Unit Commitment, jdm@linux3.ece.iastate.

edu

1.0 Introduction

The problem of unit commitment (UC) is to decide which units to

interconnect over the next T hours, where T is commonly 24 or 48
hours, although it is reasonable to solve UC for a week at a time.
The problem is complicated by the presence of inter-temporal
constraints, i.e., what you do in one period constrains what you can
do in the next period. The problem is also complicated because it
involves integer decision variables, i.e., a unit is either committed
(1) or not (0).

The UC problem forms the basis of today’s day-ahead markets

(DAMs). Most ISOs today are running so-called security-
constrained unit commitment (SCUC) 24 hours ahead of the real-
time (balancing) market.

If one has a very good solution method to solve the UC problem (or
the SCUC problem), then the good solutions that come will save a
lot of money relative to using a not-so-good solution method.
Regardless of the solution method, however, the solutions may not
save much money if the forecast of the demand that needs to be met
contains significant error. Having a “perfect” solution for a
particular demand forecast is not very valuable if the demand
forecast is very wrong. Therefore demand forecasting is very
important for solving the UC. Systems that are expecting high wind
energy penetrations are concerned about this fact, since high wind
penetration increases demand forecast uncertainty (the demand that
the thermal units must meet is load-wind). This is why so much
attention is being paid to improving wind power forecasting. It is
also why so much attention is being paid to creating UC models and
solvers that handle uncertainty.

1
We begin these notes with a motivating example in Section 2.0, then
we provide the explicit problem statement, in words, in Section 3.0.
Section 4.0 provides the analytic problem statement. Section 5.0
provides an overview of solution methods. Section 6.0 ….
Section 7.0 describes the most important solution method – branch
and bound. Section 8.0 illustrates the method on the UC problem.
Section 9.0 provides an overview of the SCUC used by several ISOs
today.

2.0 Motivating example

Assume we are operating a power system that has load characteristic

as given in Fig. 1.

Fig. 1
Consider that we have three units to supply the load. The unit cost
rates are expressed below.
C1( Pg1) = 5 + 0.5Pg1 + Pg21 , 0 ≤ Pg1 ≤ 5
C2 ( Pg 2 ) = 5 + Pg 2 + 0.5Pg22 , 0 ≤ Pg 2 ≤ 10
C3 ( Pg 3 ) = 5 + 3Pg 3 + 2 Pg23 , 0 ≤ Pg 3 ≤ 3
Note that the available capacity is 5+10+3=18.

2
In the economic dispatch problem, we identified the minimum cost
for each hour, under the assumption that all units were connected.

The UC problem differs from the economic dispatch problem in that

we no longer assume that all of the units are connected. In fact, the
essence of the UC problem is to decide which units to connect.

To begin consideration of the problem at hand, let’s make the very

significant assumption that there are no costs associated with a unit
making the transition between up (connected) and down
(disconnected).

Therefore our objective is to determine how to operate the three

units in order to
• achieve the minimum cost over the 24 hour period and
• satisfy the load.
Let’s consider two approaches for doing this.

Approach 1:

In this simple-minded approach, we will connect (commit) all units

for the entire 24 hour period, and dispatch them according to
economic dispatch at each hour.

Observe that this method will certainly satisfy the load. But it does
not achieve minimum cost because, for example, we could simply
run unit 1 by itself from 0 to 6 hours and not incur the automatic
$10/hr required by running units 2 and 3 with Pg2=Pg3=0.

So this is a very poor approach.

Approach 2:

Let’s try to run only the necessary units for each load-level. But we
need to decide which units.

3
To answer this question, let’s consider that there are 7 possible
combinations of units. We will denote each combination as Sk. They
are enumerated below.
S1: G1
S2: G2
S3: G3
S4: G1, G2
S5: G1, G3
S6: G2, G3
S7: G1, G2, G3
However, we observe that unit 3 is very expensive therefore let’s
run this unit only if we must. This means we will eliminate any of
the above combinations that have G3 except the last one. Therefore
we now only have four possibilities:
S1: G1
S2: G2
S3: G1, G2
S4: G1, G2, G3
We desire to determine which combination should be chosen at each
of the various load levels.

To accomplish this, we will plot the total cost of each combination

against total load, assuming the units committed are dispatched
according to economic dispatch (without losses).

So we want to obtain a function CTk(Pd) for each set Sk, k=1,2,3,4.

This is easy for S1 because in this case, Pd=Pg1, and also for S2,
because in this case, Pd=Pg2. Therefore, we have
CT 1( Pd ) = 5 + 0.5Pd + Pd2 , 0 ≤ Pd ≤ 5
CT 2 ( Pd ) = 5 + Pd + 0.5Pd2 , 0 ≤ Pd ≤ 10
For S3 and S4, we have more than one generator, and so how do we
get CT3(Pd)?

4
What we will do here is to write the optimality condition for each
generator, which is
∂Ci
λ=
∂Pgi
We will also use
N
Pd = ∑ Pgi
i =1
where N=2 for S3 and N=3 for S4.

We do it here for S3 and just give the result for S4.

For S3:
CT 3 ( Pd ) = 5 + 0.5Pg1 + Pg21 + 5 + Pg 2 + 0.5Pg22
, (&)
= 10 + 0.5Pg1 + Pg21 + Pg 2 + 0.5Pg22
Using the optimality condition:
∂C ∂C
, λ = ∂P = ∂P
1 2
g1 g2
we can write that
0.5 + 2 Pg1 = 1 + Pg 2 ⇒ Pg 2 = 2 Pg1 − 0.5 (*)
From power balance, we have
Pd = Pg1 + Pg 2 (**)
Substitution of (*) into (**) results in
Pd = Pg1 + 2 Pg1 − 0.5 = 3Pg1 − 0.5
Solving for Pg1 results in
Pd + 0.5
Pg1 = (#)
3
Substitution of (#) into (*) results in

5
 Pd + 0.5 4P − 1
Pg 2 = 2 Pg1 − 0.5 = 2 − 0.5 = d (##)
3 6
Substitution of (#) and (##) into (&) above results in
2 2
⎛ P + 0.5 ⎞ ⎛ Pd + 0.5 ⎞ ⎛ 4 Pd − 1 ⎞ ⎛ 4P − 1 ⎞
CT 3 ( Pd ) = 10 + 0.5⎜ d ⎟+⎜ ⎟ +⎜ ⎟ + 0.5⎜ d ⎟
⎝ 3 ⎠ ⎝ 3 ⎠ ⎝ 6 ⎠ ⎝ 6 ⎠

and the above relation is applicable for 0 ≤ Pd ≤ 15 .

For S4:
We will not go through the detailed algebra here but just give the
result, which is
CT 4 ( Pd ) =
2
⎛ P + 1.125 ⎞ ⎛ Pd + 1.125 ⎞
15 + 0.5⎜ d ⎟+⎜ ⎟ + 0.571Pd + 0.143
⎝ 3 .5 ⎠ ⎝ 3 .5 ⎠
+ 0.5(0.571Pd + 0.143)2 + 3(0.1429Pd − 0.4643) + 2(0.1429Pd − 0.4643)2
and this relation is applicable for 0 ≤ Pd ≤ 18 .

Figure 2 plots CT1, CT2, CT3, and CT4 together as a function of

demand.

6
 Use S2 Use S3 Use S4

Fig. 2

Recall the generators comprising each set, repeated below for

convenience, but now we indicate the load interval for which each
set should be used.

7
 S1: G1 NEVER
S2: G2 0-5
S3: G1, G2 5-13
S4: G1, G2, G3 13-18

Now we return to the load characteristic of Fig. 1 and use Fig. 2 to

identify the “solution” to that particular UC problem. The solution is
given in Fig. 3.

G1,G2, G3

G1,G2

G1,G2
G1,G2
G2 G2

Fig. 3

It is important to note that we have solved this problem under the

assumption that transition costs are zero. What if this is not the
case?

Transition costs include startup costs and shutdown costs. Startup

costs involve both fixed costs and variable costs and requires some
further explanation.

Shutdown costs generally involve only fixed costs (mainly labor)

and are easy to model. Sometimes they are neglected because they
are generally not very significant.

8
The fixed startup costs (generally labor) will be denoted by Cf. The
variable startup costs are denoted by CV.

Variable costs depends on the shut-down state. There are two

possibilities, depending on how “ready” we want the unit to be
during its shutdown period. These two possibilities are
• Hot reserve (banking): This is when the unit is down, but the
boiler is kept hot. The disadvantage of this state is that it costs
money to supply the fuel to heat the boiler. The advantage of this
state is the unit can be started quickly. It is also less expensive to
start a unit from a hot reserve state since no startup fuel is
required to heat the unit. A relation to estimate the variable costs
of a hot reserve state is:
CVb = [Cb t ] f
where
o Cb is the energy per hour to keep the boiler warm
(MBTU/hr)
o f is the cost per MBTU ($/MBTU)
o t is the shutdown duration
Note that Cvb increases with time, without bound. Therefore the
hot reserve state is typically more attractive if the unit will be
down for only a short time.
• Cold reserve (cooling): This is when the unit is down and the
boiler is not heated. The disadvantage of this state is that the
colder the unit is, the more costly and more time to start. The
advantage of this state is that there is no fuel cost while the unit is
down. A relation to estimate the costs of a cold reserve state is
[ ]
CVc = Cc 1 − e −t / α f
where
o Cc is the fuel required to start the unit from completely cold
(MBTU)
o α is unit thermal time constant (time constant of thermal
loss)
o f is the cost per MBTU ($/MBTU)

9
 o t is the shutdown duration
Note from the cold reserve equation that
• t=0 implies CVc=0, meaning the unit is not allowed time to cool.
• t=∞ implies CVc=Ccf, meaning the unit becomes completely cold.

