j, t) − λ4 (j, t) + λ5 (j, t) − λ2 (j, t + 1)
+ λ3 (j, t + 1) = 0,
∀j ∈ J, t = 1, . . . , T − 1,
λi (j, t) ≥ 0, i = 2, 3, 4, 5
XX X X
cj pj (t) = pj (0)(−λ2 (j, 1) + λ3 (j, 1)) + β(t)d(t)
t∈T j∈J j∈J t∈T
XXh
t∈T j∈J
i
−λ5 (j, t)K UB (j, t)
where the optimization variables are Ξ̃2 = {d(t), pj (t), β(t), λi (j, t), i =
2, 3, 4, 5, ∀j ∈ J, ∀t ∈ T }.
The resulting problem is not a linear optimization problem because
of the presence of the |T | bilinear terms d(t)β(t). However, it can be
solved using a nonlinear optimization solver.
9
Conclusions and Directions for Future Research
Mixed-integer linear optimization was first proposed for UC problems

more than 50 years ago [Garver, 1962], and it is now used daily by power
system operators. Nevertheless, plenty of challenges remain when it
comes to solving UC problems that incorporate various types of prac-
tical requirements, such as security constraints, network effects and
power losses, and different uncertainty models. All of these have been
touched on in this tutorial, and they remain active research areas.
There is also great interest in incorporating into UC models the new
requirements of smart grids, such as the integration of wind and solar
generation, the management of demand-response, and the scheduling
of electricity storage devices.
In summary, even after several decades of research, the UC problem
continues to pose challenges to practitioners and researchers alike, and
there is plenty of scope for novel models and algorithms.
295
Appendices
A
Derivation of the Dual of the Third-Level
Problem in Section 8.2.1
We start with the third-level problem as stated at the beginning of

Section 8.2.1:
XX
min cj pj (t)
pj (t)
t∈T j∈J
s.t.
X
d(t) − pj (t) = 0, ∀t ∈ T ← β(t)
j∈J
pj (t) − pj (t − 1) − K U (j, t) ≤ 0, ∀j ∈ J, ∀t ∈ T ← λ2 (j, t)

pj (t − 1) − pj (t) − K D (j, t) ≤ 0, ∀j ∈ J, ∀t ∈ T ← λ3 (j, t)
LB
K (j, t) − pj (t) ≤ 0, ∀j ∈ J, ∀t ∈ T ← λ4 (j, t)
UB
pj (t) − K (j, t) ≤ 0, ∀j ∈ J, ∀t ∈ T ← λ5 (j, t)
where β(t) and λi are the dual variables, or Lagrange multipliers, as-
sociated with each constraint, β(t) is unconstrained, the λi are non-
negative, and the following quantities are constant for fixed j ∈ J and
297
298 Derivation of the Dual of the Third-Level Problem in Section 8.2.1
t ∈ T:
K U (j, t) = RjU vj (t − 1) + SjU yj (t),
K D (j, t) = RjD vj (t) + SjD zj (t),
K LB (j, t) = P j vj (t), and
UB
K (j, t) = P j vj (t).
We start by writing down the Lagrangian function for this problem:
XX
L(pj (t), β(t), λi (j, t)) = cj pj (t) (A.1)
t∈T j∈J
 
X X
+ β(t) d(t) − pj (t)
t∈T j∈J
XX h i
+ λ2 (j, t) pj (t) − pj (t − 1) − K U (j, t)
t∈T j∈J
XX h i
+ λ3 (j, t) pj (t − 1) − pj (t) − K D (j, t)
t∈T j∈J
XX h i
+ λ4 (j, t) K LB (j, t) − pj (t)
t∈T j∈J
XX h i
+ λ5 (j, t) pj (t) − K UB (j, t)
t∈T j∈J
The dual is:
max min L(pj (t), β(t), λi (j, t)) (A.2)
λi (j,t)≥0,β(t) pj (t)
Let us group the terms of the Lagrangian function (A.1) as follows:

