A Transmission-Constrained Unit Commitment Method

Chung-Li Tseng, Shmuel S. Oren,

Department of IEOR, UC Berkeley, Berkeley, CA 94720, USA
[email protected], [email protected]
Carol S. Cheng, Chao-an Li, Alva J. Svoboda, Raymond B. Johnson
Pacic Gas & Electric Company, San Francisco, CA 94177, USA

Abstract In today's power system, generating units of a util-

This paper presents a transmission-constrained unit ity company are normally located in multiple areas
commitment method using a Lagrangian relaxation ap- that are interconnected via transmission lines. If the
proach. The transmission constraints are modeled as transmission constraints are not considered, the sched-
linear constraints based on a DC power ow model. ule obtained might cause some transmission lines to be
The transmission constraints, as well as the demand overloaded. In this paper, we will discuss the unit com-
and spinning reserve constraints, are relaxed by attach- mitment problem which takes transmission constraints
ing Lagrange multipliers. In this paper we take a new into account.
approach in the algorithmic scheme. A three-phase al- The transmission constraints will be modeled as lin-
gorithm is devised including dual optimization, a fea- ear constraints based on a DC power ow model. This
sibility phase and unit decommitment. A test problem is also the approach taken in other papers, e.g. [9, 13].
involving more than 2500 transmission lines and 2200 Methods for solving the transmission-constrained unit
buses is tested along with other test problems. commitment problem have been developed. Pang et
al have considered the problem in [12]. The approach
1. Introduction taken in [12] is a dynamic programming method. Lee
develops a sequential method in [8, 9], which sequen-
The unit commitment problem at a power utility tially determines the commitment of the next most-
like PG&E requires economically scheduling generat- advantageous unit to commit. The decision making
ing units over a planning horizon so as to meet forecast involves a price-adjustment procedure which resem-
demand and system operating constraints. It has been bles a bidding process. In [13], Shaw has proposed
an active research subject due to potential cost saving a practical method for solving the security-constrained
and the diculty of the problem. The unit commit- unit commitment problem using the Lagrangian relax-
ment problem is a mixed integer programming prob- ation approach. This approach relaxes not only the de-
lem, and is proved to be NP-hard in [15]. Many opti- mand constraints and the spinning reserve constraints,
mization methods have been proposed to solve the unit but also the transmission constraints using multipli-
commitment problem (e.g. see [5] for a survey). These ers. Shaw describes two methods in his paper [13],
methods include priority list methods [4], dynamic pro- a direct method and an indirect method. The former
gramming methods [11, 12, 17], sequential method [8] takes full account of the transmission constraints in the
and Lagrangian relaxation methods [3, 5, 6, 7], etc. La- optimization phase, while the latter does so only in lo-
grangian relaxation methods are now among the most cating a feasible solution. The conclusion of [13] favors
widely used approaches to solving unit commitment. the direct method.
At PG&E, the Hydro-Thermal Optimization (HTO) Algorithmically, the dual optimization approach in
program was developed almost a decade ago based on this paper corresponds closely to that presented in [13].
the Lagrangian relaxation approach [6]. In recent work, However, we take a new approach in the algorithmic
the Lagrangian relaxation-based algorithm has been scheme. A three-phase algorithm is proposed including
extended to schedule thermal units under ramp con- unit decommitment [10, 16]. A thorough discussion of
straints ([14]). the three-phase algorithm scheme for solving the unit
commitment problem can be found in [15]. A test prob- Dbt : forecast demand requirement at bus b in time
lem involving more than 2500 transmission lines and period t
2200 buses is tested along with other test problems.
This paper is organized as follows: Section 2 gives Rbt : spinning capacity requirement at bus b in time
the model of the unit commitment problem to be dis- period t
cussed. A three-phase Lagrangian relaxation algorithm The transmission-constrained unit commitment
is presented in Section 3. Section 4 gives some numer- problem is formulated as the following mixed-integer
ical results. We conclude the paper in Section 5. programming problem: (note that the underlined vari-
ables are vectors in this paper, e.g. the components of
2. Problem formulation u are all legitimate uit.)