3.0 Node-arc model of the UC problem

You will find in the homework problem 1 that when including start-
up costs, the peaks allow only one solution (S3) but the valleys
allow three which we designate as follows:
S2-H: This is G2 up, with G1 in hot reserve
S2-C: This is G2 up, with G1 in cold reserve
S3: This is G1, G2 up.
You can see that the problem, peaks and valleys, admits only the
above three possible states.

We will use a particular representation using nodes and arcs to

model the situation where
• node: state of the system at the beginning of a period
• arc: possible path from a state in period i to a state in period
i+1.
Notice the use of the term “state” here is to globally specify the
status of all units in the system.

For the homework problem, of our three possible states, only one,
S3, is feasible during the peaks when the demand is 11, but all three
states are feasible during the valleys when the demand is only 3 or 4.
Figures 4 and 5 represents the situation.

10
 Fig. 4

S3

S2-H

S2-C

P1 V1 P2 V2 P3 V3 P4
0 12 18 36 42 58 66 72
4 11 3 11 4 11 4 11

Fig. 5

11
We can also associate a value with each arc as the cost if the system
is in the state from which the arc begins. We can compute these
costs for each state and for each different load level.
I will not provide the expressions to make these computations (you
will need to do that in your homework). Rather, I will just provide
the results in terms of Fig. 6.

74.5 356.8 278.2 356.8 298 356.8 111.8

S3
356.8 315 352 356.8 132
356.8
S2-H
253.5 300 121
356.8 356.8 356.8

S2-C

P1 V1 P2 V2 P3 V3 P4
0 12 18 36 42 58 66 72
4 11 3 11 4 11 4 11

Fig. 6
Observe that the UC problem has been converted to a new
problem…

Assume that the values on the arcs are arc-lengths. Then we desire
to find the shortest path between the first and last node in the
network.

It is easy to see the shortest path in our node-arc model, but consider
a case where we have N units instead of 3.

12
Even if we limit the number of states per unit to two (on or off), then
at each loading level, there are 2N-1 possible states (nodes) to
consider, and we still have not thought about the many different
transitions (arcs). This represents the curse of dimensionality.
If there are m intervals of time for which we must find a solution,
each having 2N-1 possible states, then we will have a total number of
possible solutions equal to (2N-1)m.

For example, consider where N=5 and m=24. In this case

(25-1)24=6.2E35.

Two questions:
1. How do we limit the dimensionality of the problem?
2. How do we algorithmically solve the problem of how to find
the shortest path?

Question 1:
There are two approaches:
A. Limit the number of nodes at each time interval.
B. Limit the number of possible transitions (arc) between time
intervals.

Example:
Consider 3 gens with 2 possible states (nodes): on or off.
The total number of nodes possible at any time interval is 7.

But let’s now prioritize the units using the following rule:
We always turn on unit i before unit i+1. Therefore, we now
only have 3 possible states (nodes), as follows:
S1=G1
S2=G1G2
S3=G1G2G3
In general, this rule creates all states Si=G1G2…Gi

13
The prioritization rule is typically done according to economic
criteria and security criteria.

Note also that this limits the transitions if you can quantify the
maximum possible load variation for one period, see Fig. 7.

If the maximum possible

load change between
intervals is less than X, then
i we need only consider states
in the next interval that are
relevant in considering loads
of Li+1= Li+X

Fig. 7
Question 2: How to algorithmically solve the “shortest path”
problem?

Several alternatives:
• Dijkstra’s algorithm
• Dial’s algorithm
• Label correcting algorithms
• All-pair algorithm
• Dynamic programming: forward and backward

Dynamic programming was used many years ago but not has fallen
out to what are called branch and bound methods, which we will
explore more fully.

14
4.0 Problem statement

The unit commitment problem is solved over a particular time

period T; in the day-ahead market, the time period is usually 24
hours. It is articulated in [10], in words, as follows:

1. Min Objective=UnitEnergyCost+StartupCost+ShutdownCost
+TransactionCost
+VirtualBidCost+ DemandBidCost
+Wheeling Cost
Subject to:
2. Area Constraints:
a. Demand + Net Interchange
b. Spinning and Operating Reserves
3. Zonal Constraints:
a. Spinning and Operating Reserves
4. Security Constraints
5. Unit Constraints:
a. Minimum and Maximum Generation limits
b. Reserve limits
c. Minimum Up/Down times
d. Hours up/down at start of study
e. Must run schedules
f. Pre-scheduled generation schedules
g. Ramp Rates
h. Hot, Intermediate, & Cold startup costs
i. Maximum starts per day and per week
j. Maximum Energy per day and per study length

We describe the objective function and the various constraints in

what follows.

4.1 Objective function

15
a. UnitEnergyCost: This is the total costs of supply over T, based on
the supply offers made, in $/MWhr.
b. StartupCost: This is the total cost of starting units over T, based
on the startup costs
c. ShutdownCost: This is the total cost of shutting down units over
T, based on the shutdown costs.
d. TransactionCost: Transactions are bilateral agreements made
outside the market. Transaction cost for a particular transaction is
the difference between nodal prices of transaction sink and source
nodes, multiplied by the MW value of the transaction. So
TransactionCost is the total transaction costs over T.
e. VirtualBidCost: Purely financial energy bids and offers made to
arbitrage between the day ahead and real time market prices.
f. DemandBidCost: This is the total “cost” of demand over T, based
on the demand bids made, in $/MWhr.
g. WheelingCost: I do not find this defined in the PJM materials but
assume this is the transmission service cost associated with non-firm
transactions.

Revenue from transaction sales, virtual bids and demand bids are
added as negative costs so that by minimizing the objective the
profit is maximized. For Day Ahead studies, this results in a large
negative objective cost.

4.2 Area constraints

a. Demand + Net Interchange: The area demand plus the exports
from the area (which could be negative, or imports).
b. Spinning and Operating Reserves: The spinning reserve is the
amount of generation capacity Σ(Pgmax,k-Pgen,k) in MW that is on-line
and available to produce energy within 10 minutes. Operating
reserve is a broader term: the amounts of generating capacity
scheduled to be available for specified periods of an Operating Day
to ensure the security of the control area. Generally, operating
reserve includes primary (which includes spinning) and secondary
reserve, as shown in Fig. 8.

16
 Fig. 8 [1]

4.3 Zonal constraints

Some regions within the control area, called zones, may also have
spinning and operating reserve constraints, particularly if
transmission interconnecting that region with the rest of the system
is constrained.

4.4 Security constraints

These include constraints on branch flows under the no-contingency

condition and also constraints on branch flows under a specified set
of contingency conditions. The set is normally a subset of all N-1
contingencies.

4.5 Unit constraints

a. Minimum and Maximum Generation limits: Self explanatory.
b. Reserve limits: The spinning, primary, and/or secondary reserves
must exceed some value, or some percentage of the load.
c. Minimum Up/Down times: Units that are committed must remain
committed for a minimum amount of time. Likewise, units that are
de-committed must remain down for a minimum amount of time.
These constraints are due to the fact that thermal units can undergo
only gradual temperature changes.

17
d. Hours up/down at start of study: The problem must begin at some
initial time period, and it will necessarily be the case that all of the
units will have been either up or down for some number of hours at
that initial time period. These hours need to be accounted for to
ensure no unit is switched in violation of its minimum up/down
times constraint.
e. Must run schedules: There are some units that are required to run
at certain times of the day. Such requirements are most often driven
by network security issues, e.g., a unit may be required in order to
supply the reactive needs of the network to avoid voltage instability
in case of a contingency, but other factors can be involved, e.g.,
steam supply requirements of co-generation plants.
f. Pre-scheduled generation schedules: There are some units that are
required to generate certain amounts at certain times of the day. The
simplest example of this is nuclear plants which are usually required
to generate at full load all day. Import, export, and wheel
transactions may also be modeled this way.
g. Ramp Rates: The rate at which a unit may increase or decrease
generation is limited, therefore the generation level in one period is
constrained to the generation level of the previous period plus the
generation change achievable by the ramp rate over the amount of
time in the period.
h. Hot, Intermediate, & Cold startup costs: A certain amount of
energy must be used to bring a thermal plant on-line, and that
amount of energy depends on the existing state of the unit. Possible
states are: hot, intermediate, and cold. Although it costs less to start
a hot unit, it is more expensive to maintain a unit in the hot state.
Likewise, although it costs more to start a cold unit, it is less
expensive to maintain a unit in the cold state. Whether a de-
committed unit should be maintained in the hot, intermediate, or
cold state, depends on the amount of time it will be off-line.
i. Maximum starts per day and per week: Starting a unit requires
people. Depending on the number of people and the number of units
at a plant, the number of times a particular unit may be started in a
day, and/or in a week, is usually limited.

