A Fast Technique for Unit Commitment Problem
by Extended Priority List
Tomonobu Senjyu, Kai Shimabukuro, Katsumi Uezato, and Toshihisa Funabashi
Abstmct- This paper introduces a new unit commitment
problem, adapting Extended Priority List (EPL) method.
The EPL method consists of two steps, in the first step we
get rapidly some initial unit commitment problem schedules
by Priority List (PL) method. At this step, operational
constraints are disregarded. In the second step unit schedule
is modified using the problem specific heuristics t o fulfill
operational constraints. To calculate efficiently, however,
note that some heuristics is applied only to solutions can
expect improvement.
Several numerical examples demonstrate the effectiveness
of proposed method.
Keywords- power system operation and planning, unit
commitment, priority list, heat rate, priority order.
I. INTRODUCTION
N EW power supplier has entered the electric power
market since the deregulation of electric utilities
started. Therefore, power suppliers are exposed to com-
petition of sale of electric power and has to aim at an eco-
nomic operation of power systems to obtain maximum prof-
its. Then, the Unit Commitment (UC) problem is manda-
tory in planing and operation of power systems. The basic
goal of the UC problem is to properly schedule the on/off
states of all the units in the system. In addition to fulfill-
ing a large number of constraints, the optimal UC should
be meet the predicted load demand, calculated in advance
plus the spinning reserve requirement at every time interval
such that the total cost is minimum. The UC problem is
formulated as a combinatorial optimization problem with
0-1 variables which represents on/off status and continuous
variables which represents unit power. However, the num-
ber of combinations of 0-1 variables grows exponentially
as being a large scale problem. Therefore, this problem is
known as one of the problems which is the most difficult to
solve in power systems.
The solution methods being used to solve the UC prob-
lem can be divided into four categories[l-71:
0 Priority List (PL) PI
0 Branch-and-Bound (BB) P-31
0 Lagrangian Relaxation (LR) [4]
0 Meta-heuristics 15-71
For the PL methods, the units are committed in ascend-
ing order of the unit full-load cost so that the most eco-
nomic base load units are committed first and the peak-
ing units last in order to meet the load demand. The PL
Tomonobu Senjyu, Kai Shimabukuro, and Katsumi Uezato are with
the Department of Electrical and Electronics Engineering, Faculty of
Engineering, University of the Ryukyus, Okinawa, Japan. Email:
b9844520tec.u-ryukyu.ac.jp. Toshihisa Funabashi is with the Mei-
densha Corporation, Tokyo, Japan. Email: funabashi-tQhonsha-
meidensha.co.jp.
0-7803-7525-4/02/$17.00 0 2002 IEEE.
methods are very fast but they are highly heuristic and
give schedules with relatively high operation cost. The BB
methods has the danger of a deficiency of storage capacity
and increasing the calculation time enormously as being a
large scale problem. The Lagrangian Relaxation method
concentrates on finding an appropriate co-ordination tech-
nique for generating feasible primal solutions, while min-
imizing the duality gap. The main problem with the LR
methods is the difficulty encountered in obtaining feasible
solutions. The meta-heuristic methods is iterative search
technique that can search not only local optimal solution
but also global optimal solution and can treat various con-
straints. In the meta-heuristic methods the technique fre-
quently applied to the UC problem are Genetic Algorithm
(GA) and Tabu Search (TS). The GA is general-purpose
search technique based on principles inspired from the ge-
netic and evolution mechanisms observed in natural sys-
tems and populations of living beings. The feature of the
GA is to search retaining candidate for the solutions. The
TS is approach which avoids the circle of solutions and
searches in the all search space. Tabu List is used to avoid
revisiting certain areas of the search space that have al-
ready been searched. These meta-heuristic methods give
suitable calculation time for reasonably sized problems.
However, in case of large scale problem they consumes a
lot of time due to its iterative nature.
In this paper we propose a new unit commitment method
which applies fast Extended Priority List (EPL) method.
First, the PL method yields rapidly a initial solution.
Then, plurality of solutions is produced based on derived
one before. Finally, among all solutions, only potential SG