X Xh
cj pj (t) − β(t)pj (t) + λ2 (j, t)pj (t) − λ3 (j, t)pj (t) − λ4 (j, t)pj (t)
t∈T j∈J
(A.3)
i
+ λ5 (j, t)pj (t)
X Xh i
+ −λ2 (j, t)pj (t − 1) + λ3 (j, t)pj (t − 1)
t∈T j∈J
Xh X
+ β(t)d(t) + −λ2 (j, t)K U (j, t) − λ3 (j, t)K D (j, t)
t∈T j∈J
i
+λ4 (j, t)K LB (j, t) − λ5 (j, t)K UB (j, t)
299
The second grouping can be rewritten as:
Xh
(−λ2 (j, 1) + λ3 (j, 1))pj (0)
j∈J
T
X i
+ (−λ2 (j, t)pj (t − 1) + λ3 (j, t))pj (t − 1)
t=2
and hence as:
X
pj (0)(−λ2 (j, 1) + λ3 (j, 1)) (A.4)
j∈J
−1
XhTX i
+ (−λ2 (j, t + 1) + λ3 (j, t + 1))pj (t)
j∈J t=1
Noting that the first term of (A.4) is constant with respect to the
variables pj (t), let us substitute (A.4) for the second grouping in (A.3)
to obtain the following expression for the Lagrangian:
L(pj (t), β(t), λi (j, t)) = (A.5)

X Xh
cj pj (t) − β(t)pj (t) + λ2 (j, t)pj (t) − λ3 (j, t)pj (t)
t∈T j∈J
i
−λ4 (j, t)pj (t) + λ5 (j, t)pj (t)
X
+ pj (0)(−λ2 (j, 1) + λ3 (j, 1))
j∈J
−1
XhTX i
+ (−λ2 (j, t + 1) + λ3 (j, t + 1))pj (t)
j∈J t=1

X X
+ β(t)d(t) + −λ2 (j, t)K U (j, t) − λ3 (j, t)K D (j, t)
t∈T j∈J
i
Regrouping the terms in A.5, we obtain:
L(pj (t), β(t), λi (j, t)) = (A.6)

−1
XhTX i
(cj − β(t) + λ2 (j, t) − λ3 (j, t) − λ4 (j, t) + λ5 (j, t))pj (t)
j∈J t=1
Xh i
+ (cj − β(T ) + λ2 (j, T ) − λ3 (j, T ) − λ4 (j, T ) + λ5 (j, T ))pj (T )
j∈J
−1
XhTX i
+ (−λ2 (j, t + 1) + λ3 (j, t + 1))pj (t)
j∈J t=1
X
+ pj (0)(−λ2 (j, 1) + λ3 (j, 1))
j∈J

X X
+ β(t)d(t) + −λ2 (j, t)K U (j, t) − λ3 (j, t)K D (j, t)
t∈T j∈J
i
+λ4 (j, t)K LB (j, t) − λ5 (j, t)K UB (j, t) (A.7)
Hence,
L(pj (t), β(t), λi (j, t)) = (A.8)

"T −1
X X
t=1 j∈J
#

−λ2 (j, t + 1) + λ3 (j, t + 1) pj (t)
" #
X
+ cj − β(T ) + λ2 (j, T ) − λ3 (j, T ) − λ4 (j, T ) + λ5 (j, T ) pj (T )
j∈J
"
X X
+ pj (0)(−λ2 (j, 1) + λ3 (j, 1)) + β(t)d(t)
j∈J t∈T
XX
t∈T j∈J
#

UB
−λ5 (j, t)K (j, t) (A.9)
301
The point of these manipulations is that we can now explicitly

rewrite the dual (A.2) without the inner minimization by the following
reasoning:
• The inner problem of (A.2) minimizes (A.8) over the variables pj (t),
with the variables β and λ taken as fixed (by the outer problem).
• Because the inner problem has no explicit constraints on the variables
pj (t), it follows that if the coefficient of pj (t) is nonzero, i.e., the
following expression is nonzero:
cj −β(t)+λ2 (j, t)−λ3 (j, t)−λ4 (j, t)+λ5 (j, t)−λ2 (j, t+1)+λ3 (j, t+1)
then the inner (minimization) problem can attain an objective value
arbitrarily close to −∞ by setting pj (t) arbitrarily large or arbitrarily
small, depending on the sign of its coefficient.
• However, the outer problem is a maximization, so it wants to prevent
the inner problem from doing this. Therefore, the outer problem is
only interested in choices of β and λ such that the above coefficient is
zero, thus preventing the inner problem from having objective values
arbitrarily close to −∞.
Hence, we can add to the outer problem the following hidden con-
straints:
cj −β(t)+λ2 (j, t)−λ3 (j, t)−λ4 (j, t)+λ5 (j, t)−λ2 (j, t+1)+λ3 (j, t+1) = 0
for j ∈ J and t = 1, . . . , T − 1, and
cj − β(T ) + λ2 (j, T ) − λ3 (j, T ) − λ4 (j, T ) + λ5 (j, T ) = 0
for j ∈ J.
Under these constraints, it follows from (A.8) that the objective
function of the inner problem becomes constant with respect to the
variables pj (t), and hence its optimal value is that constant value:
X X
pj (0)(−λ2 (j, 1) + λ3 (j, 1)) + β(t)d(t)
j∈J t∈T
XX
t∈T j∈J