18
j. Maximum Energy per day and per study length: The amount of
energy produced by a thermal plant over a day, or over a certain
study time T, may be less than Pmax×T, due to limitations of other
facilities in the plant besides the electric generator, e.g., the coal
processing facilities. The amount of energy produced by a reservoir
hydro plant over a time period may be similarly constrained due to
the availability of water.

5.0 The UC problem (analytic statement)

The unit commitment problem is a mathematical program

characterized by the following basic features.
• Dynamic: It obtains decisions for a sequence of time periods.
• Inter-temporal constraints: What happens in one time period
affects what happens in another time period. So we may not solve
each time period independent of solutions in other time periods.
• Mixed Integer: Decision variables are of two kinds:
o Integer variables: For example, we must decide whether a unit
will be up (1) or down (0). This is actually a special type of
integer variable in that it is binary.
o Continuous variables: For example, given a unit is up, we must
decide what its generation level should be. This variable may
be any number between the minimum and maximum
generation levels for the unit.

There are many papers that have articulated an analytical statement

of the unit commitment problem, more recent ones include [7, 8, 2,
3], but there are also more dated efforts that pose the problem well,
although the solution method is not as effective as what we have
today, an example is [4].

We provide a mathematical model of the security-constrained unit

commitment problem in what follows. This model was adapted from
the one given in [5, ch 1]. This model is a mixed integer linear
program.

19
 min ∑∑ zit Fi + ∑∑ git Ci + ∑∑ yit Si + ∑∑ xit H i
t 42
1 i 43 t 4i243
1 t 42
1 i 43 t 4
1 i243
Fixed Costs Production Costs Startup Costs Shutdown Costs
(1)
subject to
power balance ∑ git = Dt = ∑ dit ∀ t, (2)
i i

reserve ∑ rit = SDt ∀ t, (3)

i

min generation git ≥ zit MINi ∀ i, t , (4)

max generation git + rit ≤ zit MAX i ∀ i, t , (5)
max spinning reserve rit ≤ zit MAXSPi ∀ i, t , (6)
ramp rate pos limit git ≤ git −1 + MxInci ∀ i, t , (7)
ramp rate neg limit git ≥ git −1 − MxDeci ∀ i, t , (8)
start if off-then-on zit ≤ zit −1 + yit ∀ i, t , (9)
shut if on-then-off zit ≥ zit −1 − xit ∀ i, t , (10)
normal line flow limit ∑ aki ( git − dit ) ≤ MxFlowk ∀ k , t, (11)
i

security line flow limits ∑ aki( j ) ( git − dit ) ≤ MxFlowk( j ) ∀ k , j, t , (12)

i

where the decision variables are:

• git is the MW produced by generator i in period t,
• rit is the MW of spinning reserves from generator i in period t,
• zit is 1 if generator i is dispatched during t, 0 otherwise,
• yit is 1 if generator i starts at beginning of period t, 0 otherwise,
• xit is 1 if generator i shuts at beginning of period t, 0 otherwise,

Other parameters are

• Dt is the total demand in period t,
• SDt is the spinning reserve required in period t,
• Fit is fixed cost ($/period) of operating generator i in period t,
• Cit is prod. cost ($/MW/period) of operating gen i in period t;
• Sit is startup cost ($) of starting gen i in period t.
• Hit is shutdown cost ($) of shuting gen i in period t.
• MxInci is max ramprate (MW/period) for increasing gen i output
• MxDeci is max ramprate (MW/period) for decreasing gen i output

20
• aij is linearized coefficient relating bus i injection to line k flow
• MxFlowk is the maximum MW flow on line k
• aki is linearized coefficient relating bus i injection to line k flow
( j)

under contingency j,
• MxFlowk is the maximum MW flow on line k under contingency j
( j)

The above problem statement is identical to the one given in [5]

with the exception that here, we have added eqs. (11) and (12).
ÎThe addition of eq. (11) alone provides that this problem is a
transmission-constrained unit commitment problem.
Î The addition of eqs. (11) and (12) together provides that this
problem is a security-constrained unit commitment problem.

One should note that our problem is entirely linear in the decision
variables. Therefore this problem is a linear mixed integer program,
and it can be compactly written as
T
min c x
Subject to
Ax ≤ b
We have already had some discussion about solution methods to the
above problem, in that we have illustrated priority list methods and
we have described dynamic programming and Lagrangian
Relaxation. There is one more to discuss. In summary, then, there
have four basic solution methods used in the past few years:
• Priority list methods
• Dynamic programming
• Lagrangian relaxation
• Branch and bound
The last method, branch and bound, is what the industry means
when it says “MIP.” It is useful to understand that the chosen
method can have very large financial implications. This point is
well-made in the chart [6] of Fig. 9.

21
 Fig. 9

6.0 UC and Day-ahead market

The main tool used to implement the day-ahead-markets (DAM) is

the security-constrained unit commitment program, or SCUC. In this
section, we review some basics about the DAM by looking at some
descriptions given by a few industry authors. You are encouraged to
review the papers from which these quotes were taken. Notice that
any references made inside the quotations are given only in the
bibliography of the subject paper and not in the bibliography of
these notes. References made outside of the quotations are given in
the bibliography of these notes.

6.1 Paper by Chow & De Mello:

Reference [7] offers an overall view of the sequence of functions
used by an ISO, as given in Fig. 10. Observe that the “day-ahead

22
scheduling” and the “real time commitment and dispatch” both
utilize the SCUC.

Fig. 10

They state:
“Electricity is a commodity that cannot be effectively stored and the
energy-supplying generators have limits on how quickly they can be
started and ramped up or down. As a result, both the supply and
demand become more inelastic and the electricity market becomes
more volatile and vulnerable as it gets closer to real time [34]. To
achieve a stable margin as well as to maintain the system reliability,
a forward market is needed to provide buyers and sellers the
opportunity to lock in energy prices and quantities and the ISO to
secure adequate resources to meet predicted energy demand well in
advance of real time. Thus architecturally, many ISOs (e.g. PJM,
ISO New England, New York ISO) take a multisettlement approach
for market design….”

“The two main energy markets, each producing a financial

settlement, in a multisettlement system, are the following.
1) DAM: schedules resources and determines the LMPs for the 24 h
of the following day based on offers to sell and bids to purchase
energy from the market participants.
2) Real-time market: optimizes the clearing of bids for energy so
that the real-time system load matching and reliability requirements
are satisfied based on actual system operations. LMPs are computed
for settlement at shorter intervals, such as 5–10 min….”

23
“Fig. 6 shows the timeline of the multiple-settlement systems used
in NYISO, PJM, and ISO-NE, which are typical of those used in
practice. Supply and demand bids are submitted for the DAM,
typically 12–24 h ahead of the real-time operation. Then the day-
ahead energy prices are computed and posted, 6–12 h ahead of real-
time operation….”

“The DAM typically consists of supply and demand bids on an

hourly basis, usually from midnight to the following midnight. The
supply bids include generation supply offers with start-up and no-
load costs, incremental and decremental bids1, and external
transactions schedules. The demand bids are submitted by loads
individually or collectively through load-serving entities. In
scheduling the supply to meet the demand, all the operating
constraints such as transmission network constraints, reserve
requirements, and external transmission limits must not be violated.
This process is commonly referred to as an SCUC problem, which is
to determine hourly commitment schedules with the objective of
minimizing the total cost of energy, start-up, and spinning at no-load
while observing transmission constraints and physical resources’
minimum runtime, minimum downtime, equipment ramp rates, and
energy limits of energy-constrained resources. Based on the
commitment schedules for physical resources, SCUC is used to clear
energy supply offers, demand bids, and transaction schedules, and to
determine LMPs and their components at all defined price nodes
including the hubs, zones, and aggregated price nodes for the DAM
settlement. The SCUC problem is usually optimized using a
Lagrangian relaxation (LR) or a mixed-integer programming (MIP)
solver….”

“A critical part of the DAM is the bid-in loads, which is a day-ahead

forecast of the real-time load. The load estimate depends on the

1
Decremental bids are similar to price-sensitive demand bids. They allow a marketer or other similar entity
without physical demand to place a bid to purchase a certain quantity of energy at a certain location if the
day-ahead price is at or below a certain price. Incremental offers are the flip side of decremental bids.

24
season, day type (weekday, weekend, holiday), and hour of the day.
Most ISOs have sophisticated load forecasting programs, some with
neural network components [36], [37], to predict the day-ahead load
to within 3%–5% accuracy and the load forecasts are posted. LSEs
with fully hedged loads through long-term bilateral contracts tend to
bid in the amount corresponding to the ISO predicted loads. Some
other LSEs may bid in loads that are different from those posted by
the ISO. In such cases, if the LSE bid load exceeds the ISO load, the
LSE bid load is taken as the load to be dispatched. Otherwise, the
ISO load will supersede the LSE bid load and the SCUC will
commit generators to supply the ISO forecasted load in a reliability
stage. Then the generation levels of the committed generators will
be allocated to supply LSE bid loads. Committing extra generators
outside the DAM will be treated as uplifts and be paid by the
LSEs….”

6.2 Paper by Papalexopoulos:

Reference [8] states:
“The Must Offer Waiver (MOW) process is basically a process of
determining which Must Offer units should be committed in order to
have enough additional capacity to meet the system energy net short
which is the difference between the forecast system load and the
Day-Ahead Market energy schedules. This commitment process
ensures that the resulting unit schedule is feasible with respect to
network and system resource constraints. Mathematically, this can
be stated as a type of a SCUC problem [3]. The objective is to
minimize the total start up and minimum load costs of the
committed units while satisfying the power balance constraint, the
transmission interface constraints, and the system resource
constraints, including unit inter-temporal constraints….”

