IEEE Transactions on Power Sy?tems,,Vol. 13, No.

3, August 1998 857

Dual Programming Methods for

Large-scale Thermal Generation Scheduling
G. Demartini T.R. De Simone G.P. Granelli M. Montagna K. Rob0
ENEL S.p.A., DSNCRA Member, IEEE Member, IEEE CESI S.p.A., Milan, Italy
Cologno Monzese, Italy University of Pavia, Italy

Abstract - The problem of thermal generation scheduling: is account thermal generation ramp-rate limits neglected in the
considered in the framework of the short-term hydro-thermal model of [ 11.
coordination problem. Dual programming methods are applied In the present paper dual programming methods are used in
to the large-scale problem deriving from a fine subdivision of the the solution of the thermal scheduling subproblem incorporat-
daily optimization horizon for networks with hundreds of tber- ing ramp-rate constraints. Large-scale problems are consid-
mal units. The starting point for the dual approach is obtained
from the solution of a thermal scheduling problem with dis- ered deriving from a fine subdivision of the optimization hori-
carded generation ramp-rate constraints. The relaxed daily zon for networks with hundreds of thermal units. In the past
scheduling decouples into as many smaller dispatch problemi as primal methods were proposed to solve the thermal generation
the number of subintervals. Two dual programming methods are dispatch problem. Optimization techniques included, among
implemented: the former is the dual active set algorithm by others, penalty function methods, dynamic programming, re-
Goldfarb and Idnani while the latter is based on the application cursive nonlinear programming and gradient projection algo-
of continuation method techniques. These approaches are exten- rithms [4]-[ 81.
sively tested with reference to both a small sample system and to The choice of using dual programming is suggested by the
the daily thermal generation scheduling of the Italian (ENISL) fact that it is easy to find an optimal solution for the thermal
system (over 100 thermal units and 96 quarter hour subinter-
scheduling subproblem after relaxing the ramp-rate con-
vals). Incorporating the dual programming approach within the
ENEL hydro-thermalcoordination procedure is also considered. straints. Indeed, the relaxed problem can be solved as a se-
quence of independent thermal dispatch problems each corre-
sponding to one subinterval of the optimization time horizon.
I. INTRODUCTION Then dual programming is employed to enforce violated con-
straints while maintaining optimality.
The problem of hydro-thermal generation scheduling is lone A further reason in favour of the dual programming ap-
of the most challenging large-scale optimization probhem proach is that the code implementing the relaxed problem so-
arising in power system analysis. Most of the proposed ap- lution was already available being coincident with the thermal
proaches are based on some form of decomposition between scheduling procedure of [l]. Moreover, the number of vio-
the hydro and thermal scheduling phases. The current practice lated ramp-rate constraints is expected to be limited in normal
at the Italian Electricity Board (ENEL) is described in [I]. operating conditions, considering also that hydro scheduling
The thermal scheduling subproblem consists in minimizing has the effect of flattening the residual thermal load.
the cost of thermal generation corresponding to a given hydro Two dual programming approaches are considered. The
scheduling. Constraints include power balance equations and former implements the method of Goldfarb and Idnani [9]
operating limits on the transmission and thermal generation which consists in taking gradient projection steps which re-
subsystems. The hydro scheduling subproblem minimizes the duce the amount of violation while retaining the optimality of
thermal generation cost subject to a set of hydraulic con- the current solution point. The latter approach makes use of
straints by employing the Frank-Wolfe algorithm [2].Coordi- continuation method techniques; a whole family of optimal
nation between the two subproblems is achieved by exploiting solution points is generated as a function of a continuation pa-
the Lagrange multipliers of the thermal scheduling phase. A rameter whose value is progressively increased from zero
similar approach can be found in [3]. (corresponding to the optimal relaxed solution) to one
The short-term hydro-thermal scheduling implemented at (corresponding to the optimal solution considering all con-
ENEL typically considers a daily optimization horizon dis- straints).
cretized on an hourly basis. A finer subdivision of the optimi- Both algorithms were applied to power system analysis
zation interval (e.g. every 15 minutes) requires to take into problems. The Golfarb-Idnani method was recently employed
in the framework of probabilistic adequacy assessment of
PE-068-PWRS-I 6-09-1997 A paper recommended and approved by large interconnected systems [lo]. Continuation methods were
the lEEE Power System Engineering Committee of the IEEE Power applied to the optimal power flow [ 111 and hydro-thermal co-
Engineering Society for publication in the IEEE Transactions on Power ordination [ 121 problems.
Systems. Manuscript submitted May 27, 1997: made available! for The characteristics of the two approaches are compared
printing September 30, 1997. with reference to a small sample system. More significant re-
sults are presented for the ‘ENEL production system with more
than 100 thermal and 200 hydro units.