The notational convention and problem formulations Minimize the total generating cost:
in this paper have been in uenced by the work of Shaw X
T X
B X
[13]. We rst dene the following notation. min [Ci(pit)uit +Si (xi;t,1; uit; ui;t,1)] (1)
u;x;p
t=1 b=1 i2
b
t : index for time (t = 0;


; T)
subject to the following constraints:
b : index for the number of buses (b = 1;


; B) Demand constraints

b : index set of units at bus b X

B X X
B
pituit = Dt  Dbt ; 8t; (2)
i : index for the number of units (i 2
b ; b = 1;


; B) b=1 i2
b b=1
` : index for transmission lines (` = 1;


; L) X
B X
,`b : line ow distribution factor for transmission line , F`  ,`b( pituit , Dbt )  F` ; 8`; t; (3)
b=1 i2
b
` due to the net injection at bus b
Spinning capacity constraints
F` : the transmission capacity on the transmission line
` X
B X X B
i uit  Rt 
pmax Rbt; 8t; (4)
uit : zero-one decision variable indicating whether unit b=1 i2
b b=1
i is up or down in time period t. Local reserve constraint
xit : state variable indicating the length of time that X
unit i has been up or down in time period t pituit  rbtDbt ; rbt 2 [0; 1]; 8b; (5)
i2
b
ton
i : the minimum number of periods unit i must re- X
main on after it has been turned on i uit  sbt Rbt; sbt 2 [0; 1]; 8b;
pmax (6)
i2
b
to
i : the minimum number of periods unit i must re- where sbt and rbt are scalars used to dene the local
main o after it has been turned o minimum generation level and local spinning capacity
pit : variable indicating the amount of power unit i is level within bus b. Although constraints (5) and (6) are
generating in time period t normally considered within an area instead of a bus, we
intend to present a method suitable for a generalized
pmin
i : minimum rated capacity of unit i multi-area model.
Local unit constraints
pmax
i : maximum rated capacity of unit i Unit capacity constraints, for all i 2
b, 8b and t =
Ci(pit ) : fuel cost for operating unit i at output level 1;


; T :
i  pit  pi ;
pmin max (7)
pit in time period t
the state transition equations
Si (xi;t,1; uit; ui;t,1) : startup cost associated with 
turning on unit i at the beginning of time period xit = max(x i;t,1; 0) + 1; if uit = 1; (8)
t min(xi;t,1; 0) , 1; if uit = 0;
the minimum uptime and downtime constraints, X
T X
B X
8 , [bt( pit uit , rbtDbt)
< 1; if 1  xi;t,1 < ton
i ; t=1 b=1 i2
b
uit = : 0; if , 1  xi;t,1 > ,to
i ; (9) X
0 or 1; otherwise; + bt( i uit , sbt Rbt)]
pmax (11)
i2
b
and the initial conditions on xit at t = 0 for 8i. subject to initial conditions, and the unit constraints.
After rearrangement of the terms in (11), the separa-
Remark bility of d(; ; ; ; ; ) appears.
For simplicity, ramp constraints are not covered in
this paper. The reader who is interested in ramp- X
B X
constrained unit commitment is referred to [14]. If in- d(; ; ; ; ; ) = di(b; ; ; ; ; ; )+
corporated, the ramp constraints should be taken care b=1 i2
b
of as are other local unit constraints like (7) within the (; ; ; ; ; ); (12)
corresponding bus subproblem, which will be detailed
later. where
X
T
3. The Lagrangian relaxation (; ; ; ; ; ) = (t Dt + tRt)
algorithm X
T X
L X
B
t=1

X
B
The Lagrangian Relaxation (LR) approach relaxes , [t`(F` , ,b`Dbt ) + t` (F` + ,b` Dbt)]
the demand constraints, the spinning reserve con- t=1 `=1 b=1 b=1
straints and the bus-interchange constraints using La- XX
T B
grange multipliers. Algorithmically, this dual approach + (btrbtDbt + bt sbt Rbt) (13)
closely corresponds to the direct method in [13]. t=1 b=1
is a constant term given the multipliers, and
3.1. Phase 1: the dual optimization di(b; ; ; ; ; ; ) =
Letting t , t be the corresponding nonnegative La-
grange multiplier of (2) and (4), respectively. And tl , X
T
tl are corresponding to (3), bt and bt to (5) and (6), min
uit;xit ;pit
[Ci(pit)+Si (xi;t,1; uit; ui;t,1),t pit,t pmax
i
respectively. We now have the following dual problem: t=1

(D) max d(; ; ; ; ; ) (10) X

L
(;;;;;)0 , (t` , t` ),`b pit , btpit , bt pmax
i ]uit: (14)
`=1
where Once again, the minimization in (14) is subject to ini-
d(; ; ; ; ; ) tial conditions and the unit constraints.
Note that di is a unit subproblem corresponding to
XT X B X unit i. We can further dene a bus subproblem db such
 uitmin
;xit ;pit
[Ci(pit )uit+Si (xi;t,1; uit; ui;t,1)] that
t=1 b=1 i2
b
X
XX X
T B X max d b (; ; ; ; ; )  di(b; ; ; ; ; ; ); (15)
, [t( pit uit , Dbt )+t ( pi uit , Rbt )] i 2
b

