See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3265929

Short-Term Hydrothermal Coordination by Lagrangian Relaxation: Solution

of the Dual Problem

Article  in  Power Systems, IEEE Transactions on · March 1999

DOI: 10.1109/59.744490 · Source: IEEE Xplore

CITATIONS READS

194 703

2 authors, including:

Noemi Jimenez-Redondo
CEMOSA
31 PUBLICATIONS   404 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Acem-rail View project

Ebalance-plus (www.ebalanceplus.eu) View project

All content following this page was uploaded by Noemi Jimenez-Redondo on 18 April 2017.

The user has requested enhancement of the downloaded file.

IEEE Transactions on Pc er Systems, Vol. 14, No. 1, February 1999 89
SH ORT-TE .M HYDRO-THERMAL COORDINATION BY LAGRANGIAN
R LAXATION: SOLUTION OF THE DUAL PROBLEM

N. Jimknez Redondo A. J. Conejo

1 F.S. Ingenieros Industriales E.T.S. Ingenieros Industriales
Universidad de MBlaga Universidad de Castilla - La Mancha
MBlaga, Spain Ciudad Real, Spain

Abstract - This pap addresses the short-term hydro- subproblem per thermal plant and one subproblem per hy-
thermal coordination roblem. This problem is large- droelectric system. This decomposition allows a precise
scale, combinatorial a 1 nonlinear. It is usually solved modeling of the generation system resulting in a very ac-
using a Lagrangian R axation approach. In the f r a m e curate solution technique.
work of the Lagrangi Relaxation, this paper provides
Instead of solving the original problem, the LR technique
a novel, non-oscillatin and efficient multiplier updating solves its dual problem. The difference between the ob-
procedure. This proce ire advantageously compares with
jective function optimal values for the primal and dual
previously reported p cedures such us subgradient and problems is called the duality gap. I n most practical cases
bundle methods. A re stic large-scale case study is used
the duality gap is below 0.1%. As a byproduct of the solu-
to illustrate the behav r of the proposed procedure. tion of the dual problem a solution for the primal problem
Keywords: Short-t n Hydro-thermal Coordination, is obtained. However, more often than not, this primal
Large-scale Optimizat n, Lagrangian Relaxation, Multi- solution is infeasible. Heuristic procedures are required to
plier Updating. get a feasible primal solution.
1 Introduction The LR technique consists of the three phases below:
This paper addresses e short-term hydro-t hermal coor-
dination problem. Thi Iroblern is solved to determine the Phase 1. To solve the dual problem.
start-up and shut-dow schedule of thermal plants, as well
as the power output o hermal and hydro plants during a Phase 2. To obtain a primal feasible solution.
short-term planning hc zon. The objective is to meet cus-
tomer demand with a] ropriate levels of spinning reserve Phase 3. To exactly dispatch committed generation to
so that total operatin) are minimized. meet the demand.
This is a large-scale rr ed-integer nonlinear problem. As
recognized in the tech :a1 literature, the Lagrangian Re- Phase 2 is easily accomplished as stated in most previous
laxation (LR) techniq is the most promising procedure LR works [2, 41. Heuristic techniques are successfully ap-
to solve the short-terr hydro-thermal coordination prob- plied to obtain a primal feasible solution, i.e. a primal
lem [l,2, 3, 4, 5, 6, 7, . Dynamic programming requires feasible set of commitment decisios. Phase 3 is a multi-
discretization and dras c simplifying assumptions to make period economic dispatch whose solution is well stated in
the problem computal nally tractable [9]. Mixed integer the technical literature [13].
linear programming re tires linearization and the solution
However, Phase 1 requires the solution of a non-
of large-scale mixed-in ger linear problems which is not a
trivial task [lo, 11, 12 differentiable maximization problem which is not an easy
task. The procedures reported in the literature to solve
The LR procedure dec nposes the original problem in one this type of problems are either oscillating or computa-
tionally inefficient, or both, oscillating and computation-
PE-333-PWRS-0-12-1997 L paper recommended and approved by ally inefficient.
the IEEE Power Systen Analysis, Computing and Economics This paper focuses on Phase 1. It provides a solution tech-
Committee of the IEEE POM Engineering Society for publication in the nique which is non-oscillating, computationally efficient
IEEE Transactions on Pov Systems. Manuscript submitted July 1,
and which does not rely on “messy” heuristics.
1997; made available for pri ng December 12,1997.
The most commonly used technique to address Phase 1 is
the subgradient [2, 4, 5, 61 which is computationally inef-
ficient, and produces oscillating solutions which makes it
complicated to define appropriate stopping criteria. The
cutting plane technique [14], under mild assumptions, con-
verges to the optimum but makes it very slowly. The re
cently proposed primal bundle method [15, 161 is a cutting