“The most popular algorithms for the solutions of the unit

commitment problems are Priority-List schemes [4], Dynamic
Programming [5], and Mixed Integer Linear Programming [6].
Among these approaches the MILP technique has achieved

25
significant progress in the recent years [7]. The MILP methodology
has been applied to the SCUC formulation to solve this MOW
problem. Recent developments in the implementation of MILP-
based algorithms and careful attention to the specific problem
formulation have made it possible to meet accuracy and
performance requirements for solving such large scale problems in a
practical competitive energy market environment. In this section the
MILP-based SCUC formulation is presented in detail….”

6.3 Paper by Ott:

Reference [9] states:

“In addition to the LMP concept, the fundamental design objectives
of the PJM day-ahead energy market are: 1) to provide a mechanism
in which all participants have the opportunity to lock in day-ahead
financial schedules for energy and transmission; 2) to coordinate the
day-ahead financial schedules with system reliability requirements;
3) to provide incentive for resources and demand to submit day-
ahead schedules; and 4) to provide incentive for resources to follow
real-time dispatch instructions….”

6.4 Paper by AREVA and PJM:

Reference [10] states:

“As the operator of the world’s largest wholesale market for
electricity, PJM must ensure that market-priced electricity flows
reliably, securely and cost-effectively from more than 1100
Generating resources to serve a peak load in excess of 100,000 MW.
In doing so, PJM must balance the market’s needs with thousands of
reliability-based constraints and conditions before it can schedule
and commit units to generate power the next day. The PJM market
design is based on the Two Settlement concept [4]. The Two-
Settlement System provides a Day-ahead forward market and a real-
time balancing market for use by PJM market participants to
schedule energy purchases, energy sales and bilateral contracts. Unit

26
commitment software is used to perform optimal resource
scheduling in both the Day-ahead market and in the subsequent
Reliability Analysis….”

“As the market was projected to more than double its original size,
PJM identified the need to develop a more robust approach for
solving the unit commitment problem. The LR algorithm was
adequate for the original market size, but as the market size
increased, PJM desired an approach that had more flexibility in
modeling transmission constraints. In addition, PJM has seen an
increasing need to model Combined-cycle plant operation more
accurately. While these enhancements present a challenge to the LR
formulation, the use of a MIP formulation provides much more
flexibility. For these reasons, PJM began discussion with its
software vendors, in late 2002, concerning the need to develop a
production grade MIP-based approach for large-scale unit
commitment problems….”

“The Day-ahead market clearing problem includes next-day

generation offers, demand bids, virtual bids and offers, and bilateral
transactions schedules. The objective of the problem is to minimize
costs subject to system constraints. The Day-ahead market is a
financial market that provides participants an operating plan with
known compensation: If their generation (or load) is the same in the
real-time market, their revenue (or cost) is the same. Compensation
for any real-time deviations is based on real-time prices, providing
participants with opportunities to improve profit (or reduce cost) if
they have flexibility to adjust their schedules….”

“In both problems, unit commitment accepts data that define bids
(e.g., generator constraints, generator costs, and costs for other
resources) and the physical system (e.g., load forecast, reserve
requirements, security constraints). In real time, the limited
responsiveness of units and additional physical data (e.g., state

27
estimator solution, net-interchange forecast) further constrains the
unit commitment problem.”

“The Unit Commitment problem is a large-scale non-linear mixed

integer programming problem. Integer variables are required for
modeling: 1) Generator hourly On/Off-line status, 2) generator
Startups/Shutdowns, 3) conditional startup costs (hot, intermediate
& cold). Due to the large number of integer variables in this
problem, it has long been viewed as an intractable optimization
problem. Most existing solution methods make use of simplifying
assumptions to reduce the dimensionality of the problem and the
number of combinations that need to be evaluated. Examples
include priority-based methods, decomposition schemes (LR) and
stochastic (genetic) methods. While many of these schemes have
worked well in the past, there is an increasing need to solve larger
(RTO-size) problems with more complex (e.g. security) constraints,
to a greater degree of accuracy. Over the last several years, the
number of units being scheduled by RTOs has increased
dramatically. PJM started with about 500 units a few years ago, and
is now clearing over 1100 each day. MISO cases will be larger
still….”

“The classical MIP implementation utilizes a Branch and Bound

scheme. This method attempts to perform an implicit enumeration of
all combinations of integer variables to locate the optimal solution.
In theory, the MIP is the only method that can make this claim. It
can, in fact, solve non-convex problems with multiple local minima.
Since the MIP methods utilize multiple Linear Programming (LP)
executions, they have benefited from recent advances in both
computer hardware and software [6]…”

“This section presents results from using the CPLEX 7.1 and
CPLEX 9.0 MIP solvers on a large-scale RTO Day Ahead Unit
Commitment problem. This problem has 593 units and a 48 hour
time horizon….”

28
7.0 Solution methods for mixed integer programs

To begin with, we need to discuss a very fundamental principle

related to linear programming. Consider a standard linear program,
given below. Notice that variables are continuous, not integer.
max Z = x1 + 5 x2
s.t.
x1 + 10 x2 ≤ 20
x1 ≤ 2
x1 ≥ 0, x2 ≥ 0
We can use Matlab-linprog or CPLEX to solve this, but it is so
simple that you we can also visualize the solution in x1-x2 space, as
in Fig. 11 below. The shaded area indicates the feasible region.

Observe the following in Fig. 11:

• The parallel lines correspond to different values of the
objective function.
• The objective function is increasing from bottom to top, as
indicated by the heavy arrow. The parallel lines are contours of
constant Z. For example, note there are two dark dotted
parallel lines. The lower one corresponds to Z=5. The upper
one corresponds to Z=10.
• The solid dark parallel line is the highest Z-contour (parallel
line) which touches one feasible point.
• Unless the Z-contours are parallel to a binding constraint, the
highest Z-contour which touches only one feasible point will
always do so at a corner point. A corner point is a point of
intersection between two or more constraints.

29
 ↑
x2
Solution is where the
3 Increasing Z highest Z-contour
touches a feasible point.

Z=10=x1+5x2
x1+ 10x2<20
1
Z=5=x1+5x2

x1 →
1 2 3

Fig. 11

7.1 Integer programs

In Sections 5 and 6, we observed that our problem is a linear mixed

integer programming problem. Before we discuss our chosen
method of solving this problem, it is useful to see where our
problem lies in the general area of optimization.

Linear mixed integer programs are a particular type of problem that

falls under the more general heading of integer programs (IPs). IPs
may be pure, in which case all the decision variables are integer
(PIP), or they may be mixed, in which case some decision variables
are integer and some are continuous valued (MIP).

It is also possible to have problems where the integer variables may

take on one of just two values: 0 and 1. Such problems occur
frequently when the decisions to be made are of the “yes or no”
type. When all integer variables are this way, the problem is
considered to be a binary integer program (BIP).

30
The remaining material in this section is adapted from [11, ch 13.3].

There are two obvious approaches that come to mind for solving
IPs. One is to check every possible solution. We call this exhaustive
enumeration. Another is to solve the problem as a linear program
without integrality requirements, and then round the values we get
to the nearest integer. We call this LP-relaxation with rounding.
Let’s look at these two approaches.

7.1.1 Exhaustive enumeration

Consider a BIP problem with 3 variables: x1, x2, and x3, each of
which can be 1 or 0. The possible solutions are
(0,0,0), (0,0,1), (0,1,0), (0,1,1),
(1,0,0), (1,0,1), (1,1,0), (1,1,1).
3
Thus, there are 2 =8 solutions. It would not be too hard to check
them all. But consider more typical problems with 30 variables, or
even 300.
230=1.0737×109=1,073,700,000 (over a billion possible solutions)
2300=2.037×1090

Recall our UC formulation in the previous notes. There were three

integer variables for every generator:
• zit is 1 if generator i is dispatched during t, 0 otherwise,
• yit is 1 if generator i starts at beginning of period t, 0 otherwise,
• xit is 1 if generator i shuts at beginning of period t, 0 otherwise,
PJM, for example, has over 1100 generators. For PJM’s unit
commitment problem, the number of integer decision variables is
over 3300. We see there are over 23300 possible solutions! It is for
this reason, as we have seen in Section 3, that the UC problem is
said to suffer from the “curse of dimensionality.”

The conclusion is that exhaustive enumeration is simply not

feasible, even for the fastest computers.

31
7.1.2 LP relaxation with rounding

Here, we relax the requirement that the decision variables be

integers. Then our problem becomes a standard LP, and we can
solve it efficiently using the simplex method. When we examine the
resulting solution, some or all of the variables are likely to be non-
integer. These variables are then just rounded to the nearest feasible
integer, and we call what results our solution.

This approach may in fact work with reasonable accuracy (it may
get close to the optimum) for some problems, especially if the
values of the variables are large so that rounding creates relatively
little error. (This would not be the case, however, for BIPs, as in the
case of the UC, where integer variables are either 1 or 0.)

But there are two pitfalls in this approach.