lutions of improvement are implemented some fast heuris-
tics, which can reduce the cost. Therefore, the proposed
method provides a satisfactory solution in terms of solution
accuracy and computational effort.
11. PROBLEM
FORMULATION
A . Notation
The following notation is used throughout the paper:
N : number of generators
T : total scheduling period
i : index of unit ( Z = l , 2, . . ., N )
t : index of hour (t=l, 2, e.., T)
I z ( t ) : unit off/on [O,1] status of
unit i at hour t
Pt(t): generation of unit i at hour t
P,,, : unit i maximum generation limit
PzmEn : unit i minimum generation limit
244
 D ( t ) : system load demand in the hour t
Rt : spinning reserve requirements
T,On : minimum up time of ith unit
T,off : minimum down time of ith unit
Xtm(t) : duration during which unit i
is continuously on
X p f f ( t ) : duration during which unit i
is continuously off
SC,(t) : start up cost of unit i a t hour t
Sol@): shut down cost of unit i a t hour t
F,(P,(t)): fuel cost of unit i at hour t
CF : total cost
h-cost, : ith unit hot start cost
c-cost, : ith unit cold start cost
c-s-hour, : cold start time of unit i
X : number of unit inside the window
W : number of initial solutions
S : number of solutions applying
several heuristics.
B. Formulation
1. Objective function
The total cost, CF over the entire scheduling periods is
the sum of the running cost, start up cost and shut down
cost for all the units. Accordingly, overall objective func-
tion of the UC problem is:
T N
min CF = + +
x [ F , ( P , ( t ) ) SC*(t) SDa(t)]. (1)
t=l,=l
Generally, the fuel cost, F,(P,(t)) per unit in any given
time interval is a function of the generator power output.
Most frequently used cost function is in the form of
Fa(P,(t))= a, + b, * P'(t) + c, * P,2(t) (2)
a,, b,, and c, represent the unit cost coefficients.
The generator start up cost depends on the time the
unit has been off prior to start up. In this aper, time-
dependent start up cost is simplified using HPpf defined as
equation (4) as follows:
h-cost, : T f f f 5 Xioff 5 H f f f
SCI = (3)
c-cost, : Xioff > Hioff
Hiof = qf + c-s-houri. (4)
The shut down cost, SD, is usually given a constant
value for each unit. In this paper, the shut down cost has
been taken equal to 0 for each unit.
2. Constraints
The UC problem has some constraints for the character-
istic of a generator, a technical side, etc. These include sys-
tem constraints and the generating units constraints. The
245
overall objective is to minimize CF subject to a number of
constraints:
Svstem constraints
(a) System power balance
System power balance on hourly basis, where the total
power generated must supply the load demand D ( t ) ,
N
D ( t )= Pa(t). (5)
i=l
(b) System reserve requirements
Hourly spinning reserve requirements Rt must be met
N
i= 1
Unit constraints
(c) Generation limits
Unit rated minimum and maximum capacities must not
violate,
Pimin IPi(t) IPimaze (7)
(d) Unit minimum up/down time
Minimum up/down time limits of units must satisfy the
following:
(e) Unit initial status
The initial status a t the start of the scheduling period
must be taken into account.
111. EXPLANATION
OF THE EPL METHOD
A . Overview
We explain the proposed method briefly. The EPL
method consists of two steps, in the first step we get rapidly
plurality of initial solutions by PL method. Advantage of
using PL method is to be obtained solution which forms
base of unit commitment schedule close to optimal sched-
ule, in addition t o obtaining rapidly solution. That is, the
schedule obtained by PL method decides that the more
costly unit for generation cost is off status, whereas more
cheap unit for generating cost is settled on status. In this
time all constraints are not satisfied but the PL method
decides a rough framework which is costly units are off sta-
tus and cheap units are on status. In the second step we
apply several heuristics to plurality of initial solutions. To
apply heuristics to all solutions is a time consuming pro-
cesses. Hence, in this paper using a proper condition, the
proposed method carries out the reduction of number of so-
lutions applying heuristics. Then, only the solutions which
are not potentially improvement are reduced.
The details of the generation techniques for plurality of
initial solutions obtained by PL method and incorporation
of several heuristics are explained below. Note that actual Table 1. Heat rate(at maximum power).
introduction order of the heuristics in programming is the I Unit I HR II Unit I
same order as the explanation of heuristics, given in the
following lines.