t=1 b=1 i2
b i2
b
and the dual objective (12) is equivalent to
XXT L X X
B
, [t`(F` + ,`b( pituit , Dbt)) d(; ; ; ; ; )
t=1 `=1 b=1 i2
b P B
= b=1 db(; ; ; ; ; ) + (; ; ; ; ; )
X XB
Each unit subproblem di (14) can be solved us-
+ t` (F` , ,`b ( pituit , Dbt ))]
b=1 i2
b ing dynamic programming. The dual function
d(; ; ; ; ; ) is concave, but not necessarily dier-
entiable at all points (e.g. [2]). A subgradient method Denition 1 With respect to a given spinning capac-
is employed to solve the dual maximization problem ity requirement fRtg, a reserve-feasible commitment is
(D). a unit commitment fuitg that satises the spinning re-
The dual optimization algorithm is summarized as serve capacity constraints (4), (6), unit constraints (8),
follows. (9) and the initial conditions xi0.
Phase 1: Dual optimization
Step 0: k 0; initialize 0 , 0, 0 ,  0 ,  0 , 0 ; b 1.
Denition 2 With respect to a given load require-
ment fDt g, a unit commitment fuitg is said to be dis-
Step 1: If stopping criteria are met, stop. If b > B, patchable if there exists a set of dispatches fpitg such
go to Step 4. Otherwise solve the b-th bus that (2), (3), (5) and (7) can be satised.
subproblem db (k ; k ; k ;  k ;  k ; k ) to ob-
tain (uk ; pk ). In [18], Zhuang proposed a method for obtaining
feasible solutions to the single-area unit commitment
Step 2: If (uk ; pk ) satises (5) and (6) for bus b, go problem. The basic idea is to obtain a reserve-feasible
to Step 3. Otherwise update btk and kbt as commitment (in the single-area case, ignore (6) in the
follows then go to StepP1: for 8t, denition). In [18] a reserve-feasible commitment is
btk max[0; btk , sk (P i2
b pkitukit , rbtDbt)] found by projecting the subgradient onto the hour cor-
kbt max[0; kbt , sk ( i2
b pmax responding to the most unsatised capacity constraint
i uit , sbt Rbt)]
k
and increasing only the multiplier t of this hour. A
Step 3: b b + 1, go to Step 1. method to calculate the exact amount of the increase
in the value of the corresponding t is also provided
Step 4: For 8t; ` in [18]. It can be shown that a (single-area) reserve-
btk+1 btk ; kbt+1 kbt;P P feasible commitment is dispatchable (ignoring (3) and
kt +1 max[0; kt , sk ( Bb=1 i2
b pkitukit, (5) in the denition in the single-area case) if and only
Dt )]; P P if the minimum load condition below is satised.
kt +1 max[0; kt , sk ( Bb=1 i2
b pmax i uit,
k X min
Rt)]; pi uit  Dt ; 8t: (17)
PB ,t` (,Ps (F` +pk uk , D ))];
kt`+1 max[0; k k i
b=1 `b i2
b it it bt However, in the presence of transmission constraints,
t`k+1 max[0;  k , sk (F` , this is not the case. The feasibility phase presented in
PB ,t` (P pk uk , D ))]; this paper contains two parts. Part 1 seeks a reserve-
b=1 `b i2
b it it bt
feasible commitment by increasing the multipliers t
Step 5: k k + 1, b 1, go to Step 1. in hours with insucient spinning capacities. Based
The step size in the algorithm used in our implemen- on the obtained reserve-feasible commitment, Part 2
tation has the following form. looks for a dispatchable commitment while maintaining
reserve-feasibility by solving the following linear pro-
sk = mk =kgk k; m > 0; (16) gram (LP) for all hours, (LP(t)) for t = 1;