088.5-8950/98/$10.00 0 1997 IEEE

858

11. THE HYDRO-THERMAL SCHEDULING PROBLEM The hydro scheduling subproblem is defined as:

Hydro-thermal scheduling requires to minimize the genera-

SJY
[
min FH HL ( q i )]
tion cost of thermal units while satisfying the operating con-
straints of the hydro, thermal and transmission subsystems. subject to a set of linear hydraulic continuity equations and to
For the short-term scheduling problem considered, it is as- upperllower bounds on water releases and storage volumes.
sumed that the results of thermal unit commitment are avail- The problem variables are hydro unit discharge rates q ,
able. spillage discharge rates s and water storage volumes V .
The problem is decomposed into the two phases of hydro
According to the Frank-Wolfe algorithm, each iteration of
and thermal scheduling. The hydro scheduling objective func-
the hydro scheduling subproblem consists in minimizing a lin-
tion FH is defined as the minimum cost of the thermal gen-
earization of FH (with respect to hydro unit discharge rates)
eration corresponding to the residual load.
subject to the hydraulic constraints. The resulting linear pro-
For a given hydro scheduling fii , the value of FH is ob- gram decouples into smaller problems each corresponding to
tained by solving the following thermal generation scheduling a system of hydraulically coupled plants. Efficient network
(TGS) subproblem: flow programming techniques are used in each linear pro-
gramming solution.
F,(k;)= A solution of TGS is necessary to evaluate the derivatives
of FH which are obtained as a linear combination of the La-
grange multipliers corresponding to the power balance equa-
tions (1) and to the active constraints of (2). Other TGS solu-
tions are needed to evaluate FH in the linear search process
required by each iteration of the Frank-Wolfe algorithm.
TGS is a very large quadratic programming problem; it is
solved here by employing dual active set methods.
j=1 k=l Dual programming approaches require a (dual feasible)
starting point which is the optimal solution of a relaxed prob-
lem TGS-R obtained by discarding some constraints of TGS.
The current solution is iteratively modified so as to enforce
those of the relaxed constraints that result violated while
keeping the optimality conditions satisfied. When no more
--I < p'-I 2 p ;
pi. - J = 1,2,...,N, (3) violated constraints are left, the solution of TGS is obtained.
The choice of dual programming methods was suggested
j = 1,2,...,N, mainly by the following considerations:
(4)
0 If the thermal generation ramp-rate constraints (4) are se-
In the TGS model, the optimization horizon is subdivided lected as the ones to be relaxed, TGS-R decouples into N
into N subintervals of equal length At ; constraints (1)-(4) are separate scheduling subproblems of smaller dimensions.
0 The program performing the solution of TGS-R was al-
relative to all subintervals i = 1,2,. ..,N . $ ( ki) is the real ready available and working [ l].
power generation of thermal (hydro) unit J ( k ) and D; is the * The number of violations of constraints (4) corresponding
load at bus 1 during the i-th subinterval. N, (NH ) and NL are to the solution of TGS-R is not expected to be high in nor-
mal operating conditions.
the number of thermal (hydro) units and load buses.
Equations (1) express the balance between real hydro-ther- To reduce the number of ramp-rate violations and ease the
N . computation burden of the whole hydro-thermal coordination
mal generation and the total load demand Di =
/=I 0,' . problem, it is convenient to employ a two-stage procedure. In
Constraints ( 2 ) refer to a subdivision of the network into the first stage the problem is solved by discarding ramp-rate
tree-connected areas. A limit 31is set on the real power in- constraints. The complete TGS model is introduced only in
terchange between any two contiguous areas, with h belong- the second stage hydro-thermal dispatch that is when a re-
ing to the set of interconnections s,. Coefficients a,p and y markable flattening of the residual thermal load has already
may be 1, -1 or 0 according to the location of individual gen- been achieved.
erations and loads and to which power interchange is evalu-
ated. 111. SOLUTION OF THE TGS PROBLEM
Constraints (3) set the technical limits on power generation
for the thermal units. A . Duality Based Quadratic Programming
Constraints (4) set lowerlupper bounds y j and 5 on the A quadratic programming problem can be expressed in ma-
thermal generation ramp-rate. trix form as follows:
 859
straint i to decrease since c r h < 0. Because of linearity, the
subject to
constraints in V, will remain satisfied.
