Thermal Unit Commitment Solution Using an Improved Lagrangian Relaxation

Farid Benhamida, E. N. Abdallah and A. H. Rashed

Department of Electrical Engineering,
Alexandria University
Faculty of Engineering, P.O.BOX 21544, Alexandria (Egypt)

email: [email protected]

Abstract.

An improved Lagrangian relaxation (LR) solution

to the thermal unit commitment problem (UCP) is proposed in
this paper. The algorithm is characterized by: (1) a new Matlab
function to determine the optimal path of the dual problem, (2)
new initialization procedure of Lagrangian multipliers, based on
both unit and time interval classification, (3) a flexible adjustment
of Lagrangian multipliers, and (4) a dynamic search for uncertain
stage scheduling, using a Lagrangian relaxation - dynamic
programming method (LR-DP). After the LR best feasible
solution is reached, and when identical or similar units exist, a
unit decommitment is used to adjust the solution. The proposed
algorithm is tested and compared to conventional Lagrangian
relaxation (LR), genetic algorithm (GA), evolutionary
programming (EP), Lagrangian relaxation and genetic algorithm
(LRGA), and genetic algorithm based on unit characteristic
classification (GAUC) on systems with the number of generating
units in the range of 10 to 100. The total system production cost
of the proposed algorithm is less than the others especially for the
larger number of generating units. Computational time was found
to increase almost linearly with system size, which is favorable
for large-scale implementation.

Key Words
Unit commitment, Generation Scheduling, Lagrangian
Relaxation, Unit classification
List of symbols
Pit: output power of unit i at period t (MW);
Fi (Pit): fuel cost of unit i when its output power is Pit ($);
Sit: start-up price of unit i at period t ($);
uit : commitment state of unit i at period t,(uit =1: unit is
on-line and uit = 0 unit is off-line);
N: total number of generating units;
T: total number of scheduling periods;
ai,, bi,, ci,: Coefficients for the quadratic cost curve of
generating unit i;
Xtoff, i, Xton i: number of hours the unit has been offline/on-line (h);
X i0 : Initial condition of a unit i at t = 0, X i0 > 0: on-line
unit, X i0 < 0: off-line unit (h);
Tiup minimum up time (h);
Tidown minimum down time (h);
HSi, CSi : the units hot/cold startup cost ($);

CHi : is the cold start hour (h);

Dt: customers demand in time interval t.
Rt: The spinning reserve requirements
t , t : Lagrangian multipliers at hour t ($/MWh);
J, Q: primal and dual solution of the LR based UC
algorithm;
: Pre-specified tolerance.

1. Introduction
Unit commitment problem (UCP) is a nonlinear, mixed
integer combinatorial optimization problem. It is defined
as the problem of how to schedule generators
economically in a power system in order to meet the
requirements of load and spinning reserve. Usually this
problem is considered over some period of time, such as
the 24 hours of a day or the 168 hours of a week. It is a
difficult problem to solve in which the solution
procedures involve the economic dispatch problem as a
sub-problem.
Since the problem was introduced, several solution
methods have been developed. However, they differ in
the solution quality, computational efficiency and the size
of the problem they can solve. These methods or
approaches have ranged from highly complex and
theoretically complicated methods to simplified methods.
In the past, various approaches such as DP [1], B&B [2]
and Lagrangian relaxation (LR) [3] were proposed for
solving the UCP. However, not all of these methods are
regarded as feasible and/or practical as the size of the
system increases.
For moderately sized production systems, exact methods,
such as dynamic programming (DP) or branch-and-bound
(B&B) [2] can be used to solve the UCP, successfully.
For larger systems, exact methods fail because the size of
the solution space increases exponentially with the
number of time periods and units in the system. As a
result, the computation time of exact methods becomes
impractical. In these cases heuristic methods
(evolutionary programming (EP), Tabu Search (TS),
Simulated Annealing (SA), Genetic Algorithms (GA),
etc) can be used to produce near optimal solutions in a