; T . Sup-
pose Part 1 of the feasibility phase terminates at itera-
where gk is the subgradient of the dual function at iter- tion k with an obtained commitment fukitg. Let qbt be
ation k and mk is an adaptive constant. Our stopping the variable representing the total generation to be dis-
criteria used in the algorithm combine the maximum patched to bus
P b in time t. Given a commitment
P fukitg
number of iteration, the change of norm of subgradi- min min max
let qbt = i2
b pi uit and qbt = i2
b pi ukit,
k max
ents at two consecutive iterations and the number of which constitute an immediate lower bound and upper
iterations without improvement in the dual objective bound for qbt.
value.
(LP(t))
3.2. Phase 2: the feasibility phase min
PB (y , qmax ) (18)
qbt ;ybt b=1 bt bt
It is well known that the solution obtained through
the dual optimization is generally not feasible. A fea- s:t: qbt  qbt  ybt ; 8b
min (19a)
P
sibility phase is therefore required. qbt  ybt  i2
b pmax
max
i ; 8b (19b)
 PB PB
b=1 qbt = b=1 Dbt (19c) 3. One easy way to test the robustness of the feasibil-
PB ity phase of a transmission-constrained unit com-
,F`  b=1 ,`b(qbt , Dbt )  F` ; 8` (19d) mitment method is to apply it to an instance in
qbt  rbtDbt; 8b (19e) which each single bus has enough capacity to han-
In (LP(t)), ybt is a variable upper bound of qbt. Let dle its own load over the planning horizon. Then
ybt be the solution of (LP (t)). If fukitg is dispatchable the transmission line capacities are graduately re-
in time t, ybt = qbtmax solves (LP (t)) and the optimal duced to zero. In such a problem, a feasible solu-
objective value of (LP (t)) is zero. If fukitg is not dis- tion obviously exists, and it is equivalent to solv-
patchable in time t, (LP (t)) yields a positive solution, ing many single-area unit commitment problems.
which indicates that to be dispatchable, more units It can be easily veried that our feasibility phase
should be committed in time t, and the target value of reduces to solving many single-area problems in
the new (bus) capacity requirement is now increased such an instance.
to ybt for each bus b in time t. Part 1 of the feasibility 4. Should (LP(t)) be transformed to the standard
phase is therefore rerun with respect to the updated form of LP by adding slack or surplus variables:
capacity requirement fsbt Rbtg fybt g. The phase 2 min cx subject to Ax = b, x  0, it can be seen
algorithm is summarized below. that c and A are xed independent of t. This prop-
erty can be taken into account in implementation
to improve the algorithm's performance. For ex-
Phase 2 Algorithm ample, if the dual simplex method is used to solve
Step 0: k 0; 0 and 0 and other multipliers are from (LP(t)), the optimal basis of the LP in the previ-
Phase 1. ous hour can be stored and used as an initial (dual
feasible) basis.
Step 1: Given k and k and other multipliers, solve
the dual subproblems (14) to obtain (uk ; pk ). 3.3. Economic dispatch
Step 2: If (4) and (6) are satised, go to Step 5 if Step With the DC power ow model, the transmission
5 has not been visited, otherwise stop. constraints are formulated as linear constraints. If
Step 3: For 8b; t, the fuel cost is represented by a quadratic function
kt +1 P k
min(0; Rt,
kt + sP of the power generation, the transmission-constrained
B max k
b=1 i2
b pi uit); P economic dispatch is a quadratic programming prob-
kbt+1 kbt +sk
min(0; sbtRbt , i2
b pmax i uit):
k lem. Many methods can be used to solve this type of
(other multipliers are kept unchanged.) problem, including primal methods, e.g. the reduced
gradient method, and dual method, e.g. LR. The draw-
Step 4: k k + 1, go to Step 1. back of any dual method is in its being unable to de-
Step 5: Solve (LP (t)) for all t. If the optimal value of tect the (primal) infeasibility of the problem before-
(LP (t)) is 0 for all t, stop and uk is reserve- hand, which normally results in unboundedness dur-
feasible and dispatchable. Otherwise, update ing the dual optimization. In our implementation, the
the (bus) capacity requirements fsbt Rbtg transmission-constrained economic dispatch is solved
fybt g, 8b, where ybt solves (LP (t)). (Note fRtg by the commercial software MINOS, which does gen-
remains unchanged.) Then go to Step 1. eral nonlinear programming. We have also tested an-
other method using LR, which can generally handle
Remarks both the cases with smooth and piecewise quadratic
1. The loop between Step 1 and Step 4 corresponds functions as fuel cost functions. If using LR, although