A linear search along the direction ( Ax, ) is required.
CTxI d (5) Three different step sizes so, s1 and s2 are evaluated:
where equality constraints are omitted for simplicity.
di - crxo
For the TGS problem, D = d i a g { 2 ~ , ~ }is a positive definite so = -
CTAX
diagonal matrix of dimension n = N x NT .
According to the dual programming perspective, problem d k - CEXo
TGS is approached by first solving the relaxed problem TCiS- s - min{tk}, tk =--
- kcV,
R. The solution of TGS-R satisfies the following optimality
conditions:
Dx, + Ah, = -b
ATxo= a 0 so is the step size to be taken in the direction Ax to enforce
the feasibility of the most violated constraint i .
in which x, is the generation scheduling vector and h, (with
0 s1 is the largest step size that can be taken in the Ax direc-
elements 10)is the vector of the Lagrange multipliers corre-
tion without violating any strictly satisfied constraint.
sponding to the active constraints of (5). MatrixAT (vector
Computation of tk is required only if c c h > 0 .
a ) is a partition of C (d ) corresponding to those constraints
0 s, is the largest step size that can be taken in the A?L di-
of TGS-R which are active at the solution. AT is a full rank rection before a Lagrange multiplier changes from positive
mX n matrix with m 5 n . to zero. Only active constraints with < 0 must be con-
If no ramp-rate constraints are violated, then x, is the op- sidered. In (9), A,, is the multiplier corresponding to con-
timal solution of the TGS problem as well. In general, how- straint h .
ever, some ramp-rate constraint violations will be present.
The dual programming approach consists in modifying the The step size to be taken is defined as s = min{so,sl , s 2 } .
current solution from the dual feasible starting point xo (with The varied point is obtained as: x1 = x, + SAX.The constraint
the associated multiplier vector 1,) until all violated con- sets V, ,VI ,Vv and the multiplier vector are updated according
straints are enforced while retaining optimality. to which of the step sizes so ,s,,s2 is the minimum.
Two dual programming algorithms are considered: the for- If s = so, the varied point is the solution of the following
mer implements the approach of Goldfarb and Idnani while
problem:
the latter belongs to the continuation methods.
min{ bTx+ sxTDx}
1
subject to
B. The Goldfarb-IdnaniApproach
A brief outline of the Goldfarb-Idnani (G-I) method is
given here; a complete description can be found in [9].
Corresponding to the dual feasible starting point x, , three
mutually exclusive constraint sets are defined:
If s = sl or s = s,, set V, is changed by the addition or
0 set V, contains all the active constraints; elimination of a constraint. In this case the varied point is the
set VI contains all the strictly satisfied constraints; solution of the problem:
0 set Vv contains all the violated constraints.
mini (bT+ S C ~ ) X1+ ~ X ' D X } subject to
Starting from xo, the search direction ( h , A h ) is obtained
by solving the following linear system:
i T X = i
DAX + AAL = -ci
(7) where iT and % are relative to the modified V, .
=Q
A~AX
A new iteration of the G-I algorithm is performed starting
from the varied point.
where ci is the column vector corresponding to the most vio-
The resulting G-I procedure can be stated as follows:
lated constraint i E V,.
Pre-multiplying the first equation of (7) by hTand ex-
0. Corresponding to the dual feasible starting point x,, de-
ploiting the fact that matrix D is positive definite, it results termine the constraint sets V, ,VI ,Vv; let m be the number
that a step along the direction Ax causes the violation of Icon- of active constraints.
860
1. If V, = 0, the current iterate is optimal and the procedure After solving (11), a linear search is performed to deter-
is over; otherwise, find the most violated constraint i E V, mine step sizes s, and s2 according to (8) and (9). The actual
and set r = 0. step size is taken as s = min{l,s, , s 2 } and the varied point is:
2. Determine the search direction (Ax,Ah) by solving system x1 = xo + SAX.The relaxation vector is modified in the form:
(7) relative to the current V, . zt(1-s)z.
3. Perform the linear search to find s = min{ so, sl,s 2 } If s = 1, all the violated constraints are satisfied in the var-
Update the current iterate: xo t x, + SAX ied point.
If s = s, , a new constraint enters the active set V, .
Compute: h, = h, + sAh
If s = s2, the active constraint corresponding to the vanish-
Ifs=so: h o t [ ~ ~ , r + s ~ ; m t m + l ing Lagrange multiplier is eliminated from the active set V,
Update the constraint sets: and added to V I .
V, t V, u {i}; V, t V, - {i}; return to Step I A new iteration of the CM algorithm is then performed