reasonable computation time. For heuristic methods

optimality is not given such a high priority but the
emphasis is on finding good solutions in a short time.
This often results in the solution method being more
simple and transparent than exact solution methods [4].
The application of LR in the scheduling of power
generations was proposed in the late 1970s. These earlier
methods used LR to substitute the common linear
programming (LP) relaxation approach as a lower bound
in the B&B technique [5]. In this regard, great
improvement of computational efficiency was achieved
compared with previous B&B algorithms.
In recent years, methods based on LR, have become the
most dominant ones. This approach has shown some
potential in dealing with systems that consist of hundreds
of generating units and is motivated by the separable
nature of the problem, and several examples have been
reported in the literature.
Based on the sharp bound provided by the Lagrangian
dual optimum, it is expected that a suboptimal feasible
solution near the dual optimal point can be accepted as a
proper solution for the primal problem. A more direct and
fairly efficient methodology which has used this idea was
presented in [6] by Merlin, for UCP using LR method
and validated at Electricite De France. Due to its
reasonable performance, the successive improvement of
the LR algorithm, in the last few years, has mainly
followed the work in [6]. The problem which is supposed
to be handled by this algorithm consists of thermal units
only.
In [7], they combined LR, sequential UC based on the
least reserve cost index and unit decommitment (UD)
based on the highest average spinning reserve cost index.
However, this method could not decommit some units
that violate the minimum up time constraints even though
the excessive reserve exists, leading to a higher
production cost.
In the advent of heuristic approaches, GA [8], EP [9], SA
[10], and TS [11] have been proposed to solve the UC
problems. Nevertheless, the obtained results by GA, EP,
and SA required a considerable amount of computational
time especially for a large system size. There was an
attempt to combine the LR and GA method (LRGA) to
obtain a higher quality of UC solution in a shorter time
by using normalized Lagrange multipliers as the encoded
parameter [12].

2. Unit Commitment Problem Formulation

The objective of the UCP is to minimize the system
operating costs, which is the sum of production and startup costs of all units over the entire study time span (e.g.,
24 h), under the generator operational and spinning
reserve constraints. Mathematically, the objective
function, or the total operating cost of the system can be
written as follows:
J = min
f (Pi t ,u it ) =
t t
Pi u i

T N t

min
u [Fi (Pi t ) + S it (1 u it 1 )]
t t i
Pi u i
t =1 i =1

(1)

Subject to:
(1) The start-up cost is modeled by the following function
of the form:
t
down
HS i , if X off
+ CH i
, i T i
t
(2)
Si = i
t
down
+ CH i
CS , if X off , i >T i
(2) Power balance
N

u P
t
i

= Dt

(3)

(3) Spinning reserve requirements:

N

u P
t
i

max

Dt + Rt

(4)

Generating limits:
u it Pi min Pi t u it Pi max

(5)

Minimum up time constraint:

(X ont ,1i T i up )(u it 1 u it ) 0

(6)

X ont , i = (X ont ,1i + 1) u it

(7)

Minimum down time constraint:

t 1
down
(X off
)(u it u it 1 ) 0
, i T i

(8)

t
t 1
t
X off
, i = ( X off , i + 1) (1 u i )

(9)

Fuel cost functions Fi ( Pi t ) is frequently represented by

the polynomial function:
Fi (Pit) = ai + bi Pit + ci (Pit)2

3. An Improved Flexible
Relaxation Technique

(10)

Lagrangian

In the Lagrangian relaxation approach, the system

operating cost function of (1) of the unit-commitment
problem is related to the power balance and the spinning
reserve constraints via two sets of Lagrangian multipliers
to form a Lagrangian dual function.
T
N

L (P , u , , ) = f (P , u ) + t D t u it Pi t
t =1
i =1

T
N

+ t D t + R t u it Pi max
t =1
i =1

(11)

The LR procedure solves the UC problem through the

dual problem optimization procedure attempting to reach
the constrained optimum.
The dual procedure attempts to maximize the Lagrangian
with respect to the Lagrangian multipliers t and t, while
minimizing it with respect to the other variables Pi t , u it
subject to the unit constraints in (5) through (9). The dual

problem is thus the search of the dual solution (Q)

expressed as:
Q = max(min
L (P , u , , )), t 0 and t 0
(12)
t
t
t
t
,

Pi , u i

The Lagrangian function of (11) is rewritten as

T

L (P , u , , ) = f (P , u ) t u it Pi t
t =1

u P
t =1

i =1

t
i

max

i =1

+ D + t (D t + R t )
t

t =1

(13)

t =1

When the Lagrangian multipliers t(k) and t(k) are fixed

for iteration k, the last two terms of the
Lagrangian in (13) are constant and can be dropped from
the
minimization
problem.
Hence, the system (coupling) constrains can be relaxed
and
the
search
for
the
dual
solution can be done through the minimization of the
Lagrangian as:
T

min
L (P , u , ( k ) , ( k ) ) = min
u t {Fi (Pi t ) +
t
t
t
t i
Pi , u i

Pi , u i

(14)

Then, the minimum of the Lagrangian function is solved

for each generating unit over the time horizon, that is
min
L (P ,u ,
t
t

(k )

Pi , u i

(k )

) = min
u
t
t
i =1

Pi , u i

t =1

t
i

{F ( P
i

)+

S it (1 u it 1 )] t ( k ) Pi t t ( k ) Pi max }

(15)

Subject to constraints in (5) through (9).