B. Generation of initial solution

The generation of initial solution is important, particu-
larly, for the UC problem. The initial solution is usually
generated at random. However, this technique is difficult
to get feasible solution for the UC problem with many con-
straints, resulting in quality of solution obtained will go
bad. Ftecently, in order to solve this problem, the tech-
nique of starting search from the solution which satisfies
all the constraints in the stage of an initial solution has
been proposed.
In this paper, initial solution is generated using the PL
method. However, as it is also in the abovementioned, the
solution which obtained by PL method cannot fulfill all
constraints. Merely characteristics which costly units are
off status and cheap units are on status, can be expected
to obtain approximate solution close to optimal solution by PriorityList H 1 H+1 , H+2 j H+3 I Period
modification of solution in after section. Fig. 1. Initial solution by Priority List method.
First, priority list as shown in Fig. 1 is created based
on each unit parameters. A unit, which is the highest of
maximum output power, is located in the bottom of the
list, and the other units are located in descending order
of their maximum output power towards the top of the
list. The most costly unit which is on the top of the list
would have the lowest priority. Then, commit units until
the load demand plus the spinning reserve requirements
represented by ( 6 ) are fulfilled in the priority list order at
every time interval. In addition, units of equal maximum
output power are started up in ascending order of heat rate
( h r ) ,which means average fuel piece rate per output power
given in Table 1. The hr is given by the following formula

hr = l q P $ ( t ) ) / P i ( t ) . (9) PriorityList H H+1 f H+2 H+3 i Period

Finally, plurality of solutions are generated based on ob- Fig. 2. Unit located in a window(case of X = 3).
tained one by PL method at this point as shown in Fig. 1.
The technique of generating a solution sets up at random D. Incorporated heuristics
on/off status of X units located in a window (shaded area) 1. Start up of the base load unit
shown in Fig. 2 at every time interval, and one solution In power system, the units which must run at any time
is generated when a setup of the last period is completed. exists. We desire these units, which becomes the base of
The same procedure is repeated several times and W initial load, to start up at all scheduling period than repeating on
solutions are generated. or off because these units has relatively large output power.
Since the solution by PL method is started up from the
C. The condition for solution reduction
unit of a large output power, although most units used as
In order t o reduce computational efforts, the plurality of a base will be in a on state, all base units does not become
solutions are reduced by adding following condition, only so. In the proposed method, for 10-unit systems, two units
remaining solutions are applied heuristics. Actually, eq. which are located from the bottom of priority list to toward
(10) is calculated about each solution and solutions are the top are specified as the base load unit as shown in
rearranged into the small order of Y.After rearrangement, Fig. 3, these two units are fixed on status. hrthermore,
each heuristics explained below is applied to the solution when expanding the number of units, the base units were
of S higher ranks. assumed to be 20 % of all units.
2. Consideration of the minimum up/down time at the
period inserted into the load peak
In the period inserted into the load peak, how the con-

246
 a a a
straint represented by eq. (8) for relatively large units of a 0
a
a
a
minimum up/down time is satisfied poses an important
problem. For example, when assuming that the minimum
up/down time of unit e shown in Fig. 4 is relatively large
five hours, since unit e does not need to generate electricity
1 2 3 4 5 ' 1 2 3 4 5
at hour (H+2), unit e is shut down usually but since it must Fig. 3. Starting of the base load unit.
restart in the next hour (H+3), unit e must start up also in (Case of scheduling period T = 5)
a previous hour (H+2) to satisfy the minimum up time. On
the contrary, it is also possible to satisfy minimum down Unit Power
time by shutting down the unit e. However, other costly
units are started up instead of shutting down unit e, as . ... ___ .. . . ...
i D(r)t Ri
a result the total cost will be high. Hence, this heuristic
starts up units in large order of minimum up/down time,
i.e., priority list order at the period inserted into the load
peak. From prior simulation, we assume that started unit
is until one that the total cost becomes more cheap.
3. Consideration of the minimum up/down
times of all units
At this moment in time, the minimum up/down times
of all units has not considered. In order to be satisfied the PnontyList U H+1 H+2 H+3 Penod