Ifs=s,:hot [h:,~
IT ; r t r + s ; m t m + ~
starting from the varied point.
At the first CM iteration the continuation parameter 0 is
equal to the step size s . For subsequent iterations the follow-
Let k be the index of the constraint met in the linear
search process; update the constraint sets: ing assignment is to be used: 0 t 0 + s( 1- . The algorithm
e)
V, t V , U{k};V, t V , - { k } ; r e t u r n t o S t e p 2 . terminates when 0 = 1 , that is when the right-hand-sides of the
violated constraints have been moved back to their original
Ifs=s2:rtr+s;mtm-l
values.
Let k be the index of the constraint associated with the The CM procedure can be stated as follows:
vanishing multiplier A,, ; set:
0. Corresponding to the dual feasible starting point xo, de-
h, = bel'..., hem]'
,h,,l+,,...> termine constraint sets V, ,VI ,Vv ; let m be the number of
Update the constraint sets: active constraints.
V, t V , -{k};V, t V I u { k } ; r e t u r n t o S t e p 2 . 1 . Compute vector z and modify the right-hand-sides of vio-
lated constraints; add the most violated constraint to V,
C. The Continuation Method Approach and all other violated constraints to VI ; set 0 = 0.
The idea behind the continuation method (CM) approach is 2. Determine the search direction (Ax,Ah) by solving system
that of modifying the right-hand-side of the violated con- (1 1) relative to the current V, .
straints as follows: 3. Perform the linear search to find s = min{ l,s, ,s2}
Update the current iterate:
c;x I (d, + zj,) - zj,8 (10) xo t x, + S A X ;8 t 8 + s(1- 8) ; z t (1- s)z
where c l is the row of CT corresponding to constraint k ; 0 Compute: h, = ho+ sAh
is a continuation parameter ranging between 0 and 1 ; z, is the If s = 1 (8 = 1) the current iterate is optimal and the CM
element of a positive valued relaxation vector z chosen so as procedure is over.
to satisfy all the violated inequalities for x = xo and 0 = 0. 0 Ifs=s,: h , t [ ~ ; , o r ;m t m + l
Assuming that i is the most violated constraint, the right- Let k be the index of the constraint met in the linear
hand-side modification of (10) is carried out by choosing:
search process; update the constraint sets:
zk=6 VkcV, V, t V , u { k } ; V , t V , - { k } ; r e t u r n t o S t e p 2 .
0 Ifs=s,:mtm-l
being 6 = C ~ X ,- d, the violation of constraint i .
Let k be the index of the constraint associated with the
After the modification, all the originally violated con- vanishing multiplier A,, ; set:
straints become strictly satisfied and enter set VI with the ex-
ception of the most violated constraint i which enters the ac- A, = [h,,>...>hc,,-l>&,,+l 2...> heJ
tive set V, . Update the constraint sets:
The optimality conditions for the modified problem, ex- V, t V, - { k } ; VI t VI u { k } ; return to Step 2.
pressed in incremental form, are:
D h + AAh + ciALm+,= 0 D. Comparative Considerations
ATAX = 0 The G-I and CM algorithms present substantial analogies.
CTAx = -6 The evaluation of step sizes s, and s2 is identical for both
methods. Updating the constraint sets required at each itera-
in which: x = xo + 0Ax ; h = ho+ 8Ah ; Am+, = OAh,,,. tion is quite similar. The same approach is also employed in
 861