A. The Dual problem optimization
In the Lagrangian relaxation method, the dual solution is
obtained for each unit separately.
When the state uit = 0, the value of the function to be
minimized is equal zero (the unit is off-line).
When the state uit = 1, the value to be minimized is:
(16)
Fi (Pi t ) t ( k ) Pi t
The startup cost and the last term in (15) are dropped
since the minimization is with respect to Pi t .
When the units fuel cost functions are represented as
polynomial functions as in (10), the minimum of (16) can
be found by taking its first derivative.
d ( Fi (Pi t ) t Pi t ) dPi t = dFi (Pi t ) dPi t t = 0

To minimize the term in (15) for each unit, over the

scheduled time T, subject to minimum up and down time
constraints in (6) through (9) , DP is often used to
determine the optimal schedule. Dynamic programming
CPU time increases at least linearly with N and T (upper
bounded by N [4 (T - 1) + 2] additions and 2 N (T - 1)
comparisons), [3] and [13].
A reduction of the search domain, which is defined by 2T
combinations, can be made by discarding the infeasible
combinations from the domain. The optimal combination
which minimize (15) for a unit i can be determined by
direct evaluation of all feasible combinations. A Matlab
function is developed for this purpose.
This function gives all feasible combinations (mi) of unit i
over the scheduling period T which satisfy the minimum
up and down time constraints given its initial state and
condition.
Function input: X i0 , T, Tiup, Tidown.
Function output: [u it , j ] which is a (T m i ) matrix

t =1 i =1

S it (1 u it 1 ) t ( k ) Pi t t ( k ) Pi max }

1). A new Matlab function to determine the optimal path

(17)

containing all feasible combination as

u i1,1 u i1,2 .... u i1, m i

u 2 u i2,2 .... u i2, m i

t
[u i , j ] = i ,1
M
M
M
T
u T
u
u Ti , m i
...
i ,2
i ,1

u it , j {Fi (Pi t ( k ) ) + S it (1 u it ,j1 ) t ( k ) Pi t ( k ) t ( k ) Pi max }

j = 1,..., m i

t = 1,...,T

(18)

then Pit(k) = Pimin

(19)

j = 1,..., m i

If Pit(k) > Pimax then Pit(k) = Pimax

(20)

= (

If Pit(k) < Pimin

For known t ( k ) , then Pi t ( k ) is obtained by (18) through

(20).

(22)

Step 3: For each ui,j, j = 1,,mi , calculate the

contribution term which correspond to unit i over the
total period T using the following equation:

bi ) / 2c i

Hence, Pi

t (k )

(21)

Then the optimal solution Pi ( k ) and its corresponding

path ui,j (combination), given the Lagrangian multipliers
( k ) , ( k ) of iteration k, is obtained by the following
procedure:
Step 1: Running the Matlab function to obtain all feasible
combinations (mi) of unit i over the scheduling period T
which satisfies the minimum up and down time
constraints given its initial state and condition.
Step 2: For each ui,j, j = 1,,mi, t = 1,,T calculate the
contribution term of unit i in a specific period t using the
following equation

u {F (P

t (k )

, i = 1,..., N ;

t =1

t
i,j

t (k )

) + S it (1 u it ,j1 ) t ( k ) Pi t ( k ) t ( k ) Pi max }

(23)

Step 4: Obtain optimal solution Pi ( k ) and its

corresponding path ui,j (combination), by taking the least
valued contribution terms obtained in step 3. We have not
to check the path vis--vis the minimum up and down
time since it is a feasible one.

Step 5: repeat step 1 to step 4 for all units to

obtain P ( k ) ,u ( k ) .
The
values
of
the
system
variables
P ( k ) , u ( k ) , ( k ) , ( k ) are substituted back into the

Lagrangian (11) , L (P ( k ) , u ( k ) , ( k ) , ( k ) ) to determine the

dual solution Q(k) :
L (P ( k ) , u ( k ) , ( k ) , ( k ) ) = f (Pi t ( k ) , u it ( k ) ) +
T

t (k )

t =1

1). Unit Classification

t
t (k ) t (k )
D u i Pi +
=
i
1

t =1

t (k )

N
t

t
+

D
R
u it ( k ) Pi max

i =1

(24)

Provided that the dual solution is feasible with respect to

the spinning reserve constraint (4) and the following
constraint regarding the minimum output power of the
scheduled units is satisfied:
N

u P
i =1

t
i

min

cost of the schedule to satisfy both the power balance and

spinning reserve constraint, whereas the initial multiplier
t was set to zero which was generally lower than the
optimal value [16].
An initialization procedure which intends to create a high
quality feasible schedule in the first iteration is described
here, based on unit and time interval classification.