to the Part 1 of the feasibility phase, and Step 5 its decomposition feature is very ecient, it sometimes
initiates Part 2. suers the drawback mentioned above.
2. Note that our proposed method for searching for a 3.4. Phase 3: unit decommitment
reserve-feasible commitment in Part 1 of the feasi-
bility phase at each iteration updates the t corre- Given a feasible schedule, a unit decommitment
sponding to the hours that the spinning capacity (UD) method [10, 16, 15] is used to improve this so-
requirement is violated simultaneously, instead of lution while maintaining feasibility. In [16, 15], a
only updating t in one hour at a time as in [18]. single-area unit decommitment method was devised as
a postprocessing method for the LR approach for solv- Table 1: Summary of test results
ing the unit commitment problem. A transmission- method Dual Total CPU time Duality
constrained unit decommitment algorithm can be de- value ($) Cost ($) (sec) Cap (%)
vised by repeatedly solving the single area unit decom- Three-phase 1100030 1100153 7.410 0.011
mitment. Given a feasible solution (~uk ; p~k ), at itera- Shaw 1110319 1111049 4.400 0.065
tion k the schedule of bus b can be improved by the
unit decommitment with respect to the current load
assigned toPbe the bus demand at this iteration, i.e. 1. We must emphasize that the results of our proposed
fDbtk g f i2
b p~kitu~kit g. After the commitments of all method (three-phase) in Table 1 and the results in [13]
buses are rened, a transmission-constrained economic (Shaw) are obtained under dierent bases: dierent
dispatch is performed to redispatch the units in all spinning reserve requirements and dierent computer
buses. The decommitment procedure can thus proceed facilities. However, the results of the duality gap re-
until no improvement in the total cost is made. Since at garded as a measure of solution error suggest that both
each iteration the decommitment procedure improves methods obtained very accurate solutions in this test
the commitment schedule without aecting the bus in- problem.
jection, the transmission feasibility is retained at each
iteration. The algorithm is as follows. 4.1. The test system
Transmission-constrained UD algorithm The second test problem is based on an IEEE test
Step 0: a feasible schedule (~u0; p~0) is given, k 0. problem [1]. This test system contains 24 buses, 34
transmission lines and 32 generating units. The pa-
Step 1: Given (~uk ; p~k ), each bus performs the unit de- rameters of the generating units are summarized in Ta-
commitment presented in [16] independently ble 2, in which the unit fuel costs are assume to be a
with the current load assigned toP be the de- quadratic function of power generation as: Ci(pit ) =
mand requirement fDbtk g f i2
b p~kit u~kitg ai0 + ai1 pit + ai2p2it, and the unit startup costs Si are
also subject to the satisfaction of (6). The re- assumed to be constants. This test system contains
sultant commitment is denoted by u^k . steam units and combustion turbines as indicated in
Table 2.
Step 2: Apply the transmission-constrained economic The transmission line parameters are given in Table
dispatch with respect to u^k to obtain a dispatch 3. In Table 3 two sets of line capacities are presented,
p^k . Let (~uk+1; p~k+1) (^uk ; p^k ): the one with smaller line capacities results in more se-
rious congestion than the other. A 24 hour planning
Step 3: If the total cost of (~uk+1; p~k+1 ) is no better than horizon is adopted in the test problem, the system load
that of (~uk ; p~k ), stop; otherwise k k +1, and information can be found in Tables 4. The reader can
go to Step 1. nd that the load taken here corresponding to the day
The above method works best for the cases with with peak load in the year in the IEEE test problem.
many generators in each bus. For other transmission- In this test, the spinning capacity requirements are set
constrained unit decommitment approaches, the inter- to be Rt = 1:07Dt for all t. The proposed algorithm
ested reader is referred to [15]. is applied to solve this test problem. The test result is
summarized in Table 5.
In case (i) of the transmission line capacity F` , the
4. Numerical results test result shows that line 33 is always congested and
line 11 becomes congested during peak hours. In this
The transmission-constrained unit commitment al- test case, it is noted that no combustion turbine is
gorithm presented in this paper has been implemented turned on during the planning horizon. The duality
in FORTRAN on an HP 700 workstation. This section gap is 0:47% in this case. Compared with the total