solving the linear systems (7) and (1 1). The Ax variables are TABLE I
CHARACTERISTICS OF THERMAL
UNITS
eliminated so that a reduced coefficient matrix i T D - ' i re- -
sults (being AT the matrix of the current active constraints). Units P P CO Cl 103c,
Actually only functional constraints are considered since ac-
tive upper/lower bounds are handled implicitly. The reduced
coefficient matrix is factorized by the Cholesky method rind 2 A 140 310 1135.4 28.63 7.39
all modifications required by active set changes are carried 3 A 240 640 21 18.4 33.61 5.33
out by efficient rank-one updating of the triangular factors. 4 B 120 300 860.1 41.23 2.12
There is, however, an important difference in the way the 5 B 125 300 860.1 41.23 2.12
two methods are exposed to the risk of failure due to the sin- 6 C 200 600 2885.1 32.89 4.12
7 C 65 140 913.8 32.28 88.75
gularity of the reduced coefficient matrix. Matrix L T D - ' i 8 D 150 210 1562.1 31.11 25.10
corresponding to system (6) is non-singular because D I S a 9 E 190 300 649.7 40.34 0.02
10 E 150 300 649.7 40.41 0.02
positive definite matrix (of order n ) and LT = A T is a full
,

W is a fictional monetary unit

rank m x n matrix (with m I n).
According to the G-I and CM algorithms, the reduced co-
efficient matrix changes by effect of the elimination or addi-
tion of a constraint at a time.
Elimination of a constraint from the current active set V,
leaves the reduced coefficient matrix non-singular.
Matrix ATD-'L becomes singular if a new constraint with
a coefficient row linearly dependent on the rows of iT
is
added to the current active set V, .
Let cfx < d, be the expression of a constraint of VI with a
coefficient row c l = TiT
linearly dependent on the rows of
iT( q being a coefficient vector). According to the G-I
method, the search direction must satisfy the equation
X T A x = 0 whereby for a linearly dependent constraint
c& = 0 results. Therefore, a step in the Ax direction of the
G-I method cannot bring any linearly dependent constraint
into the active set.
When using the CM approach, the possibility that c i h re-
sults greater than zero cannot be avoided for a linearly de-
pendent constraint of VI . Assuming cf = q T i T +qIci', it
follows, by the last two equations of (Il), that C ~ A=X-q,6.
Therefore, c ; h may be positive according to the sign of qI.
In the CM approach, the problem of conflicting constraints
is handled by allowing linearly dependent constraints, met
during the continuation method procedure, to violate their
limits. Once the continuation parameter reaches 1, still vio-
lated conflicting constraints are enforced either by successive
applications of the continuation method or by the G-I algo-
rithm.
IV. TESTS AND RESULTS
The dual programming procedures considered were imple-
TL =30%
HG = 80% $11
0 4T
0
4