0885-8950/99/$10.00 0 1997 IEEE

90

plane method whose objective function has been penal-

ized to control the path to the optimum. This penalty
technique is complicated and requires careful tune-up of
penalty parameters to be efficient. The required tune-up
is problem-dependent and difficult to implement. The dual
bundle method [16, 171 exhibits the same type of problems
as the primal bundle method. However, it is a less elab-
orated procedure than the primal one. It is close to the
subgradient method, and therefore it is prone to oscilla-
tions. T h e technique proposed in this paper is a cutting
plane method which incorporates adaptive control over
the feasible region. This is easily accomplished because
the feasibility region is simple: just bounds on multipliers. where (z, y) = (z,, y,; Vz, V J ) , variables z2 are the vari-
This adaptive control can be performed without involved ables related to thermal plant 2, and vJ are the variables
parameter tune-up, resulting in a non-oscillating, compu- related to the hydroelectric system 3 . G, g,(z,), g3(y,),
tationally efficient algorithm. H I h,(z,) and h,(y,) are vectors of dimension equal to the
number of subperiods in the planning horizon.
In the framework of deregulated electric energy markets,
the LR technique has a well known and simple yet elegant Equation (1) is the production cost to b e minimized, equa-
economic interpretation (see, for instance, [IS]) which is tions (2) are the constraints related to every thermal unit,
described below. Hourly energy price proposals are speci- equations (3) are the constraints related to every hydro-
fied by the market operator for the planning horizon (e.g. electric system, equation (4)enforces spinning reserve con-
the 168 hours of one week). Each generator schedules in- straints, and equation ( 5 ) enforces demand constraints. It
dependently its production along the planning horizon t o should be noted that time is embedded in the above for-
maximize its benefit. Analogously, each hydroelectric sys- mulation.
tem is scheduled so that its benefit is maximum. After Equations (4) and (5) are global constraints which couple
the submission of production proposals by all generators together thermal-related and hydro-related variables.
the power balance equation is evaluated in each hour. Ex-
cess of committed power will occur in some hours, whereas Problem (1)-(5) is denominated the primal problem (PP).
deficit will occur in some other hours. Hourly prices are Global constraints are incorporated in the objective func-
updated by the market operator and the previous proce- tion of the primal problem through Lagrangian multipliers
dure repeated until the demand is satisfied. The above to obtain the Lagrangian function:
mechanism constitutes a competitive energy market. Sim-
ilarly, a spinning reserve market can be established.
The price updating or multiplier updating is in fact a so-
lution procedure for the dual problem. If multipliers are
updated proportionally to the power mismatch, the sub-
gradient technique is obtained. If the updating procedure
keeps track of the information of all steps to reconstruct where A and p are vectors of dimension equal to the num-
total cost as a function of multiplier values, the cutting ber of subperiods in the planning horizon. T h e dual prob-
plane method is obtained. By proper modification of the lem (DP) of the original primal problem (1)-(5) has the
objective function of the problem to be solved when using form
the cutting plane technique, the bundle method is derived.
And finally, by proper adaptive modification of the feasi-
bility region of the problem to be solved when using the
cutting plane technique, the method proposed in this pa-
per is derived.
This paper is organized as follows. Section 2 gives a formu-
lation of the short-term hydro-thermal coordination prob-
lem. Section 3 provides a novel approach to update the where B = (A, p ) , and d ( A , p ) is the solution of the decom-
multipliers in the framework of a LR solution procedure. posed primal problem (DPP) stated below
This updating procedure is compared with previous pro-
cedures. Section 4 includes two case studies. The first one
is designed to compare the behavior of the different proce-
dures to update multipliers, the second one is a large-scale Ci
case study based on the electric energy system of mainland
Spain. Section 5 provides conclusions.
2 Problem formulation subject to
The short-term hydro-thermal coordination problem can
 91
The above problem de Imposes in one subproblem per [20]. However, it retains the same drawbacks as the origi-
thermal unit and one E bproblem per hydroelectric sys- nal subgradient procedure.
tem.
3.2 Cutting plane method (CP)
The subproblem associa ?d to thermal unit i is
The updated multiplier vector is obtained solving the lin-
Minimize,, fi(xi) - ear programming problem below.
subject to Maximize,,e, c z (20)
and the subproblem ass ciated to hydroelectric system j
is
subject to +
z 5 d ( k ) ,(k)T(B - dk)) (21)
k = l , . . ., v
Maximizeyj XThj(y”
where C is a convex and compact set. It is made up of the
subject to Sj( ranges
- of variation of the multipliers, i.e. C = {e, 5 0 5
The Lagrangian Relaxs on procedure to solve the dual 6).
problem works as follow It should be noted that equation (21) represents a half-
space (hyperplane) on the multiplier space.
1. Initialize multiplier pector B = (x,p).
It should also be noted that the number of constraints of
2. Solve the decompo ?d primal problem by solving 1 the problem above grows with the number of iterations.
subproblem per t h ,mal plant (13)-(14) and 1 sub- The above problem (20)-(21) is a relaxed dual problem
problem per hydroc xtric system (15)-(16). which gets closer to the actual dual problem as the number
of iterations grows.
3. Update the multipl ‘r vector using any of the proce
dures stated below Section 3). The CP method achieves a dual optimum by “reconstruct-
in$ the dual function. It reconstructs the region of inter-
4. If the convergence iterion is satisfied, stop. Other- est as well as other regions of no interest. This reconstruc-
wise go to 2. Cow rgence criteria are stated below tion is computationally expensive and therefore the CP
(Section 3). method computational burden is high.
This algorithm is typically stopped when the multiplier
3 Multiplier upda ing vector difference between two consecutive iterations is be-
The column vector of c is traint mismatches at iteration low a pre-specified threshold.
v constitutes a subgradj nt as stated in [14], i.e. 3.3 Bundle method (BD)
The updated multiplier vector is obtained solving the re
laxed dual quadratic programming problem below.