−λ5 (j, t)K UB (j, t)
Thus, we have obtained the dual of the third-level problem:

X X
max pj (0)(−λ2 (j, 1) + λ3 (j, 1)) + β(t)d(t)
λi (j,t),β(t)
j∈J t∈T
XX
+ −λ2 (j, t)K (j, t) − λ3 (j, t)K D (j, t)
U
t∈T j∈J

s.t.
− λ2 (j, t + 1) + λ3 (j, t + 1) = 0, ∀j ∈ J, t = 1, . . . , T − 1,
λi (j, t) ≥ 0, i = 2, 3, 4, 5,
β(t) free
B
Notation
B.1 Sets
J Generating units.
T Time periods.
Ξ Optimization variables.
N Nodes in the electric network.
ΛLn Nodes directly connected to node n.
ΛG
n Generating units located at node n.
ΛD
n Demands located at node n.
R Reserve areas.
ΩG
r Generating units in reserve area r.
ΩD
r Demands in reserve area r.
Ω Scenarios.
303
304 Notation
B.2 Constants
Πj (t) Feasible region for production schedules for generating unit j in

time period t.
cU
j (t) Startup cost of unit j in time period t ($).
ccj , hcj , tcold

j Coefficients for the startup cost function for unit j.
Dj Number of hours that unit j is required to be off at the start of the

planning period (h).
D(t) Demand in time period t (MW).
Dexp (t) Expected net demand in time period t (MW).
Di (t) Active power load of demand i in time period t (MW).
TjD Minimum number of time periods required for unit j to be off

before it can be turned on (h).
Fj Equal to Dj unless Dj exceeds the planning horizon, in which case

Fj equals |T |.
Kjk Cost of turning on unit j after it has been inactive for k time
periods ($).
Lj Equal to Uj unless Uj exceeds the planning horizon, in which case

Lj equals |T |.
P j Maximum power output of unit j (MW).
P j Minimum power output of unit j (MW).
Qi (t) Reactive power load of demand i in time period t (VAr).
R(t) Spinning reserve in time period t (MW).
RjD Maximum ramp-down rate of unit j (MW/h).
RjU Maximum ramp-up rate of unit j (MW/h).
SjD Maximum shutdown rate of unit j (MW).

B.2. Constants 305
SjU Maximum startup rate of unit j (MW).
XjU Maximum increase in the power provided by unit j to transition

to a safe post-contingency state (MW/h).
XjD Maximum decrease in the power provided by unit j to transition

to a safe post-contingency state (MW/h).
S nm Apparent power capacity of line nm (VA).
Uj Number of hours that unit j is required to be on at the start of the

planning period (h).
TjU Minimum number of time periods required for unit j to be on

before it can be turned off (h).
nm Line connecting nodes n and m.
Bnm Susceptance (imaginary part of the admittance) of line nm.
Gnm Real part of the admittance of line nm.
bshunt
nm Half of the shunt susceptance of line nm.
P nm Transmission capacity of line nm.
Rr (t) Reserve required in reserve area r in time period t (MW).
n̂ Node of the reference angle.
Vnmin Minimum voltage magnitude at node n.
Vnmax Maximum voltage magnitude at node n.
qjmin Minimum reactive power output (negative) of unit j (VAr).
qjmax Maximum reactive power output (positive) of unit j (VAr).
ω Scenario index.
πω Probability of occurrence of scenario ω.
MjU Maximum up-reserve provided by unit j (MW).