mented in FORTRAN using double precision arithmetic and

were tested on the Digital VAX 9000-210 computer at the
Automatica Research Center of ENEL.
The test cases refer to a small sample system with 10 ther-
mal units and to the ENEL production system with over 100
thermal units. All data required to solve the TGS problem are
given for the sample system.
862
TABLE I1 TABLE IV
DUALALGORITHM (SAMPLESYSTEM)
CHARACTERISTICS DUALALGORITHM
CHARACTERISTICS (ENELSYSTEM)
Methods Methods
G-I CM G-I CM
(3 No.of CPUtime No,of CPUtime n, (3 No.of CPUtime No.of CPUtime n,
(%) iterations (s) iterations (SI % iterations (s) iterations 6)
0.5 218 0.85 53+118 0.41+0.46 18 1 498 109.4 321+324 104.2+75.2 49
1 64 0.22 22+32 0.17+0.10 5 1.5 198 42.2 124+69 34.6+15.5 15
1.5 20 0.09 12+9 0.12+0.04 5 2 80 17.6 60+27 16.8+6.3 7
1.8 3 0.04 3 0.04 0
22176 functional constraints (96 balance equations, 768
The solution of the TGS-R problem provides the dual fea- power interchange constraints, 213 12 ramp-rate con-
sible starting point. A number of 18 functional constraints and straints).
21 1 upperllower bounds (including 120 lower bounds on arti- A number of 177 functional constraints and 10454 up-
ficial unit generations) are active at the TGS-R solution. The perllower bounds (including 960 lower bounds on artificial
number of violated ramp-rate constraints is 24, 1 1, 7 and 2 for unit generations) are active at the solution of TGS-R. The
(3= 0.5, 1, 1.5 and 1.8% respectively. number of violated ramp-rate constraints is 139, 60 and 34 for
Table I1 shows the number of iterations and CPU times re- O = 1, 1.5 and 2% respectively.
quired to perform the dual programming procedures. In the Table IV shows the number of iterations and CPU times re-
test cases with O = 0.5, 1, 1.5%the CM procedure incurred in quired to perform the dual programming procedures. In all
the conflicting constraint problem: n, linearly dependent con- test cases the CM procedure incurred in the conflicting con-
straints met during the continuation procedure are checked at straint problem. A successive G-I computation allows the en-
the end of a first application of the continuation algorithm. A forcement of all constraints.
successive G-I computation is adopted to enforce all con- The composition of the active constraint set relative to the
straints. The number of iterations and the CPU times of the TGS solution is shown in Table V. For all test cases, no arti-
two phases are shown separately in Table 11. In the test case ficial unit is loaded at the end of the computation.
with 0=0.5% it was possible to reduce the number of con- The test results prove that both the G-I and CM procedures
flicting constraints up to 11 by successive applications of the are well suited to the problem of large-scale thermal genera-
continuation algorithm. Enforcement of these last constraints, tion scheduling. G-I appears more efficient than CM regard-
however, required application of the G-I algorithm. ing CPU times. This is an outcome of the conflicting con-
The composition of the active constraint set relative to the straint phenomenon. Indeed, according to the present CM
TGS solutions is shown in Table 111. For all test cases, no ar- implementation, matrix XTDp'X is refactorized when singu-
tificial unit is loaded at the end of the computation. larity occurs. Improved efficiency is expected by replacing re-
factorization by direct modification of the current triangular
B. Tests on the Italian System factors.
The schematization of the ENEL hydro-thermal production The task of introducing the dual programming techniques
system considered for testing purposes comprises over 100 into the short-term hydro-thermal coordination procedure is
thermal and 200 hydro units. The system consists of five areas presently in progress. The two-stage computation mentioned
arranged in a tree-connected network as in Fig. 1 (with in Section I1 is being implemented. The enforcement of ramp-
A,B,C,D and E corresponding to Northern, Central, Southern rate constraints is performed only after the first stage hydro-
Italy, Sardinia and Sicily). thermal dispatch has achieved a remarkable flattening of the
The largest thermal dispatch problem tested so far is rela- residual thermal load. In this way a reduced number of ramp-
tive to a case with 111 thermal units and 96 quarter hour sub- rate violations is expected at the end of the first stage. As a re-
intervals. sult, the second stage computation should take only a few
The dimensions of the problem are: Frank-Wolfe iterations, thus reducing the CPU time overhead
needed by the dual programming procedure. The solution of
e 1 1616 variables (including artificial unit generations) with the TGS-R, required in both stages, is performed by an effi-
upperllower bounds; cient primal quadratic programming algorithm based on the
active set strategy.
TABLE I11
ACTIVES E T AT THE SOLUTION OF TGS (SAMPLE SYSTEM) TABLE V
ACTIVESET AT THE SOLUTION OF TGS (ENELSYSTEM)
Number of constraints
(3 Functional U/L bounds Number of constraints
(%) Balance Ramp-rate Interchange CJ Functional U/L bounds
0.5 12 53 1 169 (%I Balance Ramp-rate Interchange
1 12 14 6 199 1. 96 203 77 10311
1.5 12 8 6 205 1.5 96 74 79 10410
1.8 12 2 6 210 2 96 35 80 10428
 863