Dt

t = 1,...,T

(25)

The inequalities related to the spinning reserve

constraints (4) do not impose an upper bound on the
amount of reserve. However, common sense for an
economic schedule indicates that there should not be too
much excess MW reserve because it would certainly
increase the cost associated with the corresponding dual
solution. Therefore, in the searching algorithm, a slack
term (st) is included in the reserve constraint to assess the
quality of the dual solution. The upper-bound limit
introduced by the slack term restricts the solution space
and therefore may prevent the optimal solution to be
found. In addition, the value of the slack term may affect
the convergence of the process. Unfortunately, there is no
mathematical guideline for properly selecting the value of
slack term (st) [13].
Hence, in the searching algorithm, the following
constraints are included implicitly to test the validity of
the commitment schedule.
N

D t + R t u it Pi max D t + R t + s t

(26)

In this paper st is specified using a new heuristic

algorithm based on both unit and time interval
classification.
B. A new initial scheduling of UC

The initial values of Lagrangian multipliers are very

critical to the LR solution since they may prevent LR
from reaching the optimal solution or require a longer
computational time to reach one [14]. Different initial
values may also lead LR to different solutions. In [15],
the initial multiplier t was set to the hourly system
marginal cost of the schedule to satisfy the power balance
constraint and the initial multiplier t was set to zero,
leading to an infeasible initial solution. Alternatively, the
initial multiplier t was set to the hourly system marginal

In general, generation units can be classified into three

types: base load units with low operation cost Fi, high
startup cost Si, and long minimum up/down
times T i up , T i down ; intermediate load units with medium
operating cost, medium startup cost and medium
minimum up/down time, and peak load units with high
operation cost, low startup cost and short minimum
up/down time. Base load units should not be shut down.
In other words they constitute the must run constraint.
Intermediate load units could be committed during onpeak and decommitted during off-peak periods. Finally,
peak load units could be frequently turned on and off.
Following this classification, the N units of an N-unit
system can be classified into a set Nb of base load units, a
set NI of intermediate load units and a set Np of peak load
units according to unit full load average production costs
(flac) and unit operational constraints where:
Where flac = F (Pi max ) Pi max

(27)

2). Time interval classification

The overall study period is decomposed into several

interval classes as follow:
(1) Tbd presents the set of scheduling intervals t where
t T b and the upper-bound limit of the spinning reserve
is satisfied:

i N b

max

D t R t s t , Tb being the set of

scheduling intervals t where base units can produce

enough
power
to
satisfy
the
inequality Pi max D t + R t . Hence, during the
i N b

intervals (Tbd) only base units are committed.

(2) TId presents the set of scheduling intervals t where
t T I T I 0 , and in which Pi max D t R t s t .
i N b N I

Here TI presents the set of scheduling intervals t where

the group of base and intermediate units cannot produce
enough
power
to
satisfy
the
constraint
max
t
t
max
P

D
+
R
+
min(
P
)
,
while
T
presents
I0
i
i
i N I

i N b U N I

the set of scheduling hour's t where the base and

intermediate units grouped can produce enough power to
satisfy
the
spinning
reserve
max
t
t
constraint Pi D + R . At these intervals (TId)
i N b U N I

both base and intermediate units are committed. Note that

Tb T I 0 .

(3) TI -TId give the set of scheduling hour's t, where peak

units must be committed, for these scheduling periods the
peak units are selected one by one, based on the flac,
until enough capacity is reached to fulfill the spinning
reserve constraints.
(4) TI0-TId-Tbd give the set of scheduling hour's t, where
intermediate units must be committed, for these
scheduling periods intermediate units are selected one by
one, based on the flac, until enough capacity is reached to
fulfill the spinning reserve constraints.
The slack term (st) is defined as
If t T b , s t = max(Pi max )

(28)

If t T I , s t = max(Pi max )

(29)

i Nb
i Np

If t T I 0 and t T b , s = max(Pi
t

max

i N I

(30)

Where is a tuning constant. Through the application of

the method, its most suitable value was found to be 2.
Table 1 gives the initial commitment states of different
sets Nb, NI and Np during different time interval classes.