presents numerical results of some test problems. cost in the unconstrained case, the increased cost due
We rst apply the proposed algorithm to solve the to the transmission network is about 0.17% of the total
test problem in [13]. Not knowing the spinning reserve cost of the unconstrained case. In the second case (ii) of
requirements set in [13], we set the spinning capacity F` , the congestion of the network involves seven lines.
requirements to be Rt = 1:07Dt for all t in (4). The There are 12 hours on the planning horizon in which
result along with that in [13] is summarized in Table there are 5 lines congested (during the peak hours), 7
hours with 4 lines, 4 hours with 2 lines and 1 hour with 4.2. Direct and indirect approaches
1 line congested. The increased cost due to the trans- In [13], two methods, the direct and indirect meth-
mission system constraints increases to 1.5% of the to- ods, are dened. Those methods which take transmis-
tal cost of the unconstrained case. The congestion is sion constraints into account in the optimization stage
much more serious than in the case (i). It can be ob- are called direct methods and those which do not are
served that when the transmission network gets more called indirect. Indirect methods ignore the transmis-
congested, the algorithm takes longer time to converge, sion constraints in the dual optimization, and deal with
especially in the dual optimization and in obtaining a them only in the feasibility phase. In this section, we
feasible solution in the feasibility phase, with larger du- shall compare two algorithms. Using the same nam-
ality gap. We also found that the duality gap of 1.43% ing convention as in [13], we will call them the direct
in this test case is primarily due to the poor dual ob- approach and the indirect approach. The direct ap-
jective value obtained. In another test run, tripling the proach is the three-phase algorithm proposed in this
CPU time spent on the dual optimization improves the paper. The indirect approach is essentially the direct
dual value to 902758 with the primal value reduced to approach except that the multiplier t` = t` = 0, for
912576 for the duality gap of 1.09%. Overall, our ex- 8t; ` through out the algorithm.
perience show that our feasibility phase is very robust The indirect approach ignores the transmission con-
in obtaining feasible solutions. straints by setting the corresponding multipliers to be
zeros. During the feasibility phase, transmission con-
straints are enforced through (19d). The major moti-
vations for applying the indirect approach are to speed
up the dual optimization, and to save memory space
(the direct method essentially requires an additional
2T  (1 + L) multipliers). The indirect approach is
Table 2: Generating unit parameters
bi typey pmini pmax
i xi0 tup tdown
i tcold
i ai0 ai1 ai2 Si
1.00010,6
i
11 C 4 20 1 1 1 2 63.999 20.000 40
12 C 4 20 2 1 1 2 63.999 20.000 1.00010,6 40
13 S2 10 76 3 6 6 12 133.919 16.193 1.50810,2 45
14 S2 10 76 -2 6 6 12 133.919 16.193 1.50810,2 45
21 C 4 20 1 1 1 2 63.999 20.000 1.00010,6 40
22 C 4 20 2 1 1 2 63.999 20.000 1.00010,6 40
23 S2 10 76 10 6 6 12 133.919 16.193 1.50810,2 45
24 S2 10 76 -2 6 6 12 133.919 16.193 1.50810,2 45
71 S1 15 100 10 10 10 20 199.124 12.468 1.53210,2 45
72 S1 15 100 10 10 10 20 199.124 12.468 1.53210,2 110
73 S1 15 100 10 10 10 20 199.124 12.468 1.53210,2 110
13 1 S1 20 197 12 12 12 24 209.546 13.928 2.08510,3 100
13 2 S1 20 197 12 12 12 24 209.546 13.928 2.08510,3 100
13 3 S1 20 197 12 12 12 24 209.546 13.928 2.08510,3 100
15 1 S1 3 12 4 2 2 4 21.145 16.193 9.55310,2 30
15 2 S2 3 12 -9 2 2 4 21.145 16.193 9.55310,2 30
15 3 S1 3 12 4 2 2 4 21.145 16.193 9.55310,2 30
15 4 S1 3 12 3 2 2 4 21.145 16.193 9.55310,2 30
15 5 S1 3 12 4 2 2 4 21.145 16.193 9.55310,2 30
15 6 S2 20 155 -20 12 12 24 275.606 12.360 8.89810,3 100
16 1 S2 20 155 5 12 12 24 275.606 12.360 8.89810,3 100
18 1 S3 40 400 100 48 48 60 577.537 14.253 7.36510,4 440
21 1 S3 40 400 100 48 48 60 577.537 14.253 7.36510,4 440
22 1 S2 10 76 -2 6 6 12 133.919 16.193 1.50810,2 45
22 2 S2 10 76 -2 6 6 12 133.919 16.193 1.50810,2 45
22 3 S2 10 76 -2 6 6 12 133.919 16.193 1.50810,2 45
22 4 S2 10 76 -2 6 6 12 133.919 16.193 1.50810,2 45
22 5 S2 10 76 -2 6 6 12 133.919 16.193 1.50810,2 45
22 6 S2 10 76 -2 6 6 12 133.919 16.193 1.50810,2 45
23 1 S2 20 155 5 12 12 24 275.606 12.360 8.89810,3 100
23 2 S2 20 155 3 12 12 24 275.606 12.360 8.89810,3 100
23 3 S2 35 350 11 24 24 36 517.669 11.892 5.22010,3 250
y S1: Steam: fossil-oil; S2: Steam: fossil-coal;
S3: Steam: nuclear; C: Combustion turbine