The two-stage procedure has been tested on some daily hy- [4] A.D. Patton, “Dynamic optimal dispatch of real power for ther-
dro-thermal scheduling problems. Reference is made here to mal generating units”, Proc. 8th PICA Con$, Minneapolis,
the situation occurred on July 25th, 1996 at the ENEL system Minnesota, June 1973, pp. 403-411.
featuring 85 committed thermal units (with 0 = 2 % ) and 222 [5] D.W. Ross and S. Kim, “Dynamic economic dispatch of genera-
tions”, IEEE Trans. Power Appar. & Syst., Vol. PAS-99, No. 6,
hydro units (arranged into 8 1 hydraulically coupled systems). NovembedDecember 1980, pp. 2060-2068.
A complete hydro-thermal coordination study (with 96 quarter [6] P.P.J. Van den Bosch, “Optimal dynamic dispatch owing to
hour subintervals) required a total CPU time of 150 s. The spinning-reserve and power rate limits”, IEEE Trans. Power
first stage took 9 Frank-Wolfe iterations and a CPU time of Appar. & Syst., Vol. PAS-104, No. 12, December 1985, pp.
95 s, while the second stage required only 2 Frank-Wolfe it- 3395-3401,
erations and a CPU time of 55 s. Execution of the G-I dual [7] M. Innorta, P. Marannino, G.P. Granelli, M. Montagna and A.
programming procedure took about 80% of the second stage Silvestri, “Security constrained dynamic dispatch of real power
CPU time. for thermal groups”, 1EE.E Trans. Power Syst., Vol. 3, No. 2,
May 1988, pp. 774-781.
V. CONCLUSIONS [8] P. Marannino, G.P. Granelli, M. Montagna and A. Silvestri,
“Different time-scale approaches to the real power dispatch of
In the present paper dual programming algorithms are ap- thermal units”, IEEE Traizs. Power Syst., Vol. 5, No. 1, Febru-
plied to the problem of large-scale thermal generation ary 1990, pp. 169-176.
scheduling arising in the framework of short-term hydro- [9] D. Goldfarb and A. Idnani, “A numerically stable dual method
thermal coordination. The choice of dual programming tech- for solving convex quadratic problems”, Mathematical Pro-
niques is suggested by the availability of an efficient code for gramming, Vol. 27, 1983, pp. 1-33.
the solution of the thermal scheduling problem with relaxed [lo] 0. Bertoldi, M.V. Cazzol, A. Garzillo and M. Innorta, “A dual
quadratic programming algorithm oriented to the probabilistic
thermal generation ramp-rate constraints. Moreover, the num- analysis of large interconnected networks”, Proc. 12th PSCC,
ber of ramp-rate constraint violations is not very large in nor- Dresden, Germany, August 19-23, 1996, pp. 1249-1256.
mal operating conditions. M Huneault, A. Fahmideh-Vojdani, M. Juman, R. Calderon and
Two algorithms are compared: the former derives frorn a F.D. Galiana, “The continuation method in power system opti-
work by Goldfarb and Idnani, while the latter is based on mization: applications to economy-security functions”, IEEE
continuation method techniques. Trans. Power App. & Syst., Vol. PAS-104, No. 1, January
The two approaches present significant similarities and are 1985, pp. 114-124.
roughly equivalent from the point of view of computation ef- L.R. Calderon and F.D. Galiana, “Continuous solution simula-
tion in the short-term hydrothermal coordination problem”,