Pitfall #1: The solution obtained by rounding the optimal LP

solution is not necessarily feasible. To illustrate, consider a problem
having the below two constraints.
− x1 + x2 ≤ 3.5
x1 + x2 ≤ 16.5
Assume the LP-relaxation found a solution that has x1=6.5 and
x2=10, so that
− 6.5 + 10 = 3.5 ≤ 3.5
6.5 + 10 = 16.5 ≤ 16.5
If we round x1 to 7, then
− 7 + 10 = 3 ≤ 3.5
7 + 10 = 17 not ≤ 16.5
If we round x1 to 6, then
− 6 + 10 = 4 not ≤ 3.5
6 + 10 = 16 ≤ 16.5

32
Thus, either way we go, rounding up or down, we result in an
infeasible solution. The only way we can make x1 an integer is if we
also change x2. This situation is illustrated using Fig. 12 below.

Fig. 12
In Fig. 12, the feasible region is between the x-axis and the two
constraint-lines. The “corner point” of (6.5, 10) is the solution to the
relaxed IP. When x1 is rounded to 7, the solution is X in Fig. 12,
which is clearly above the feasible region. When x1 is rounded to 6,
the solution is O in Fig. 12, again, clearly above the feasible region.

Pitfall #2: Even if the rounded solution is feasible, there is no

guarantee it will be optimal or that it will even be reasonably
accurate (reasonably close to the optimal). This is illustrated by the
following problem.

33
 max Z = x1 + 5 x2
s.t.
x1 + 10 x2 ≤ 20
x1 ≤ 2
x1 ≥ 0, x2 ≥ 0
x1 , x2 integers

This problem is illustrated in Fig. 12.

↑
x2 Actual integer LP-relaxed
solution, solution,
Z*=10 Z*=11
3 ● ● ● ●

2 ● ● ● ●
Z*=10=x1+5x2

1 ● ● ● ●
LP-relaxed,
rounded solution
Z*=7
● ● ● ●
x1 →
1 2 3

Fig. 12

In Fig. 12, the dots are the possible integer solutions, and the shaded
region is the feasible region. The LP-relaxed solution is (2, 9/5)
where Z*=11. If we rounded, then we would get (2, 1) where Z*=7.
But it is easy to see that the point (0, 2) is on the Z*=10 line.
Because (2,0) is integer and feasible, and higher than Z*=7, we see
that the rounding approach has failed us miserably.

34
7.2 Some other methods
There are two broad classes of methods to solving IPs:
• Cutting plane methods
• Tree-search methods

7.2.1 Cutting plane methods

Cutting plane methods generate additional constraints. There are
several kinds of cutting plane methods. One of the simplest to
understand is the fractional integer programming algorithm [12, ch
13]. In this method, we shrink the feasible region the minimum
amount possible so that all corner points are integer. If we are
successful in doing this, then the solution to the corresponding
relaxed LP will be integer and thus the solution to the original
integer program. Figure 13 illustrates.
↑
x2 Original
constraint
3 ● ● ● ●

2 ● ● ● ●
Shrunk
feasible
region
1 ● ● ● ●

● ● ● ●
x1 →
1 2 3

Fig. 13
Another very popular cutting plane method, particularly for MIP, is
Benders decomposition [12, Sec 15.2]. To apply this method, the
variables must be separable. If they are, one can set up a Master
problem to solve for one set of variables and subproblems to solve
for the other set. Then the algorithm iterates between Master and
subproblems.

35
7.2.2 Tree-search methods

The essence of tree-search algorithms are that they conceptualize the

problem as a huge tree of solutions, and then they try to do some
smart things to avoid searching the entire tree. Such a tree is shown
in Fig. 14.
●

● ●

● ● ● ●

● ● ● ● ● ● ● ●

Fig. 14
The common features of tree-search algorithms are [12, App C]:
1. They are easy to understand;
2. They are easy to program on a computer;
3. The upper bound on the number of steps the algorithm needs to
find the solution is O(kn), where n is the number of decision
variables (this means running time increases exponentially with
the number of variables);
4. They lack mathematical structure.

The most popular tree-search method today is called branch and

bound. We will study this method in the next section.

7.3 Branch and bound (B&B)

Two definitions are necessary.

• Predecessor: Pj
• Successor: Pk
Problem Pj is predecessor to Problem Pk, and Problem Pk is
successor to Problem Pj, if they are identical with the exception that

36
one continuous-valued variable in Problem Pj is constrained to be
integer in Problem Pk.

How to constrain a continuous-valued variable to be integer?

Consider the following problem [13, ch. 23]. We will call it P0.
max ζ = 17 x1 + 12 x2
s.t.
P0 10 x1 + 7 x2 ≤ 40
x1 + x2 ≤ 5
x1 , x2 ≥ 0
Solution using CPLEX yields
P0 Solution : x1 = 1.667, x2 = 3.333, ζ = 68.333
Observe this solution as the corner point labeled “solution” in Fig.
15.

Fig. 15

37
What if we now pose a problem P1 to be exactly like P0 except that
we will constrain x1≤1? Here it is:
max ζ = 17 x1 + 12 x2
s.t.
10 x1 + 7 x2 ≤ 40
P1
x1 + x2 ≤ 5
x1 ≤ 1
x1 , x2 ≥ 0
What do you expect the value of x1 to be in the optimal solution?

Because the solution without the constraint x1≤1 wanted 1.667 of x1,
we can be sure that the solution with the constraint x1≤1 will want as
much of x1 as it can get, i.e., it will want x1=1.

But now let’s ask, in terms of constraining a continuous-valued

variable to be integer: what have we done here?
1. We solved a predecessor problem as a relaxed LP and obtained
optimal values (but non-integer) for the variables.
2. Then we constructed a successor problem by indirectly imposing
integrality on one variable via constraining it to be less than or
equal to the integer just less than the value of that variable in the
optimal solution to the predecessor problem.
Let’s use CPLEX to solve P1 to see if it works. The CPLEX solution
is:
P1 Solution : x1 = 1.0, x2 = 4.0, ζ = 65.0
It worked in that x1 did in fact become integer. In fact, x2 became
integer as well, but this is by coincidence, i.e., in general, the “trick”
of imposing integrality on a variable, as we have done above, is not
guaranteed to also impose integrality on the remaining variables.

A related way to think of this is that we just made a new corner

point, as illustrated in Fig. 16, that had to be the next “best” (largest

38
ζ in this case) corner point to the optimal without the constraint
x1≤1.

Fig. 16
So we understand now how to constrain a continuous-valued
variable to be integer.

But here is another question for you… What if we want to solve the
following IP:
max ζ = 17 x1 + 12 x2
s.t.
10 x1 + 7 x2 ≤ 40
IP1
x1 + x2 ≤ 5
x1 , x2 ≥ 0
x1 , x2 integers.
This problem is identical to P0 except we require x1, x2 to be integer.
The question is: Is the P1 solution we obtained, which by chance is a
feasible solution to IP1, also the optimal solution to IP1?

39
To answer this question, let’s state a rather obvious criterion.
IP Optimality Criterion: A solution to an IP is optimal if the
corresponding objective value is better than the objective value
corresponding to every other feasible solution to that IP.

So is the P1 solution the optimal solution to IP1?

Answer: We do not know. Why?

Î Because we have not yet explored the entire solution space.

What part of the solution space remains?

Answer: The part associated with x1=2. So let’s constrain x1≥2. This
results in problem P2.
max ζ = 17 x1 + 12 x2
s.t.
10 x1 + 7 x2 ≤ 40
P2
x1 + x2 ≤ 5
x1 ≥ 2
x1 , x2 ≥ 0
Using CPLEX to solve P2 results in:
P2 Solution : x1 = 2.0, x2 = 2.857, ζ = 68.8286
This time, we were not so fortunate to obtain a feasible solution to
IP1, since x2 is not integer. And so the P2 solution is not feasible to
IP1, and therefore it is certainly not optimal to IP1. The situation is
illustrated in Fig. 17.

40
 Fig. 17
But does the P2 solution tell us anything useful?
YES! Compare the objective function value of P2, which is 68.8286,
with the objective function value of P1, which is 65. Since we are
maximizing, the objective function value of P2 is better. But the P2
solution is not feasible. However, we can constrain x2 appropriately
so that we get a feasible solution. Whether such a feasible solution
will have better objective function value we do not know. What we
do know is, because the objective function value of P2 (68.8286) is
better than the objective function value of P1 (65), it is worthwhile to
check it. Although the objective function value of successor
problems to P2 can only get worse (lower), they might be better than
P1, and if we can find a successor (or a successor’s successor,…)
that is feasible, it might be better than our best current feasible
solution, which is P1.

On the other hand, if P2 had resulted in an objective function lower

than that of P1, what would we have done?

Î We would not have evaluated any more successor problems to

P2. Why? Because successor nodes add constraints, and it is
impossible for the objective to get better by adding constraints.
Either the objective will get worse, or it will stay the same. But it

41
will not get better. Therefore if a predecessor node is already not as
good as the current best feasible solution, there is no way a
successor node will be, and we might as well terminate evaluation of
successors to that predecessor node.