is a subgradient vector 1 hich is used below. Maximizez,eGc z - a(k)1,9 - Q ( k ) ) 1 2 (22)

3.1 Subgradient met1 ,d (SG)
subject to z 5 4 ( k ) + s ( k ) T ( ~- dk)) (23)
The multiplier vector is pdated as stated below [19]. IC= l,*..,U
where a is a penalty parameter and 8,the “center of grav-
ity”, is a vector of multipliers centered in the feasibility
region so that oscillations are avoided [15].
where It should be noted that the number of constraints of the
problem above grows with the number of iterations.
The BD method is a CP method in which the ascent proce-
dure is constrained by a n objective function penalty. The
target is to center the C P method in the region of interest.
However, in order to do this, it is necessary to carefully
The SG method is simp to implement and its computa- tune-up the penalty and other parameters [15, 161. This
tional burden is small. k wever, it progresses slowly to the tune-up is problem dependent and hard to achieve.
optimum in a n oscillatii ; fashion. This is a consequence
of the non-differentiabil This algorithm is typically stopped when the multiplier
7 of the dual function. Further-
more, the oscillating bc avior makes it very difficult to vector difference between two consecutive iterations is be-
devise an appropriate s ’pping criterion. It is typically low a prespecified threshold.
stopped after a pre-spec ied number of iterations. 3.4 Dynamically constrained cutting plane method
An elaborated subgradi It procedure which is more effi- (DC-CP)
cient than the original E bgradient method is reported in The updated multiplier vector is obtained solving the re-
92

laxed dual linear programming problem below.