C. Updating of the Lagrangian Multiplier

In general, adjusting Lagrangian multiplier by subgradient method is not efficient in the presence of the
spinning reserve constraint [6]; one of the shortcomings
of this method is the slow convergence. The LR
performance is heavily dependent on the method used to
update the multipliers. In this paper, a flexible subgradient rule is proposed to update the Lagrangian
multiplier and designed such that the step size is large at
the beginning of iterations and smaller as the iteration
grows. Each nonnegative t and t are adaptively
updated by,

t ( k ) = max 0, t ( k 1) +

t ( k ) = max t ( k 1) +

PMt
( + k ) norm(PM

(32)

SR Mt
, 0 (33)
( + k ) norm(SR M )

Where
N

Table 1. Initial scheduling of UCP based on unit and time

interval classification
Time set
t T b d
t TId

Nb
uit = 1
t

ui = 1

Unit Sets
NI
t

ui = 1

t TI - TId

ui = 1

ui = 1

t TI0-TId-Tbd

uit = 1

uit initialized
besed on flac

Np

ui t = 0
t

ui t = 0
t

ui = 0
uit initialized
besed on flac
uit = 0

3). Initial value of Lagrangian multipliers

The initial value of Lagrangian multipliers t(0) are set as

follow:
(1) For each hour t TI0-TId-Tbd and t TI - TId, the group
of identical units with the least (flac) will be committed
one group by one group until the spinning reserves
constraint is satisfied as shown in Table 1. Subsequently,
economic dispatch in each hour is carried out to obtain
the hourly equal lambda which is initially set to
Lagrangian multipliers t(0).
(2) For each t Tbd TId, as at these period a predefined
UC is established as shown in Table 1. Lagrangian
multipliers t(0) are set to the hourly equal lambda, after
running an economic dispatch program for these periods.
The initial value of each non-negative Lagrangian
multipliers t (0) is set as follow:

t (0) = max( max (

(Fi (Pi t ) +

S it
t (0) Pi t )), 0)
up
Ti

Pi
t = 1,...T
(31)
Where M is the marginal unit with the highest (flac),
giving the sufficient spinning reserve at hour t.
i =1,..., M

max

PMt = D t u it Pi t

(34)

i =1

SR Mt = D t + R t u it Pi max

(35)

norm(PM ) = (PM1 ) 2 + (PM2 ) 2 + ............ + (PMT ) 2

(36)

norm(SR M ) = (SR M1 ) 2 + (SR M2 ) 2 + ... + (SR MT ) 2

(37)

i =1

The values of and are divided into four cases

depending on the signs of PM and SR M .
Case 1) PMt 0 and SR Mt 0 : updating both t and

t by using =0.03 and =0.06.

Case 2) PMt < 0 and SR Mt < 0 : updating both t and

t by using =0.5 and =0.3.

Case 3) PMt < 0 and SR Mt > 0 : updating only t by using
=0.03 and =0.06.
Case 4) PMt > 0 and SR Mt < 0 : updating only t by using
=0.5 and =0.3.
The general guidelines for selecting their values are
explained in [17].
In fact, updating the two multipliers t and t in hour t
must move them in the same direction. In hour t, if
PMt and SR Mt have the same signs, either positive or
negative, t and t will be updated (increase or
decrease) by (32) and (33), respectively.
When the total dual generation output is larger than the
load in that hour ( PMt < 0 ) but the spinning reserve is
insufficient ( SR Mt > 0 ), more committed unit(s) are
required to satisfy the spinning reserve constraints.
However, updating t by (32) will decrease its value,

resulting in committing less units. Therefore, when

( PMt < 0 ) and ( SR Mt > 0 ), only t will be updated.
On the contrary, when the spinning reserve is sufficient
( SR Mt < 0 ), but the total dual generation output is less
than the load in that hour ( SR Mt > 0 ), updating t by
(33) will decrease its value, resulting in committing less
units. Therefore, when ( PMt > 0 ) and ( SR Mt < 0 ), only

t will be updated.
Note that the sub-gradient method generally needs a large
number of iterations to converge to near the dual
optimum [17]. The proposed flexible sub-gradient
method using high-quality initial feasible multipliers
proved to require much lower number of iterations to
converge, leading to much less computational time.
D. Dynamic Economic Dispatch (DED) [18]

To replace conventional economic dispatch algorithm, a

more accurate and flexible problem formulation of DED
is developed to facilitate the interaction with UC
schedule, The DED solver use the Hopfield Neural
Network, which make it a very fast solver and suitable to
UCP.
If the 24-h schedule is feasible at iteration k, a DED is
carried out to determine the optimal generation power
outputs for each of the 24 h, and the total production cost
J(k).
E. Checking for Convergence