- -
minimum up/down times of all units, this heuristic looks
Fig. 4. Consideration of the minimum up/down times
for the hour changing from on status to off status between 1
of the period inserted into the load peak.
hour to 24 hours every unit, and corrects to on or off until
the minimum up/down time is satisfied if the minimum
up/down time is not satisfied at this hour. The example of ~ l ~ l ~ O ~ O~ o~ ~lo ~l l l o~l ol l ]o l o ~
operation in case the minimum up/down times is 3 hours
is shown in Fig. 5. ~ l ~ l ~ O ~ O ~~ O0 ] ~l l0l ~~ 1 ~ 1 ~ 1 ~ 0 ~ 0 ~
4. Commitment of the excess started or deficient units Fig. 5. Consideration of the minimum
up/down times of all units.
When the above heuristic is completed, excess started
or deficient units are generated at one period. In order to Unit Power D(t) + R' IJnit Power D(t) + R'
reduce total cost, their units must be committed certainly
without violating the constraints. The procedure of com-
mitment of the excess started or deficient units is explained
below and the example of operation is shown in Fig. 6. In
Fig. 6 , units are located in shaded area indicates the excess
started or deficient units.
The priority order for shutting down the excess started
H : H+I i H+2,H+3 Pf:rid H iH+l'H+2:H+3iPenod
units is decided by heat rate of each unit at maximum
output power as shown in Table 1. The order of shutting Fig. 6. Commitment of excess started
down the unit is the large order of the heat rate. Then, the or deficient units.
hours which are satisfied of eq. (6) are searched, starting
units are shut down every one unit in the above decided
ended when eq. (6) is satisfied.
shut down order. This heuristic judges whether the unit
schedule can satisfy eqs. (6) and (8) every shutting down 5. Modification of the grey zone for start up cost
one unit, and keeps up remaining shut down if the unit The grey zone for start up cost indicates the change point
schedule can satisfy two equations, and starts up the unit where changes from the cold start up cost to the hot start
again if not. The above operation is carried out as long as up cost. As shown in Table 3, the cold start up cost is
constraints are fulfilled. two times larger than the hot start up cost. Hence, it
Inversely, in case of no satisfying eq. (6) at one period, is desirable that the unit start up with the hot start up
i.e., deficient units exists, the units are started up every cost if possible when the unit starts up. To reduce in the
one unit in inverse order for shutting down. This heuristic cost for detail, the grey zone modification algorithm, which
judges whether the unit schedule can satisfy eqs. (6) and searches the grey zone for start up cost and modifies sched-
(8) every starting up one unit, and keeps up remaining ule, is incorporated in this paper. In order to expedite the
starting up if the unit schedule can satisfy two equations, reduction of the cost, the unit schedule is modified in order
and shuts down the unit again if not. This operation is of the unit that the start up cost is high and the fuel cost

247
is low, i.e., priority list order. The grey zone modification Table 2. Load demand.
algorithm incorporated in this paper is explained briefly
below.
STEP1 : The S solutions obtained by all heuristics men-
tioned above are calculated each total cost and the cost
which is the smallest cost in S solutions is set to as a in-
cumbent value.
STEP2 : At hours changing from off at previous hour to
on at next hour, the grey zone is searched each unit in order Unit1 Unit2 Unit3 Unit4 Unit5
P,,,, 455 455 130 130 162
of priority list by modifying from off to on. p,,,, 150 150 20 20 25
STEP3 : Modified solution is calculated the total cost if a. 1000 970 700 680 450
the grey zone exists and a incumbent value is updated if the
total cost is improved compare with a incumbent value. In
this time modified solution is updated to a incumbent solu-
tion. Inversely, if the grey zone does not exist or modified
solution is infeasible, modified part is returned.