applied to solve the IEEE test problem given in the
previous section. The result is summarized in Table Table 4: Load information
6. It can be seen in Table 6 that the direct approach
obtains a better solution and uses less CPU time. The System load Bus load
indirect approach only uses less CPU time in the dual hour load (MW) % of peak bus % of peak
optimization phase (Ph 1), but spends more time to 1 2223.0 78 1 3.8
2 2052.0 72 2 3.4
locate a feasible solution. Also the quality of the solu- 3 1938.0 68 3 6.3
tion obtained from the feasibility phase of the indirect 4 1881.0 66 4 2.6
approach tends to be worse than that from the direct 5 1824.0 64 5 2.5
6 1852.5 65 6 4.8
approach, so that it might take longer for unit decom- 7 1881.0 66 7 4.4
mitment to improve the solution. Our comparison be- 8 1995.0 70 8 6.0
tween the direct and indirect approaches agrees with 9 2280.0 80 9 6.1
what is observed in [13]. 10 2508.0 88 10 6.8
11 2565.0 90 11 0
12 2593.5 91 12 0
13 2565.0 90 13 9.3
14 2508.0 88 14 6.8
Table 3: Transmission line parameters 15 2479.5 87 15 11.1
` From To X (i) F` (ii) F` 16 2479.5 87 16 3.5
Bus Bus (p.u.) (MW) (MW) 17 2593.5 91 17 0
1 1 2 0.0139 175 175 18 2850.0 100 18 0
19 2821.5 99 19 6.4
2 1 3 0.2112 175 175 20 2764.5 97 20 4.5
3 1 5 0.0845 175 175 21 2679.0 94 21 0
4 2 4 0.1267 175 175 22 2622.0 92 22 0
5 2 6 0.1920 175 175 23 2479.5 87 23 0
6 3 9 0.1190 175 175 24 2308.5 81 24 0
7 3 24 0.0839 400 175
8 4 9 0.1037 175 175
9 5 10 0.0883 175 175
10 6 10 0.0605 175 175
11 7 8 0.0614 175 175
12 8 9 0.1651 175 175 problem with more than 2500 lines, 2200 buses and 79
13 8 10 0.1651 175 175 dispatchable generating units is derived with necessary
14 9 11 0.0839 400 175
15 9 12 0.0839 400 136 modications to t the scope discussed in this paper.
16 10 11 0.0839 400 200 The peak load is 3957 MW. The test result is given
17 10 12 0.0839 400 250 in Table 7. We emphasize that these test results used
18 11 13 0.0476 400 198 a program developed for research purpose with little
19 11 14 0.0418 400 250
20 12 13 0.0476 400 150 eort spent in speeding up the algorithm performance.
21 12 23 0.0966 400 166 The motivation is to see how the algorithm performs on
22 13 23 0.0865 400 250 a large scale problem. We observe: (i) the convergence
23 14 16 0.0389 400 220 of the dual optimization is not directly aected by the
24 15 16 0.0173 400 250
25 15 21 0.0245 400 290 increased number of multipliers corresponding to the
26 15 24 0.0519 400 270 transmission constraints but by the increased number
27 16 17 0.0259 400 270 of nonzero multipliers. The more congested the net-
28 16 19 0.0231 400 270 work is, the more iterations are required to reach a
29 17 18 0.0144 400 270
30 17 22 0.1053 400 270 near-optimal dual solution. We observe that the dual
31 18 21 0.0129 400 270 objective value tends to improve more slowly as the
32 19 20 0.0198 400 198 congestion of the network increases. (ii) In such a
33 20 23 0.0108 300 270
34 21 22 0.0678 400 270 large scale test problem the most signicant CPU time-
consuming step occurs in the calculation of the norm
of the subgradient g in (16) at each iteration. (iii)
The CPU time required in this test problem, although
4.3. A large scale test problem high, scales approximate linearly with problem size.
Algorithm performance can be improved by taking ad-
The proposed three-phase algorithm is also applied vantage of the sparsity of any large matrices including
to a test problem based on PG&E's system. The test the matrix of the distribution factors , and the mul-
Table 5: Test result of IEEE test problem (direct ap- determined. Finally the multi-area unit decommitment
proach) method is applied, which serves as a post-processing
method of the algorithm.
Case Dual Primal Duality In limited numerical tests, the proposed algorithm
Value Value Gap is found to be ecient. A large scale problem is
Unconstrained $898619 $898683 0.00%
Constrained F` (i) $895980 $900206 0.47% also tested. The result suggests that the proposed
Constrained F` (ii) $899662 $912650 1.43% algorithm can be used to deal with practical-sized
Case CPU Time (sec) transmission-constrained unit commitment problems.
Ph 1 Ph 2 Ph 3 Total
Unconstrained 3.72 0.37 2.32 6.41
Constrained F` (i) 7.81 5.77 6.28 19.86
Constrained F` (ii) 7.79 12.14 10.92 30.85 6. Acknowledgement
Table 6: Test result of the indirect approach This work was partly supported by the National
Science Foundation under Grant IRI-9120074 and
Case Dual Primal Duality PG&E's R&D Department. Their support is greatly
Value Value Gap appreciated.
Constrained F` (i) $895840 $901289 0.60%