306 Notation
MjD Maximum down-reserve provided by unit j (MW).
D(t) Minimum demand in time period t (MW).
D(t) Maximum demand in time period t (MW).
RdD Maximum ramp-down rate of net demand (MW/h).
RdU Maximum ramp-up rate of net demand (MW/h).
Γ Uncertainty budget (value is in the interval [0, |T |]).
B.3 Variables
pj (t) Active/real power produced by unit j in time period t (MW).
pj (t) Maximum available power in time period t from unit j (MW).
qj (t) Reactive power produced by generating unit j in time period t

(VAr).
vj (t) On/off status in time period t of unit j:

vj (t) equals 1 if unit j is on in time period t, and 0 if it is off.
yj (t) Startup status in time period t of unit j:

yj (t) equals 1 if unit j starts up in time period t, and 0 otherwise.
zj (t) Shutdown status in time period t of unit j:

zj (t) equals 1 if unit j shuts down in time period t, and 0 other-
wise.
θn (t) Voltage angle of node n.
θnm (t) Difference between the voltage angles at the endpoints of line
nm.
Vn (t) Voltage magnitude of node n.
pcj (t) Post-contingency c power produced in time period t of unit j

(MW).
θnc (t) Post-contingency c voltage angle of node n.

References 307
pjω (t) Active power produced by unit j in time period t under scenario
ω (MW).
U (t) Amount of up-reserve deployed by unit j in time period t under
rjω
scenario ω (MW).
D (t) Amount of down-reserve deployed by unit j in time period t
rjω
under scenario ω (MW).
d(t) Uncertain demand in time period t (MW).

References
M. F. Anjos and J. B. Lasserre, editors. Handbook on Semidefinite, Conic

and Polynomial Optimization. International Series in Operations Research
& Management Science. Springer-Verlag, 2011.
E. Balas and M. Fischetti. Integer (linear) optimization. In T. Terlaky, M. F.
Anjos, and S. Ahmed, editors, Advances and Trends in Optimization with
Engineering Applications, chapter 5, pages 49–63. 2017.
D. Bertsimas and R. Weismantel. Optimization Over Integers, volume 13.
Dynamic Ideas Belmont, 2005.
D. Bertsimas, E. Litvinov, X. A. Sun, J. Zhao, and T. Zheng. Adaptive robust
optimization for the security constrained unit commitment problem. IEEE
Transactions on Power Systems, 28(1):52–63, 2013.
F. Bouffard, F. D. Galiana, and A. J. Conejo. Market-clearing with stochastic
security- Part I: formulation. IEEE Transactions on Power Systems, 20(4):
1818–1826, 2005.
R. H. Byrd, J. Nocedal, and R. A. Waltz. KNITRO: An integrated package
for nonlinear optimization. In Large-scale Nonlinear Optimization, pages
35–59. Springer, 2006.
C. Coffrin and P. Van Hentenryck. A linear-programming approximation of
AC power flows. INFORMS Journal on Computing, 26(4):718–734, 2014.
J. Czyzyk, M. P. Mesnier, and J. J. Moré. The neos server. IEEE Journal on
Computational Science and Engineering, 5(3):68–75, 1998.
308
References 309
E. D. Dolan. The neos server 4.0 administrative guide. Technical Memoran-

dum ANL/MCS-TM-250, Mathematics and Computer Science Division,
Argonne National Laboratory, 2001.
J. Dupačová, N. Gröwe-Kuska, and W. Römisch. Scenario reduction in
stochastic programming. Mathematical Programming, 95(3):493–511, 2003.
Y. Fu, M. Shahidehpour, and Z. Li. Security-constrained unit commitment
with AC constraints. IEEE Transactions on Power Systems, 20(3):1538–
1550, 2005.
L. L. Garver. Power generation scheduling by integer programming-
development of theory. Power Apparatus and Systems, Part III. Trans-
actions of the American Institute of Electrical Engineers, 81(3):730–734,
1962.
A. M. Geoffrion. Generalized Benders decomposition. Journal of Optimization
Theory and Applications, 10(4):237–260, 1972.
W. Gropp and J. J. Moré. Optimization environments and the neos server.
In M. D. Buhman and A. Iserles, editors, Approximation Theory and Op-
timization, pages 167–182. Cambridge University Press, 1997.
H. Heitsch and W. Römisch. Scenario reduction algorithms in stochastic
programming. Computational Optimization and Applications, 24(2-3):187–
206, 2003.
J. M. Morales, A. J. Conejo, and J. Pérez-Ruiz. Economic valuation of reserves
in power systems with high penetration of wind power. IEEE Transactions
on Power Systems, 24(2):900–910, 2009.
G. Morales-España, J. M. Latorre, and A. Ramos. Tight and compact MILP
formulation of start-up and shut-down ramping in unit commitment. IEEE
A. L. Motto, F. D. Galiana, A. J. Conejo, and J. M. Arroyo. Network-
constrained multiperiod auction for a pool-based electricity market. IEEE
R. P. O’Neill, P. M. Sotkiewicz, B. F. Hobbs, M. H. Rothkopf, and W. R.
Stewart. Efficient market-clearing prices in markets with nonconvexities.
European Journal of Operational Research, 164(1):269–285, 2005.
J. Ostrowski, M. F. Anjos, and A. Vannelli. Tight mixed integer linear pro-
gramming formulations for the unit commitment problem. IEEE Transac-
tions on Power Systems, 27(1):39–46, 2012.
M. Tawarmalani and N. V. Sahinidis. A polyhedral branch-and-cut approach
to global optimization. Mathematical Programming, 103:225–249, 2005.
310 References
R. J. Vanderbei. Linear Programming: Foundations and Extensions. Springer,