E. ELD calculation
In unit commitment problem, the ELD calculation con-
sumes a large amount of calculation time. In this paper,
the ELD calculation is performed only for feasible solution.
Compared with GA and TS which needs a lot of ELD cal-
culations, the proposed method can attain shortening of
calculation time by leaps and bounds.
Only the solutions which applied several heuristics will
become feasible solution and these numbers will be about
S. The number is represented as about S because all so-
lutions which applied several heuristics are not necessarily
possible to perform. From mentioned above, the number of
ELD calculation is less than S according to the proposed
method, and in case of large problem the proposed method
can expect solution t o be obtained on practical calculation
time.
IV. SIMULATION
CONDITIONS
The unit characteristics and the load demand, which are
used in this paper, are given in Tables 2 and 3, respec-
tively[5,7]. In order to perform a simulation on the same
conditions in [5] and [7],in all simulations the spinning
reserve requirements was assumed to be 10 % of the load
demand and total scheduling period is 24 hours. The sim-
ulations included test runs for 10, 20, 40, 60, 80, and 100
unit systems.
Expansion of the number of unit is based on 10 unit sys-
tem, for the 20 unit system the base 10 units were dupli-
cated and the load demand was multiplied by 2. Similarly,
other unit systems are expanded too.
In the Table 3, ‘ini state’ represents the initial unit state
on scheduling period, and the positive sign indicates that
the unit is on, whereas the negative sign indicates that the
unit is off. For example, since initial state of unit 3 is -5,
this indicates that unit 3 had kept starting up for 8 hours.
The priority list created from the unit characteristics is
given in Table 4. In this paper, the priority l i t was created
so that the larger unit of output power could have a high
priority. Hence, unit 1 has the highest priority, whereas
unit 10 has the lowest priority. The simulation parameter
Table 4. Prioritv list.
number of initial solutions w 11 20 -
number of unit inside the window x 11 1
number of solutions applying several heuristics S 11 5 -
used in the EPL,such as X, W , and S, are given in Table
5. The number of unit inside the window is one unit which
was obtained the best result from prior simulation. The
unit started in order of priority list mentioned in previous
section is till unit 3. This number is half of all units in
Table 4. For 10 unit system, the units started in order of
priority list are 5 units, in case of 100 unit system it is 50
units.
The simulation is performed for each unit systems with
conditions given in Table 5. Furthermore, for 40 unit sys-
tem we examine the effects on the total cost and execu-
tion time by varying X. Also, we simulate in two case of
modifying grey zone (Case 1) and without modifying grey
zone (Case 2) t o examine the effect on the total cost by
modifying grey zone for start up cost. The all calculations
are run on a Intel Pentium4 CPU(1.5GHz), Linux version
2.2.18 and gcc version 2.91.66.
V. SIMULATION
RESULTS
Table 6 shows the comparison of proposed EPL method
with GA and EP of total cost. In table 6, the EPL method
gives the lowest total cost of these method in all unit sys-
tems. Table 7 shows the comparison of EPL method with

248
 Table 8. Grey zone modification effect.
No. of Total cost[$’
unit EPL I GA151 EPIV unit
563977 563977
cost[%]
10 563977 I 565825 565352
__
4n - ._
343
60 I 3366210 I 3366547 337
80 I 4489322 I 4490009 687
inn I 5 ~ 0 ~ 4 4 0i 5609040 fino

(W = 20, s = 5, x = 1)