ficiency. However, the G-I approach is immune to the prob- IEEE Trans. Power Syst., Vol. 2, No. 3, August 1987, pp. 737-
lem of having conflicting constraints in the current active set. 743.
With CM, the conflicting constraint problem is solved by suc-
cessive applications of dual programming. Conflicting con-
straints are allowed to violate their limits at a first stage; vio- Giovanni Demartini received his Dr. Ing. degree in Electrical Engi-
lated constraints are successively enforced either by the con- neering from the University of Pavia, Italy in 1974. Since 1975 he
tinuation approach or by the G-I algorithm. has been with the Italian Electricity Board (ENEL). Presently he
The two methods are tested on a small sample system, as leads the “Methodology and Tools for the Electric System Opera-
well as on a large-scale problem corresponding to a quarter tion” group at the Automatica Research Center of ENEL, Cologno
Monzese, Italy.
hour discretization of the daily thermal scheduling for the
ENEL production system (about 12,000 variables and 22,000 Tomaso Roberto De Simonr! received his Dr. Ing. degree in Elec-
functional constraints). Incorporating the dual programming tronic Engineering from the Politecnico di Milano, Italy in 1992.
algorithms within a hydro-thermal coordination procedure is From 1992 to 1994 he was with Landis & Gyr, Milan, Italy. In 1994
he joined the “Methodology and Tools for the Electric System Op-
also considered.
eration” group at the Automatica Research Center of ENEL, Colog-
ACKNOWLEDGMENT no Monzese, Italy.
This work was performed with the financial support of the Gianpietro Granelli (M’86) received his Dr. Ing. degree in Electri-
Automatica Research Center (CRA) of ENEL. The autlhors cal Engineering from the Pcilitecnico di Milano, Italy in 1968. In
thank Dr. Mario Innorta of ENEL for helpful suggestions. 1970 he was appointed Researcher and successively Assistant Pro-
fessor at the Department of Ellectrical Engineering of the University
REFERENCES of Pavia, Italy where he is now a Professor of Electrical Engineering.
Mario Montagna (M’9.5)received his Dr. Ing. degree in Electrical
[I] P. Burelli, G. D’Amora, M. Innorta and P Marannino, “Short Engineering from the University of Pavia. Italy in 1982. Since 1983,
term hydro-thermal generation scheduling with inter-area power he has been a University Researcher of Electric Power Plants at the
flow constraints”, Proc. 10th PSCC, Graz, Austria, August Department of Electrical Engineering of the University of Pavia.
1990, pp. 1129-1 135.
[2] B. Martos, Nonlinear Programming Theory and Metrzods, Kostandin Robo received his Degree in Electrical Engineering from
North-Holland, Amsterdam, 1975, pp 240-248. the University of Tirana, Albania in 1973. In 1973 he joined the Al-
[3] H. Hdbibollahzadeh, D Frances and U. Sui, “A new generation banian Electricity Board (KESH). l n 1991 he received his Ph.D. in
scheduling program at Ontario Hydro”, IEEE Trans. Power Electrical Engineering from the Albanian Academy of Science.
Syst., Vol. 5, No. 1, February 1990, pp. 65-73. Since 1992 he has been with CESI, Milan, Italy.