But in this case, the objective function value of P2 is better than that
of our current best feasible solution, so let’s solve P2 with the
additional constraint x2≤2. Call this problem P3, given below.
max ζ = 17 x1 + 12 x2
s.t.
10 x1 + 7 x2 ≤ 40
P3 x1 + x2 ≤ 5
x1 ≥ 2
x2 ≤ 2
x1 , x2 ≥ 0
Using CPLEX to solve P3 results in:
P3 Solution : x1 = 2.6, x2 = 2.0, ζ = 68.2
Note that x1 has reverted back to non-integer. We could have
expected this since we forced x2 to change, requiring a new corner
point, as shown in Fig. 18.

42
 Fig. 18

Now we need to force a feasible solution, and we do so by imposing

x1≤2, as indicated in problem P4 below.

max ζ = 17 x1 + 12 x2
s.t.
10 x1 + 7 x2 ≤ 40
P4 x1 + x2 ≤ 5
x1 ≤ 2
x2 ≤ 2
x1 , x2 ≥ 0
Using CPLEX to solve P4, we obtain:
P4 Solution : x1 = 2.0, x2 = 2.0, ζ = 58.0
This solution is displayed in Fig. 19.

43
 Fig. 19
This solution is feasible! Wonderful. However, comparing the
objective value function to our “best so-far” value of 65, we see that
this value, 58, is worse. So the P4 solution is not of interest to us
since we already have one that is better.

What next?
Let’s go back to problem P3 where we had this:
P3 Solution : x1 = 2.6, x2 = 2.0, ζ = 68.2
In P4, we pushed x1 to 2. Let’s explore the space associated with
pushing x1 to 3. So we need a new problem for this, where we
require x1≥3. Let’s call this P5, given below.

44
 max ζ = 17 x1 + 12 x2
s.t.
10 x1 + 7 x2 ≤ 40
P5 x1 + x2 ≤ 5
x1 ≥ 3
x2 ≤ 2
x1 , x2 ≥ 0
Using CPLEX to solve it, we obtain
P5 Solution : x1 = 3.0, x2 = 1.4286, ζ = 68.1429
This solution is not integer, but it has an objective function value
better than our current best feasible solution (65), and so let’s branch
again to a successor node with x2≤1. Let’s call this P6, given below.
max ζ = 17 x1 + 12 x2
s.t.
10 x1 + 7 x2 ≤ 40
P6 x1 + x2 ≤ 5
x1 ≥ 3
x2 ≤ 1
x1 , x2 ≥ 0
Using CPLEX to solve it, we obtain
P6 Solution : x1 = 3.3, x2 = 1.0, ζ = 68.1

Again, the solution is not integer, but it has an objective function

value better than our current best feasible solution (65), and so let’s
branch again to a successor node with x1≤3. Let’s call this P7, given
below:

45
 max ζ = 17 x1 + 12 x2
s.t.
10 x1 + 7 x2 ≤ 40
P7 x1 + x2 ≤ 5
x1 ≤ 3
x2 ≤ 1
x1 , x2 ≥ 0
Using CPLEX to solve it, we obtain
P7 Solution : x1 = 3.0, x2 = 1.0, ζ = 63.0
This solution is feasible! However, comparing the objective value
function to our “best so-far” value of 65, we see that this value, 63,
is worse. So the P7 solution is not of interest to us since we already
have one that is better.

So now what?

Let’s go back to problem P6 where we had this:

P6 Solution : x1 = 3.3, x2 = 1.0, ζ = 68.1
In P7, we pushed x1 to 3. Let’s explore the space associated with
pushing x1 to 4. So we need a new problem for this, where we
require x1≥4. Let’s call this P8, given below.
max ζ = 17 x1 + 12 x2
s.t.
10 x1 + 7 x2 ≤ 40
P8 x1 + x2 ≤ 5
x1 ≥ 4
x2 ≤ 1
x1 , x2 ≥ 0
Using CPLEX to solve it, we obtain

46
P8 Solution : x1 = 4.0, x2 = 0, ζ = 68.0
This solution is feasible! Not only that, but when we compare the
objective value function to our “best so-far” value of 65, we see that
this value, 68, is better. So the P8 solution becomes our new “best
solution.” And we will use 68 as our bound against which we will
test other solutions.

What other solutions do we need to test?

Let’s go back to problem P2 where we had this:

P2 Solution : x1 = 2.0, x2 = 2.857, ζ = 68.8286
In P3, we pushed x2 to 2. Let’s explore the space associated with
pushing x2 to 3. So we need a new problem for this, where we
require x2≥3. Let’s call this P9, given below.
max ζ = 17 x1 + 12 x2
s.t.
10 x1 + 7 x2 ≤ 40
P9 x1 + x2 ≤ 5
x1 ≥ 2
x2 ≥ 3
x1 , x2 ≥ 0
Using CPLEX to solve, we learn that Problem P9 is infeasible.

We have one more branch to check…

Let’s go back to problem P5 where we had this:
P5 Solution : x1 = 3.0, x2 = 1.4286, ζ = 68.1429
In P6, we pushed x2 to 1. Let’s explore the space associated with
pushing x2 to 2. So we need a new problem for this, where we
require x2≥2. Let’s call this P10, given below.

47
 max ζ = 17 x1 + 12 x2
s.t.
10 x1 + 7 x2 ≤ 40
P10 x1 + x2 ≤ 5
x1 ≥ 3
x2 ≥ 2
x1 , x2 ≥ 0
Using CPLEX to solve, we learn that Problem P10 is infeasible.

The tree-search is illustrated in Fig. 20 below.

48
 Fig. 20 [13]

There are two central ideas in the B&B method.

1. Branch: It uses LP-relaxation to decide how to branch. Each
branch will add a constraint to the previous LP-relaxation to
enforce integrality on one variable that was not integer in the
predecessor solution.
2. Bound: It maintains the best integer feasible solution obtained so
far as a bound on tree-paths that should still be searched.
a. If any tree node has an objective value less optimal than the
identified bound, no further searching from that node is
necessary, since adding constraints can never improve an
objective.
b. If any tree node has an objective value more optimal than
the identified bound, then additional searching from that
node is necessary.

In the B&B method, the first thing to do is to solve the problem as

an LP-relaxation. If the solution results so that all integer variables
are indeed integer, the problem is solved.

On the other hand, if the LP-relaxation results in a solution that

contains a particular variable xk=[xk]+sk, where 0<sk<1, and [xk] is
an integer, then we branch to two more linear programs
• The predecessor problem except with the additional constraint:
xk=[xk]
• The predecessor problem except with the additional constraint
xk=[xk]+1

7.4 Other algorithmic issues on B&B

In section 7.3, we identified the essence of the B&B algorithm:

Branch: force integrality on one variable by adding a constraint to

an LP-relaxation.

49
Bound: Continue branching only if the objective function value of
the current solution is better than the objective function value of the
best feasible solution obtained so far.

It is our primary goal to have understood this.

However, you should be aware that there are several other issues
related to B&B that need to be considered before implementation.
Below is a brief summary of two of these issues [14, ch. 6].

7.4.1 Node selection

The issue here is this:

Î Following examination of a node in which we conclude we need
to branch, should we go deeper, or should we go broader?

Consider, in the example of Section 4.0, the situation we were in at

the node corresponding to Problem P3. Reference to Fig. 9 above
will be useful at this point. The solution we obtained was
P3 Solution : x1 = 2.6, x2 = 2.0, ζ = 68.2
And so it was clear to us, because our current best feasible solution
had an objective function value of 65, and 68.2>65, we needed to
branch further on this node.

However, consider P3’s predecessor node, P2. Its solution was

P2 Solution : x1 = 2.0, x2 = 2.857, ζ = 68.8286
To get to P3, we had added the constraint x2≤2. But there was
something else we could have done, and that was added the
constraint x2≥3. This was the other branch. We had to pick one or
the other, and we decided on x2≤2.

But once at P3, after evaluating it, we had a tough decision to make.
• Depth: Do we continue from P3, requiring x1≤2, for example? or
• Breadth: Do we go back to P2 to examine its other branch, x2≥3?

50
For high-dimensional IPs, it is usually the case that feasible
solutions (meaning that all integer variables are integer) are more
likely to occur deep in a tree than at nodes near the root. Finding
multiple feasible solutions early in B&B is important because it
tightens the bound (in the above example, it increases the bound),
and therefore enables termination of branching at more nodes (and
therefore decreases computation). One can see this by considering
the bound before we find a feasible solution: the bound is infinite!
(+∞ for maximization problems and -∞ for minimization problems).
Therefore the best strategy is to go deep first; then after finding
several feasible solutions, explore breadth.

7.4.2 Branching variable selection

In the example of Section 4.0, branching variable selection was not

an issue. At every node, there was only, at most, one non-integer
variable, and so there was no decision to be made. This was because
our example problem had only two variables, x1 and x2.

However, if we consider having any more than 2 decision variables,

even 3, we run into this issue, and that is:
Î Given we are at a node that needs to branch, and there are 2 or
more non-integer variables, which one do we branch on?

This is actually a rich research question that, so far, has not been
solved generically but rather, is addressed for individual problem
types by pre-specifying an ordering of the variables that are required
to be integer. Good orderings become apparent after running the
algorithm many times for many different conditions. Sometimes,
good orderings are apparent based on some physical understanding
of the problem, e.g., perhaps the largest unit should be chosen first.

51
7.4.3 Mixed integer problems

Our example focused on a pure-integer program (PIP). However, the

UC is a mixed integer problem (MIP). Do we need to consider this
more deeply?