Constrained F` (ii) $895840 $913542 1.97%
Case CPU Time (sec)
Ph 1 Ph 2 Ph 3 Total References
Constrained F` (i) 5.69 21.74 10.45 37.88
Constrained F` (ii) 5.80 18.72 9.52 34.04 [1] P. F. Albrecht et al \IEEE reliability test system",
IEEE Trans. Power Apparatus and Systems, vol.
Table 7: Test result of a large scale problem PAS-98, no. 6, pp. 2047-2054, Nov./Dec. 1979.
Total cost ($) CPU time (sec) [2] M. S. Bazaraa and C. M. Shetty, Nonlinear Pro-
Dual optimal: 199216.56 Ph. 1: 1292.47
Primal optimal: 200214.32 Ph. 2: 116.60 gramming - Theory and Algorithms, John Wiley
Duality gap: 1.469% Ph. 3: 108.32 & Sons Inc., New York, 1979.
Total: 1512.39
[3] D. P. Bertsekas, G. S. Lauer, N. R. Sandell, Jr.,
and T. A. Posbergh, \Optimal short-term schedul-
tipliers associated with the transmission constraints. ing of large-scale power systems," IEEE Trans.
This can be shown to be equivalent to reducing a large Automatic Control, vol. AC-28, no. 1, pp. 1-11,
scale `bus-based' model to a small sized `area-based' 1983.
one without aecting problem optimality and greatly
speeding up the algorithm performance. [4] W. G. Chandler, P. L. Dandeno, A. F. Gilmn
and L. K. Kirchmayer, \Short-range operation
of a combined thermal and hydroelectric power
5. Conclusions system", AIEE, Transactions, paper no. 53-326,
1953.
In this paper we presented a transmission-
constrained unit commitment algorithm using the La- [5] A. Cohen and V. Sherkat, \Optimization-based
grangian relaxation approach. The transmission con- methods for operations scheduling," Proceedings
straints are modeled as linear constraints under the of IEEE, vol. 75, no. 12, pp. 1574-1591, 1987.
DC power ow model. The Lagrangian relaxation ap- [6] L. A. F. M. Ferreira, T. Andersson, C. F. Im-
proach relaxes not only the demand constraints and the parato, T. E. Miller, C. K. Pang, A. Svoboda, and
spinning reserve constraints, but also the transmission A. F. Vojdani, \Short-term resource scheduling in
constraints using multipliers. This algorithm contains multi-area hydrothermal power systems," Electri-
three phases. In Phase 1, the dual optimization is per- cal Power & Energy Systems, vol. 11, no. 3, pp.
formed to determine a commitment, based on which 200-212, 1989.
a feasibility phase follows to locate a feasible commit-
ment. The feasibility phase is essentially an extension [7] X. Guan, P. B. Luh, and H. Yan, \An
of that in the single area unit commitment case. A optimization-based method for unit commit-
reserve-feasible commitment is rst located, then by ment," Electrical Power & Energy Systems, vol.
solving linear programs a dispatchable commitment is 14, no. 1, pp. 9-17, February 1992.
 [8] F. N. Lee, \Short-term thermal unit commitment
- a new method", IEEE Trans. Power Systems,
vol. 3, no. 2, pp. 421-428, May 1988.
[9] F. N. Lee, J. Huang and R. Adapa, \Multi-area
unit commitment via sequential method and a DC
power ow network model", IEEE Trans. Power
Systems, vol. 9, no. 1, pp. 279-287, February 1994.
[10] C. A. Li, R. B. Johnson, and A. J. Svoboda,
\A new unit commitment method", IEEE Trans.
Power Systems, vol. 12, no. 1, pp. 113-119, Feb
1997.
[11] C. K. Pang and H. C. Chen, \Optimal short-term
thermal unit commitment", IEEE Trans. Power
Apparatus and Systems, vol. PAS-95, no. 4, pp.
1336-1342, July/August 1976.
[12] C. K. Pang, G. B. Sheble and F. Albuyeh, \Eval-
uation of dynamic programming based methods
and multiple area representation for thermal unit
commitments", IEEE Trans. Power Apparatus
and Systems, vol. PAS-100, no. 3, pp. 1212-1218,
March, 1981.
[13] J. J. Shaw, \A direct method for security-
constrained unit commitment", IEEE Trans.
Power Systems, vol. 10, no.3, pp. 1329-1342, Au-
gust, 1995.
[14] A. J. Svoboda, C. L. Tseng, C. A. Li, and R.
B. Johnson, \Short term resource scheduling with
ramp constraints", IEEE Trans. Power Systems,
vol. 12, no.1, pp. 77-83, Feb 1997.
[15] C. L. Tseng, \On Power System Generation Unit
Commitment Problems", Ph.D. Thesis, Depart-
ment of Industry Engineering and Operations Re-
search, University of California at Berkeley, 1996.
[16] C. L. Tseng, S. S. Oren, A. J. Svoboda and R.
B. Johnson, \A unit decommitment method in
power system scheduling", Electrical Power & En-
ergy Systems, vol. 19, no. 6, pp. 357-365, 1997.
[17] A. J. Wood and B. F. Wollenberg, Power Genera-
tion, Operation and Control, John Wiley & Sons,
New York, 1984.
[18] F. Zhuang and F. D. Galiana, \Towards a more
rigorous and practical unit commitment by La-
grangian relaxation," IEEE Trans. Power Sys-
tems, vol. PWRS-3, no. 2, pp. 763-770, 1988.