Maximize, ,BE (.) z (24)
subject to t 5 I‘($ + . S ( ~ ) ~-( e(’))
@ (25)
IC=i , . . . , n ; n i n
where E is the maximum number of constraints considered
when solving the problem above, and C(”) is a dynamically
updated set defining the feasibility region for the multipli-
ers. Figure 1: Upper and lower bounds for the primal and dual
optima, and upper bound for the duality gap.
When the number of iterations is larger than the specified
maximum number of constraints, the excess constraints
are eliminated as stated below.
At iteration v the difference between every hyperplane of interest which results in high efficiency. The enlarg-
evaluated at the current multiplier vector arid the actual ing/shrinking procedure is not problem-dependent and in-
value of the objective function for the current multiplier volves straightforward heuristics (see the simple updating
vector (residual) is computed as procedure above).
This algorithm is typically stopped when the multiplier
vector difference between two consecutive iterations is be-
low a pre-specified threshold. A small enough difference
As soon as n is larger than E, the “most distant” hyper- between an upper bound arid a lower bound of the dual op-
planes are not considered, so that the number of hyper- timum is also an appropriate stopping criterion (see Sub-
planes is kept constant and equal to E. It should be noted section 3.5).
that the residual is always positive because the cutting 3.5 Duality gap
plane reconstruction of the dual function overestimates the
actual dual function. This technique to limit the number When using the CP method or the DC-CP method, at ev-
of hyperplanes considered has proved to be computation- ery iteration, the objective function value of the relaxed
ally effective. dual problem (problem (20)-(21) or problem (24)-(25))
constitutes an upper bound of the optimal dual cost value.
The dynamic updating of the set d”), the feasibility region This is so because the piecewise linear reconstruction of
of the multipliers, is performed as stated below. Let 0:”) the dual function overestimates the actual dual function.
be the z component of the multiplier vector a t iteration v. On the other hand, the objective function value of the
dual problem (evaluated through the decomposed primal
If 0:”) = e,’v’ then problem using equation (9)) provides at every iteration a
lower bound of the optimal dual cost value. This can be
- mathematically stated as follows:
ej”’(1 + a ) , and
--(”+l) -
0% -

where t(”)is the objective function value of the relaxed

dual problem at iteration u , 4* is the optimal dual cost
T
value, and $(”) = A ( ” I T H + +
p(’) G d ( A ( ” ) , p ( ” ) ) is the
objective function value of the dual problem at iteration
U.

The size of the per unit gap = - d(”))/~$(”)is an

appropriate per unit cost criterion to stop the search for
Overlining indicates upper bound and underlining stands
for lower bound. T h e parameters a , b, c arid d allow to
the dual optimum. Phase 2 converts heuristically the typi-
cally prirhal infeasible solution provided by the dual prob-
enlarge and shrink the feasibility region of the multiplier
lem to a primal reserve feasible solution. This typically
vector, i.e. the convex and compact set C. This is effi- originates a decrement in the objective function value, so
ciently accomplished because the above updating proce-
that
dure is simple.
The DC-CP method is a CP method in which the ascent
procedure is dynamically cotistrained by enlarging and
shrinking the feasibility region on a coordinate bases. This
is possible because the feasibility region is simple: bounds 4
where is the objective function value for the primal r e
on every multiplier. Through this enlarging/shrinking pro- serve feasible solution (Phase 2), and $(’) the objective
cedure it is possible to center the algorithm in the area function value of the dual problem a t the last iteration
 93

2a: SGMethod large-scale case study.

" 500 1000 15

1
1 2000 2500 3000 3500 4000
iter
In both case studies the generating system is based on the
electric energy system of mainland Spain. It consists of
60 thermal plants and 30 hydraulic plants in one complex
hydroelectric system.
2b: CP Method
1I Total cost is the sum of production cost, start-up cost
1 and shut-down cost for all thermal units in all periods.
-5 The production cost of every thermal plant is considered
10 20 10 40 50 60 ita-JO a quadratic function of the output power. The start-up
7
2c: BD Method and shut-down costs of every thermal plant are considered
I constant. Ramp rate limits, minimum up time and min-
imum down time constraints are enforced. Hydro power
d I
-J of each hydroelectric unit is modeled as a piecewise linear