Reconsider our example problem, except now consider it as a mixed

integer problem as follows:

max ζ = 17 x1 + 12 x2
s.t.
10 x1 + 7 x2 ≤ 40
IP2
x1 + x2 ≤ 5
x1 , x2 ≥ 0
x1 integer.
Recall our tree-search diagram, repeated here for convenience. You
should be able to see that the solution is obtained as soon as we
solved P1 and P2.

52
 Fig. 21 [13]
The other possible MIP would be to require x2 integer only, as
follows:
max ζ = 17 x1 + 12 x2
s.t.
10 x1 + 7 x2 ≤ 40
IP3
x1 + x2 ≤ 5
x1 , x2 ≥ 0
x2 integer.
Recalling the solution to P0,
P0 Solution : x1 = 1.667, x2 = 3.333, ζ = 68.333

53
we see that we can branch by setting either x2≤3 or x2≥4, resulting in

x2≤3:
Solution : x1 = 1.9, x2 = 3.0, ζ = 68.3

x2≥4:
Solution : x1 = 1.0, x2 = 4.0, ζ = 65

And so we know the solution is the one for x2≤3.

The conclusion here is that we can very easily solve MIP within our
LP-relaxation branch and bound scheme by simply allowing the
non-integer variables to remain relaxed.

7.5 Using CPLEX to solve MIPs directly

Here is the sequence I used to call CPLEX’s MIP-solver to solve our

example problem above.

1. Created the problem statement within a text file called mip.lp as

follows:
maximize
17 x1 + 12 x2
subject to
10 x1 + 7 x2 <= 40
x1 + x2 <= 5
Bounds
0<= x1 <= 1000
0<= x2 <= 1000
Integer
x1 x2
end

3. Used WinSCP to port the file to pluto.

54
4. Used Putty to log on to pluto.
5. Typed cplex101 to call cplex.
6. Typed read mip.lp to read in problem statement.
7. Typed mipopt to call the MIP-solver. The result was as follows:

Tried aggregator 1 time.

MIP Presolve modified 3 coefficients.
Reduced MIP has 2 rows, 2 columns, and 4 nonzeros.
Presolve time = 0.00 sec.
MIP emphasis: balance optimality and feasibility.
Root relaxation solution time = 0.00 sec.

Nodes Cuts/
Node Left Objective IInf Best Integer Best Node ItCnt Gap

0 0 68.3333 2 68.3333 2
* 0+ 0 0 65.0000 68.3333 2 5.13%
* 68.0000 0 68.0000 Cuts: 3 4 0.00%

Mixed integer rounding cuts applied: 1

MIP - Integer optimal solution: Objective = 6.8000000000e+01

Solution time = 0.00 sec. Iterations = 4 Nodes = 0

8. Typed display solution variables - to display solution variables.

The result was:

Variable Name Solution Value

x1 4.000000
All other variables in the range 1-2 are 0.

8.0 UC Solution by mixed-integer programming

Here, we provide some data to use in solving our problem.

We illustrate using an example that utilizes the same system we
have been using in our previous notes, where we had 3 generator
buses in a 4 bus network supplying load at 2 different buses, but this

55
time we will model each generator with the ability to submit 3
offers.
g2
g1

1 2

y12 =-j10
y14 =-j10
y13 =-j10 Pd2 y23 =-j10

y34 =-j10
4 3

g4 Pd3

Fig. 22: One line diagram for example system

The offers, in terms of fixed costs, production costs, and

corresponding min and max generation limits are as follows

Production costs (in $/pu-hr):

Unit, Fixed Startup Shutdown Production Costs ($/pu-hr)
k costs Costs Costs ($) gk1t gk2t gk4t
($/hr) ($)
1 50 100 20 1246 1307 1358
2 50 100 20 1129 1211 1282
4 50 100 20 1183 1254 1320

56
 ⎡0⎤ ⎡ g11t ⎤ ⎡0.50⎤
⎢0⎥ ⎢ g ⎥ ⎢0.60⎥
⎢ ⎥ ⎢ 12t ⎥ ⎢ ⎥
⎢0⎥ ⎢ g13t ⎥ ⎢0.40⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
g
⎢0⎥ ⎢ 21t ⎥ ⎢0.35⎥
⎢0⎥ ≤ ⎢ g 22t ⎥ ≤ ⎢0.60⎥, ∀t
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢0⎥ ⎢ g 23t ⎥ ⎢0.20⎥
⎢0⎥ ⎢ g ⎥ ⎢0.45⎥
⎢ ⎥ ⎢ 41t ⎥ ⎢ ⎥
⎢0⎥ ⎢ g 42t ⎥ ⎢0.50⎥
⎢0⎥ ⎢ g ⎥ ⎢0.40⎥
⎣ ⎦ ⎣ 43t ⎦ ⎣ ⎦
The UC problem is for a 24 hour period, with loading data given as
below. Figure 2, the load curve, illustrates variation of load with
time over the 24 hour period.

57
 Hour, t Load, Dt (pu)
1 1.50
2 1.40
3 1.30
4 1.40
5 1.70
6 2.00
7 2.40
8 2.80
9 3.20
10 3.30
11 3.30
12 3.20
13 3.20
14 3.30
15 3.35
16 3.40
17 3.30
18 3.30
19 3.20
20 2.80
21 2.30
22 2.00
23 1.70
24 1.60
One Day Load variation

400
350
300
Load (MW)

250
200 Series1
150
100
50
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time (hr)

Fig. 23: Load curve

58
 8.1 Example – 4 hours
For this solution, we will only include startup and shutdown
constraints. In order to illustrate all data entered, we will analyze
only the first four hours. The CPLEX code to do this is given below.
minimize
50 z11 +50 z12 + 50 z13 + 50 z14
+50 z21 +50 z22 + 50 z23 +50 z24
+50 z41 +50 z42 + 50 z43 +50 z44
+1246 g111 + 1307 g121 + 1358 g131
+1129 g211 + 1211 g221 + 1282 g231
+1183 g411 + 1254 g421 + 1320 g431
+1246 g112 + 1307 g122 + 1358 g132
+1129 g212 + 1211 g222 + 1282 g232
+1183 g412 + 1254 g422 + 1320 g432
+1246 g113 + 1307 g123 + 1358 g133
+1129 g213 + 1211 g223 + 1282 g233
+1183 g413 + 1254 g423 + 1320 g433
+1246 g114 + 1307 g124 + 1358 g134
+1129 g214 + 1211 g224 + 1282 g234
+1183 g414 + 1254 g424 + 1320 g434
+100 y12 + 100 y13 +100 y14
+100 y22 + 100 y23 +100 y24
+100 y42 + 100 y43 +100 y44
+20 x12 + 20 x13 +20 x14
+20 x22 + 20 x23 + 20 x24
+20 x42 + 20 x43 +20 x44
subject to
loadhr1: g111+g121+g131+g211+g221+g231+g411+g421+g431=1.5
loadhr2: g112+g122+g132+g212+g222+g232+g412+g422+g432=1.4
loadhr3: g113+g123+g133+g213+g223+g233+g413+g423+g433=1.3
loadhr4: g114+g124+g134+g214+g224+g234+g414+g424+g434=1.4
initialu1: z11=0
initialu2: z21=1
initialu4: z41=1
starthr21u1: z12-z11-y12<=0
starthr32u1: z13-z12-y13<=0
starthr43u1: z14-z13-y14<=0
starthr21u2: z22-z21-y22<=0
starthr32u2: z23-z22-y23<=0
starthr43u2: z24-z23-y24<=0
starthr21u4: z42-z41-y42<=0
starthr32u4: z43-z42-y43<=0
starthr43u4: z44-z43-y44<=0
shuthr21u1: z12-z11+x12>=0
shuthr32u1: z13-z12+x13>=0
shuthr43u1: z14-z13+x14>=0
shuthr21u2: z22-z21+x22>=0
shuthr32u2: z23-z22+x23>=0
shuthr43u2: z24-z23+x24>=0
shuthr21u4: z42-z41+x42>=0
shuthr32u4: z43-z42+x43>=0
shuthr43u4: z44-z43+x44>=0

59
g111 - 0.5 z11<= 0
g112 - 0.5 z12<= 0
g113 - 0.5 z13<= 0
g114 - 0.5 z14<= 0
g121 - 0.6 z11<= 0
g122 - 0.6 z12<= 0
g123 - 0.6 z13<= 0
g124 - 0.6 z14<= 0
g131 - 0.4 z11<= 0
g132 - 0.4 z12<= 0
g133 - 0.4 z13<= 0
g134 - 0.4 z14<= 0

g211 - 0.35 z21<= 0

g212 - 0.35 z22<= 0
g213 - 0.35 z23<= 0
g214 - 0.35 z24<= 0
g221 - 0.6 z21<= 0
g222 - 0.6 z22<= 0
g223 - 0.6 z23<= 0
g224 - 0.6 z24<= 0
g231 - 0.2 z21<= 0
g232 - 0.2 z22<= 0
g233 - 0.2 z23<= 0
g234 - 0.2 z24<= 0

g411 - 0.45 z41<= 0

g412 - 0.45 z42<= 0
g413 - 0.45 z43<= 0
g414 - 0.45 z44<= 0
g421 - 0.5 z41<= 0
g422 - 0.5 z42<= 0
g423 - 0.5 z43<= 0
g424 - 0.5 z44<= 0
g431 - 0.4 z41<= 0
g432 - 0.4 z42<= 0
g433 - 0.4 z43<= 0
g434 - 0.4 z44<= 0