The convergence of the proposed LR-UC algorithm can

be measured by the relative duality gap between the
primal and dual solutions.
Relative duality gap =

( (J

(k )

- Q ( k ) )/Q ( k ) ) 100 (38)

The process stops when the relative duality gap is smaller

than a pre-specified tolerance , or when a pre-specified
maximum number of iterations is reached.
The sensitivity of the integer variables corresponding to
the generating unit statuses (uit) to small adjustments in
the Lagrangian multipliers may cause the algorithm to
oscillate around the optimal solution. As such, there is no
guarantee that the solution achieved in the last iteration of
the iterative process will be feasible or optimal. Hence, in
the computational model developed in the paper, a
running record of the feasible solutions is kept so that the
final solution is the one corresponding to the most
economical schedule, i.e., the one with the minimum
primal solution (J).

4. Identical Unit Decommitment

When identical or similar units exist the LR could find
only sub-optimal solutions [14]. These units have the
identical cost parameters ai, bi, ci, and startup cost which
will be simultaneously committed or decommitted. This
will not lead to the optimal solution because committing

one unit at a time will be less expensive than committing

a whole group of units, which may lead to over
commitment. Thus, after committing a group of identicall
units, a unit of which is decommitted one at a time if it
does not violate the minimum up time constraint until
either the spinning reserve requirement is not satisfied or
there is only one unit left. The identical unit
decommitment procedure is as follows:
Step 1) Get the initial feasible solution [u it ] , i = 1,,N,
t = 1,,T.
Step 2) Calculate the excess spinning reserve of every hours,
N

R ext = u it Pi max D t + R t

(39)

Step 3) Initialize t =1
Step 4) Initialize i =1
Step 5) If the excess spinning reserve R ext is greater than
the maximum generation of unit i, and this unit is already
committed, check if decommitting the unit would violate
its minimum up time constraints. Decommit the unit i,
up

If, X ont , i = 1, and X ont +,Tii = T i up +1

or if X

t
on , i

up

>T i , and X

t +1
off , i

(40)

=1

(41)

or if X ont , i = 1, and u ik = T k + 1

(42)

k =t

or if T i up = 1
(43)
Otherwise, let the unit committed.
Step 6) If t = T stop, else go to step 7
Step 7) Update [u it ] and R ext , replace i by i + 1,
Step 8) If i = N, replace t by t + 1, and go to step 4.
Otherwise, go to step 5.

5. Numerical Results
A 10-unit system [8] is selected as a test system. System
data and load demand are given in Tables 2 and 3. The
spinning reserve is assumed to be 10% of the demand.
The 20, 40, 60, 80, and 100 unit systems are obtained by
duplicating the 10-unit base case, whereas the load
demand are adjusted in proportion to the system size. The
proposed LRUC uses the developed Matlab function to
determine the optimal path. A maximum allowable
number of 50 iterations was set as a stopping criteria.
Table 2. Unit data of the 10-unit 24 hour test system
Pmax (MW)
Pmin (MW)
a ($/h)
b ($/MWh)
c ($/MW2h)
Tiup (h)
Tidown (h)
HS
CS
CH
X0 i
flac

Unit 1
455
150
1000
16.19
0.00048
8
8
4500
9000
5
8
18.61

Unit 2
455
150
970
17.26
0.00031
8
8
5000
10000
5
8
19.53

Unit 3
130
20
700
16.60
0.0020
5
5
550
1100
4
-5
22.24

Unit 4
130
20
680
16.50
0.00211
5
5
560
1120
4
-5
22.01

Unit 5
162
25
450
19.70
0.00398
6
6
900
1800
4
-6
23.12

Pmax (MW)
Pmin (MW)
a ($/h)
b ($/MWh)
c ($/MW2h)
Tiup (h)
Tidown (h)
HS
CS
CH
X0 i
flac