2001.

Anjos 2017

Uploaded by

Copyright:

Available Formats

You might also like

Anjos 2017

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Anjos 2017

Uploaded by

Copyright:

Available Formats

Foundations and Trends

Unit Commitment in Electric Energy Systems

1 Introduction to the Unit Commitment Problem 221

2 Deterministic Unit Commitment 224

3 Network-Constrained Deterministic Unit Commitment 237

4 Security-Constrained Deterministic Unit Commitment 246

4.2 Recapitulation of the Security-Constrained Deterministic

5 AC Network-Constrained Unit Commitment 256

6 Stochastic Unit Commitment 268

7 Robust Unit Commitment 277

8 Computational Aspects 285

9 Conclusions and Directions for Future Research 295

A Derivation of the Dual of the Third-Level Problem in Section

The unit commitment problem is a fundamental problem in the elec-

M. F. Anjos and A. J. Conejo. Unit Commitment in Electric Energy Systems.

The unit commitment (UC) problem addresses a fundamental decision

the nonconvexity of the UC problem allows the existence of multiple

1.1 Outline of this Book

We introduce in the next six chapters a selection of formulations of

pj (t), pj (t) ∈ Πj (t), ∀j ∈ J, ∀t ∈ T, (2.4)

where each value of t ∈ T is a time period in the planning horizon,

where Cj is the number of linear pieces in the cost function of unit j,

is an equality requiring that the amount of energy generated is precisely

2.1 Binary Decision Variables

We introduce three binary variables per generating unit to model the

To ensure the logical coherence of the values assigned to these vari-

vj (t − 1) − vj (t) + yj (t) − zj (t) = 0 ∀j ∈ J, ∀t ∈ T. (2.5)

2.2 Startup Cost

While the startup cost function is nonlinear, it can be discretized

2.3 Ramping Constraints

The power output of a generating unit cannot change arbitrarily

2.4 Uptime and Downtime Constraints

where Lj = min{|T |, Uj } and Uj is the number of time periods that j

where Fj = min{|T |, Dj } and Dj is the number of time periods that

2.5 Generation Limits

If a unit is on in time period t then its power output must satisfy

2.6 Recapitulation of the Basic UC Formulation

cj (pj (t)) ≥ αjs pj (t) + βjs , s = 1, . . . , Cj , ∀j ∈ J

P j vj (t) ≤ pj (t) ≤ pj (t) ≤ P j vj (t), ∀j ∈ J, ∀t ∈ T

2.7 Illustrative Example

We consider a small example with three generators. For simplicity we

The parameters for the generators are as follows:

2.7.1 Initial Demand

We see that unit 1 is committed throughout, unit 2 is turned on for the

2.7.2 Impact of Operational Constraints

As would be expected, the higher load leads to a significantly higher

The basic UC formulation in Section 2.6 is concerned only with allocat-

3.1 DC Power Flow

Bnm (θn (t) − θm (t)) ,

where ΛLn is the set of nodes directly connected to node n.

3.2 Integration of Network Constraints in UC

To incorporate the network effects, we change the constraints (2.2)

−P nm ≤ Bnm (θn (t) − θm (t)) ≤ P nm , ∀n ∈ N, ∀m ∈ ΛLn , ∀t ∈ T,

where P nm is the transmission capacity of line nm.

θn̂ (t) = 0 ∀t ∈ T, (3.3)

where n̂ is the node of the reference angle.

where R is the set of reserve areas, Rr (t) is the reserve required in

3.3 Recapitulation of the Network-Constrained Determinis-