60
Bounds
0<= g111
0<= g112
0<= g113
0<= g114
0<= g121
0<= g122
0<= g123
0<= g124
0<= g131
0<= g132
0<= g133
0<= g134

0<= g211
0<= g212
0<= g213
0<= g214
0<= g221
0<= g222
0<= g223
0<= g224
0<= g231
0<= g232
0<= g233
0<= g234

0<= g411
0<= g412
0<= g413
0<= g414
0<= g421
0<= g422
0<= g423
0<= g424
0<= g431
0<= g432
0<= g433
0<= g434
Integer
z11 z12 z13 z14
z21 z22 z23 z24
z41 z42 z43 z44
y12 y13 y14
y22 y23 y24
y42 y43 y44
x12 x13 x14
x22 x23 x24
x42 x43 x44
end

61
Result: CPLEX gives an objective function value of 7020.7 $.

CPLEX> display solution variables -

Variable Name Solution Value
z21 1.000000
z22 1.000000
z23 1.000000
z24 1.000000
z41 1.000000
z42 1.000000
z43 1.000000
z44 1.000000
g211 0.350000
g221 0.600000
g411 0.450000
g421 0.100000
g212 0.350000
g222 0.600000
g412 0.450000
g213 0.350000
g223 0.500000
g413 0.450000
g214 0.350000
g224 0.600000
g414 0.450000
All other variables in the range 1-66 are 0.

Note that all y- and x-variables are zero, therefore there is no

starting up or shutting down.
One should check that the generation in each hour equals the
demand in that hour:
g211+g221+g411+g421=0.35+0.6+0.45+0.1=1.5
g212+g222+g412=0.35+0.6+0.45=1.4
g213+g223+g413=0.35+0.5+0.45=1.3
g214+g224+g414=0.35+0.6+0.45=1.4

62
This very simple solution was obtained as a result of the fact that the
initial solution of
initialu1: z11=0
initialu2: z21=1
initialu4: z41=1
was in fact the best one for the initial loading condition, and since
the loading condition hardly changed during the first four hours,
there was no reason to change any of the units.

Let’s try a different initial condition:

initialu1: z11=1
initialu2: z21=0
initialu4: z41=1

Result: CPLEX gives an objective function value of 7208.9 $.

CPLEX> display solution variables -

Variable Name Solution Value
z11 1.000000
z22 1.000000
z23 1.000000
z24 1.000000
z41 1.000000
z42 1.000000
z43 1.000000
z44 1.000000
g111 0.500000
g121 0.050000
g411 0.450000
g421 0.500000
g212 0.350000
g222 0.600000
g412 0.450000
g213 0.350000
g223 0.500000
g413 0.450000

63
g214 0.350000
g224 0.600000
g414 0.450000
y22 1.000000
x12 1.000000
All other variables in the range 1-66 are 0.
Why was this solution more expensive?

Î Because we initialized the solution with more expensive units, to

get back to the less expensive solution, notice that the program
forces unit 2 to start up (y22=1) and unit 1 to shut down (x12=1) at
the beginning of period 2. Apparently, the additional cost of starting
unit 2 ($100) and shutting down unit 1 ($20) was not enough to
offset the savings associated with running the more efficient unit
over the remaining three hours of the simulation.

Let’s test our theory by increasing the startup costs of unit 2 from
$100 to $10,000. The objective function value in this case is
$7281.25 (higher than the last solution). The decision variables are:
Variable Name Solution Value
z11 1.000000
z12 1.000000
z13 1.000000
z14 1.000000
z41 1.000000
z42 1.000000
z43 1.000000
z44 1.000000
g111 0.500000
g121 0.050000
g411 0.450000
g421 0.500000
g112 0.500000
g412 0.450000
g422 0.450000

64
g113 0.500000
g413 0.450000
g423 0.350000
g114 0.500000
g414 0.450000
g424 0.450000
All other variables in the range 1-66 are 0.

We observe that unit 1 was on-line the entire four hours, i.e, there
was no switching, something we expected since the start-up cost of
unit 2 was so very high.

8.2 Example – 24 hours

We refrain from providing the data in this case because it is

extensive, having 426 variables:
72 z-variables
69 y-variables
69 x-variables
216 g-variables

Rather, we have posted the dataset on the web page under “UC24
Data” for the 04/16/08 date.

The solution was initialized at

initialu1: z11=0
initialu2: z21=1
initialu4: z41=1
which is the most economic solution for this loading level.

The output is most easily analyzed by using

“display solution variables -”
and then searching the output variables for y-variables and/or x-
variables that are listed (and therefore 1). These variables indicate
changes in the unit commitment. In studying the load curve, what
kind of changes do you expect?

65
The result, objective value=$77667.3, shows that the only x and y
variables that are non-zero are y18 and x121. This means that the
changes in the unit commitment occur only for unit 1 and only at
hours 18 and 21. A pictorial representation of the unit commitment
through the 24 hour period is shown below.
One Day Load variation

400
350
300
Load (MW)

250
200
150
100
50
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time (hr)

Unit 1

1.2

0.8
Up or down

0.6 Unit 1

0.4

0.2

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Hour

Unit 2

1.2

0.8
Up or down

0.6 Unit 2

0.4

0.2

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Hour

Unit 3

1.2

0.8
Up or down

0.6 Unit 3

0.4

0.2

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Hour

66
To perform additional investigation, the load curve was modified as
shown below (UC24a.lp). All other data remained as before. The
result, with objective function value of $, shows that the only x and
y variables that are non-zero are y18, x120, and x424. A pictorial
representation of the UC through the 24 hour period is shown below.
One day load variation

400

350

300

250
load

200

150

100

50

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
hour

Unit 1

1.2

0.8

0.6

0.4

0.2

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Unit 2

1.2

0.8

0.6

0.4

0.2

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Unit 3

1.2

0.8

0.6

0.4

0.2

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

67
In a last investigation, the load curve remained modified, and startup
costs were reduced to $10, shutdown costs reduced to $2. All other data
remained as before (UC24b.lp). The result, with objective function value
of $66,867.95, shows that the only x and y variables that are non-0 are
y18, x112, y45, x111, x120, x42, x424. A pictorial representation of the
UC through the 24 hour period is shown below.
One day load variation

400

350

300

250
load

200

150

100

50

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
hour

Unit 1

1.2

0.8

0.6

0.4

0.2

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Unit 2

1.2

0.8

0.6

0.4

0.2

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Unit 3

1.2

0.8

0.6

0.4

0.2

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

68
[1] “PJM Emergency Procedures,”
www.pjm.com/etools/downloads/edart/edart-training-pres/edart-training-
instantaneous-revserve-check.pdf.
[2] H. Pinto, F. Magnago, S. Brignone, O. Alsaç, B. Stott, “Security
Constrained Unit Commitment: Network Modeling and Solution Issues,”
Proc. of the 2006 IEEE PES Power Systems Conference and Exposition, Oct.
29 2006-Nov. 1 2006, pp. 1759 – 1766.
[3] R. Chhetri, B. Venkatesh, E. Hill, “Security Constraints Unit Commitment
for a Multi-Regional Electricity Market,” Proc. of the 2006 Large Engineering
Systems Conference on Power Engineering, July 2006, pp. 47 – 52.
[4] J. Guy, “Security Constrained Unit Commitment,” IEEE Transactions on
Power Apparatus and Systems Vol. PAS-90, Issue 3, May 1971, pp. 1385-
1390.
[5] B. Hobbs, M. Rothkopf, R. O’Neill, and H. Chao, editors, “The Next
Generation of Electric Power Unit Commitment Models,” Kluwer, 2001.
[6] M. Rothleder, presentation to the Harvard Energy Policy Group, Dec 7,
2007.
[7] J. Chow, R. De Mello, K. Cheung, “Electricity Market Design: An
Integrated Approach to Reliability Assurance,” Proceedings of the IEEE, Vol.
93, No. 11, November 2005.
[8] Q. Zhou, D. Lamb, R. Frowd, E. Ledesma, A. Papalexopoulos,
“Minimizing Market Operation Costs Using A Security-Constrained Unit
Commitment Approach,” 2005 IEEE/PES Transmission and Distribution
Conference & Exhibition: Asia and Pacific Dalian, China.
[9] A. Ott, “Experience with PJM Market Operation, System Design, and
Implementation,” IEEE Transactions on Power Systems, Vol. 18, No. 2, May
2003, pp. 528-534.
[10] D. Streiffert, R. Philbrick, and A. Ott, “A Mixed Integer Programming
Solution for Market Clearing and Reliability Analysis,” Power Engineering
Society General Meeting, 2005. IEEE 12-16 June 2005 , pp. 2724 - 2731 Vol.
3.
[11] F. Hillier and G. Lieberman, “Introduction to Operations Research,”
fourth edition, Holden-Day, 1986.
[12] T. Hu, “Integer Programming and Network Flows,” Addison-Wesley,
1970.
[13] R. Vanderbei, “Linear Programming, Foundations and Extensions,” third
edition, Springer, 2008.

69
[14] G. Nemhauser, A. Kan, and M. Todd, editors, “Handbook in Operations
Research and Management Science, Volume 1: Optimization,” North
Holland, Amsterdam, 1989.

70