Unit 6
80
20
370
22.26
0.00712
3
3
170
340
2
-3
27.45

Unit 7
85
25
480
27.74
0.00079
3
3
260
520
2
-3
33.45

Unit 8
55
10
660
25.92
0.00413
1
1
30
60
0
-1
38.14

Unit 9
55
10
665
27.27
0.00222
1
1
30
60
0
-1
39.48

Unit10
55
10
670
27.79
0.00173
1
1
30
60
0
-1
40.06

Table 3. Demand of 10 unit 24 hour test system

Hour
1
2
3
4
5
6
7
8

Load
(MW)
700
750
850
950
1000
1100
1150
1200

Hour
9
10
11
12
13
14
15
16

Load
(MW)
1300
1400
1450
1500
1400
1300
1200
1050

Hour
17
18
19
20
21
22
23
24

Load
(MW)
1000
1100
1200
1400
1300
1100
900
800

A. An improvement to the method

The behavior of the units during the iterative search of the LR
based solution and the preliminary schedule itself is assessed to
define the uncertain intervals, in which commitment states of
some units are not certain. In this example these stages are [22,
23]. Then, a dynamic search is performed at these stages, using
a DP solution to UC combined with LR (LR-DP), as shown in
figure 2, where all possible and feasible paths with respect to
minimum up and down time constraints are shown. The
optimum path is distinguished by bold lines. The UC solution
schedule using the proposed Lagrangian Relaxation combined
to DP is shown in table 5.
Table 4 shows simulation results (production costs) obtained by
the proposed LR method compared with results obtained by LR
[8], GA [8], EP [19], and the combined LRGA [20] and DPLR
[21] methods. Table 6 shows the simulation time obtained by
the proposed LR method which is carried out on Pentium M
1.73 GHz processor. Because simulations were carried out on
different types of computers, simulation times are not
compared. It can be seen that the results of the proposed method
is better than other methods in term of total production cost. It
can be seen that computational time increases almost linearly
with system size.

6. Conclusion
This paper presents a Lagrangian relaxation solution to the
thermal UCP. An initialization procedure intends to create a
high quality feasible schedule in the first iteration is proposed,
based on unit and time interval classification. The proposed LR
is efficiently and effectively implemented to solve the UC
problem. The proposed LR total production costs over the
scheduled time horizon are less than conventional LR, GA, EP,
LRGA, and GAUC especially for the larger number of
generating units. Moreover, the proposed LR CPU times
increase almost linearly with the system size, which is favorable
for large-scale implementation.

References
[1] W. J. Hobbs, et. al., An enhanced dynamic programming
approach for unit commitment, IEEE Trans. Power Syst.,
Vol. 3, No. 3, pp. 1201-1205, August 1988.
[2] C. L. Chen and S. C. Wang, Branch-and-Bound Scheduling
for Thermal Generating Units, IEEE Trans. on Energy
Conversion, Vol. 8, No. 2, pp. 184-189, June 1993.
[3] F. Zhuang and F.D. Galiana, Towards a more rigorous and
practical unit commitment by Lagrangian relaxation, IEEE
Trans. Power Syst., Vol. 3, No. 2, pp. 763-773, May 1988.
[4] S. Matsuda et. al., The representation of large numbers in
neural networks and its application to economical load
dispatching of electric power, ICNN, Vol. 1, pp. 587-592,
1989.
[5] G. S. Lauer, et, al., Solution of large-scale optimal unit
commitment problems, IEEE Trans. Power Apparatus
Syst. PAS-101 (1, 1982), pp. 79-86.
[6] A. Merlin and P. Sandrin, A New Method for the Unit
Commitment at the Electricit de France, IEEE Trans.
Power Apparatus Syst., Vol. PAS-102, No. 5, pp. 12181225, 1983.
[7] C. Li, R. B. Johnson, A. J. Svoboda, C. Tseng, and E. Hsu,
A robust unit commitment algorithm for hydro-thermal
optimization, IEEE Trans. Power Syst., Vol. 13, No. 3, pp.
1051-1056, Aug. 1998.
[8] S. A. Kazarlis, A. G. Bakirtzis, and V. Petridis, A genetic
algorithm solution to the unit commitment problem, IEEE
Trans. Power Syst., Vol. 11, No. 1, pp. 83-92, Feb. 1996.
[9] P. Attaviriyanupap, H. Kita, E. Tanaka and J. Hasegawa,
Evolutionary Programming based Dynamic Economic
Dispatch, in Joint Conference Record, The Hokkaido
Chapter of The Institute of Electrical and Information
Engineers, Japan, pp. 74-75, 2000.
[10] A. H. Mantawy, et. al., A Simulated Annealing Algorithm
for Unit Commitment, IEEE Trans. Power Syst., Vol. 13,
No. 1, pp. 197-204, Feb. 1998.
[11] A. H. Mantawy et. al., Unit Commitment by Tabu
Search, lEE Proc. -Gener. Distrib.,Vol 145, No. 1, pp. 5664, Jan. 1998.
[12] K. Aoki, T. Satoh, M. Itoh, T. Ichimori, and K. Masegi,
Unit Commitment in a Large-Scale Power System
Including Fuel Constrained Thermal and Pumping- Storage
Hydro, IEEE Trans. Power Syst., Vol. PWRS-2, No. 4, pp.
1077-1084, Nov. 1987.
[13] S. K. Tong and S. M. Shahidehpour, Combination of
lagrangian-relaxation and linear programming approaches
for fuel-constrained unit-commitment problems, lEE
Proceedings-C, Vol. 136, No. 3, pp. 162-174, May 1989.
[14] S. Dekrajangpetch, G. B. Sheble, and A. J. Conejo,
Auction implementation problems using Lagrangian
relaxation, IEEE Trans. Power Syst., Vol. 14, No. 1, pp.
82-88, Feb. 1999.
[15] K. H. Abdul-Rahman, et. al., A practical resource
scheduling with OPF constraints, IEEE Trans. Power Syst.,
Vol 11, No. 1, pp. 254-259, Feb 1995.
[16] H. Yan, P. B. Luh, X. Guan, and P. M. Rogan, Scheduling
of hydrothermal power systems, IEEE Trans. Power Syst.,
Vol. 8, pp. 1358-1365, Aug. 1993.
[17] M. L. Fisher, The Lagrangian Relaxation Method for
Solving integer Programming Problems, Management
Science, Vol. 27, No.1, pp. 1-18, Jan. 1981.
[18] F. Benhamida, E N. Abdallah, A. H. Rashed, Solving the
Dynamic Economic Dispatch as Part of Unit commitment
problem, AMSE, Modeling General Physics and Electrical
Applications, Vol. 78, No. 5, pp. 49-63, 2005.