(W = 20, s = 5,x = 1)
2.245

0
t 2 4 6 8 10 12 14 16
Number of unit inside the window X

GA and EP of execution time. According to Table 7, the (a) effect on total cost.
EPL method can obtain better solution rapidly even if the
problem is large because the generation of the feasible so-
lution is limited by the solution reduction, and the number
of ELD calculation is reduced.
In the proposed method, the generation of plurality of
solutions is performed by setting up a t random on/off sta-
tus of X units in Fig. 2 at every time interval. Hence, we
have an interest in the effects on the solution quality and
execution time by varying X. Figs. 7(a) and 7(b) show
the simulation results of the total cost and execution time Number of unit inside the window X
for 40 unit system in the same conditions as Table 5 ex- (b) effect on execution time.
cept varying X. In Fig. 7(a), the total cost is minimum as
X = 1, and it indicates that the quality of a initial solution Fig. 7. Influence of varying X.
(40 unit system : W = 20, S = 5 , varying X)
obtained first by PL method excels very much. The total
cost increases rapidly a t X > 13 because the feasible solu- as a result the number of the ELD calculation can be de-
tion does not exist. Please note that the proposed method creased. Hence, for a larger problem the proposed method
adds large value as a penalty of the total cost to infeasible can get solution with practical calculation time. For oper-
solution. In Fig. 7(b), transit of execution time is staying ating more economically, to take modification of the grey
confirred range when the feasible solution exists but the ex- zone for the start up cost into consideration is beneficial.
ecution time decreases as well as non generating feasible
solution at X > 13. As shown in Figs. 7(a) and 7(b), both REFERENCES
H. H. Happ, R. C. Johnson, and W. J. Wright, ‘‘ Large Scale Hydro-
of the execution time and total cost changes in confined
Thermal Unit Commitment-Method and Results , IEEE ‘nuns. on
range at X 5 13, it indicates that the EPL method has a
”
Power App. and Syst., 90, 3,pp. 1373-1384,1971.
performance generates stably about S feasible solutions. A. Ohuch and I. Kaji, A Branch-and-Bound Algorithm for Startup
“

In order to verify effectiveness of the grey zone modifica-
tion discussed above, we compare with Case 1 and Case 2
in the same conditions as Table 5. The simulation results
are given in Table 8. The grey zone modification for start
up cost can modify the parts which cannot consider by the
other proposed heuristics, as a result it gives more good
unit schedule as shown in Table 8.
VI. CONCLUSIONS
In this paper we proposed a fast extended priority list
method for the UC problem. The PL method is applied
as a technique getting better initial solution rapidly. The
effectiveness to use the PL method is clear from simula-
tion results. The proposed condition for reduction of so-
lutions in this paper can expedite reducing solution and
and Shutdown Problem of Thermal Generating Units ” , IEEJ, Vol.
95-B,No. 10, pp. 461-468,1975.
G. S. Lauer, D. P. Bertsekas, N. R. Sandell,Jr., and T. A. Posbergh,
“ Solution of Large-scale Optimal Unit Commitment Problems ,
IEEE %ns. on Power Apparatus and Systems, Vol. PAS-101, No.
1, 79-86,1982.
Alva J. Svoboda, Chung-Li Tseng, C h e a n Li, and Raymond B.
Johnson, “Short-Term Resource Scheduling with Ramp Constraints”,
IEEE h n s a c t i o n s on Power Systems, Vol. 12, No. 1, pp. 77-83,
1997.
S. A. Kazarlis, A. G . Bakirtzis, and V. Petridis, “A Genetic Algorithm
Solution t o the Unit Commitment Problem” , IEEE ‘nunsactiow on
Power Systems, Vol. 11, No. 1, pp. 83-92, 1996.
H. Mori and 0. Matsuzaki, “ Application of Priority-List-Embedded
Tabu Search to Unit Commitment in Power Systems ” , IEEJ, Vol.
121-B,NO. 4, pp. 535-541,2001.
K. A. Juste, H. Kita, E. Tanaka, and J . Hasegawa, “ An Evolutionary
Programming Solution to the Unit Commitment Problem ” , IEEE
’nunsactiow on Power Systems, Vol. 14,No. 4,pp. 1452-1459,1999.

249