[19] K. A. Juste, H. Kita, E. Tanaka, and J. Hasegawa, An

evolutionary programming solution to the unit commitment
problem, IEEE Trans. Power Syst., Vol. 14, pp. 1452
1459, Nov. 1999.
[20] C. C. Ping, L. C. Wen, and L. C. Chang, Unit
Commitment by Lagrangian Relaxation and Genetic
Algorithms, IEEE Trans. Power Syst., Vol. 15, No. 2, pp.
707-714, May 2000.

[21] T. Senjyu, H. Yamashiro, K. Uezato, and T. Funabashi, A

unit commitment problem by using genetic algorithm based
on characteristic classification, in Proc. IEEE/Power Eng.
Soc. Winter Meet., Vol. 1, pp. 58-63, 2002.

Table 4. Comparison of total production costs

METHOD
LR [8]
GA [8]
EP [19]
LRGA [20]
DPLR [21]
GAUC [21]
Proposed LR

10
565,825
565,825
564,551
564,800
564,049
563,977
563937.69

COST($)
No of units
40
2,258,503
2,251,911
2,249,093
2,242,178
2,256,195
2,249,715
2,243,245

20
1,130,660
1,126,243
1,125,494
1.122.622
1.128.098
1.125.516
1,122,637

Table 5. Solution of 10 unit 24-hour using the

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

1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1

2
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1

3
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0
0
0

4
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0
0
0

Unit Number
5
6
0
0
0
0
1
0
1
0
1
0
1
0
1
0
1
0
1
1
1
1
1
1
1
1
1
1
1
1
1
0
1
0
1
0
1
0
1
0
1
1
1
1
1
1
0
1
0
0

7
0
0
0
0
0
0
0
0
1
1
1
1
1
1
0
0
0
0
0
1
1
1
0
0

1 1 1 1 1
Intermediate unit
coding at t = 21
21

80
4,526,022
4,504,933
4,498,479
4,501,844
4,512,391
4,505,614
4,484,915

100
5,657,277
5,627,437
5,623,885
5,613,127
5,640,488
5,640,488
5,604,470

Table 6. CPU time comparison

8
0
0
0
0
0
0
0
0
0
1
1
1
1
0
0
0
0
0
0
1
0
0
0
0

9
0
0
0
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0

10
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0

METHOD
GA [8]
EP [19]
LRGA [20]
DPLR [21]
GAUC [21]
Proposed LR

All possible
intermediate units
coding a t = 22

60
3,394,066
3,376,625
3,371,611
3,371,079
3,384,293
3,375,063
3,363,376

22

10
221
100
518
108
85
10

20
733
340
1147
299
225
14

All possible
intermediate units
coding a t = 23
1 0 0 0 0
0 1 0 0 0
0 0 1 0 0
0 0 0 1 0
0 0 0 0 1
23

CPU time
No of units
40
60
2697 5840
1176 2267
2165 2414
1200 3199
614
1085
25
39

24

Fig. 1. Dynamic programming search for uncertain stages 22 and 23.

80
10036
3584
3383
8447
1975
64

100
15733
6120
4045
12437
3547
80