2w
10 20 )O 40 50 60 70
iter-3 function of water discharge.
d: DC-CP Method
Although the number of constraints of the relaxed dual
1 problem for the C P and BD methods reported in the liter-
-0 ature grows with the number of iterations, for the sake of
10 20 50 40 50 60 70 comparison the C P and BD methods implemented main-
iter tain a pre-specified maximum number of constraints. The
procedure used to keep limited the number of constraints
Figure 2: Phas I for the testing example is explained in Subsection 3.4.
The reported CPU time refers to a SUN ULTRASPARC
2 workstation with 128 MB of RAM.
of Phase 1 which is il ,ation 7. Once 4
is known, the
Testing Example
unit commitment vari; les are set to fixed values, and a
demand adjusted prim feasible solution can be obtained
through an economic ( patch procedure (Phase 3). If
is the objective functio iralue of the demand adjusted pri-
mal feasible solution, a ?r unit upper bound of the duality
gap value is given by a : (f - 4('))/4('), because is an
upper bound for the p nal optimum and 4(') is a lower
bound for the dual opl num. In most practical cases the
duality gap is below 0.1 .
Figure 1 shows upper and lower
bounds for the primal i d dual optima, the actual duality Table 2: Costs and gaps for the large-scale case study
gap, and a n upper bou 1 for it.

Figure 2 shows the results of Phase 1. Figure 2a is a plot of

the cost of the dual problem when the SG method is used
to solve Phase 1. Figures 2b, 2c and 2d include a plot of
[MPta] the cost of the dual problem and a plot of the variable t
used in the relaxed dual problem for the CP, BD and DC-
I C P methods respectively. It should be noted that vertical
SG I 151.005 axis scales are different for the three plots of Figure 2, and

EFLG
I
DC-CP
I
150.852

150.882
0.085
0.079
I

I 49
I

I
1.378
1.378
1.000
that the horizontal axis scale of plot of Figure 2a is dif-
ferent from the horizontal axis scales of the other plots of
Figure 2. The above axis scale differences are introduced
to enhance clarity. For C P and BD methods the solution
Table 1: Cost, gap, nul her of iterations and CPU time in oscillates largely in the first few iterations. In order to
better appreciate the evolution of these methods, figures
per unit (normalizing fi tor: 6.322 seconds) for the testing
l b and I C include the cost of the dual problem and the
example value of z starting from iteration number 3. The solutions
of the dual problem for all the methods implemented are
reserve feasible primal solutions and therefore, Phase 2 is
4 Case Studies skipped. Table 1 shows, for the four methods of solution of
Two case studies are ar yzed. The first one is a six period Phase 1 described in Section 3, the cost of the dual prob-
testing example whose urpose is to compare the conver- lem (DPC) in millions of pesetas (Pta, 1 US$ 2: 145 Pta),
gence properties of thl Four methods of solution for the the size of the gap g(') (except for the SG method where
dual problem describe1 in Section 3. The planning hori- it is not defined) where ?7 is the last iteration of Phase 1
zon for the second cas is 48 periods. This is a realistic (Subsection 3.5), the number of iterations needed to solve
94

Phase 1 and the required CPilJ time normalized with re- 5 Conclusions
spect to the total CPU time in seconds required by the The objective of the short-term hydro-thermal coordina-
DC-CP method. Total CPW time for the DC-CP method tion problem is to determine the scheduling and produc-
is 6.322 seconds. tion of every hydro and thermal plant of an electric energy
system so that the customer demand is supplied at mini-
Figure 3a: CP Method mum cost and a certain level of security, and all constraints
x lo" related to the thermal and hydro subsystems are satisfied.
I
2 The most effective approach to solve this problem is La-
0 grangian Relaxation. This approach requires the solu-
-2 tion of the dual problem of the original short-term hydro-
500 1000 1500 thermal coordination problem.
iter-3
x 18 Figure 3 b BD Method This paper focuses on the updating procedure of La-
1 I
grangian multipliers which is the mechanism to solve the

I: 500 1000 1500

iter-3
1 dual problem. A novel, non-oscillating, and computation-
ally efficient procedure is presented. It outperforms previ-
ous approaches such as subgradient and bundle methods.
Extensive computational results based on the electric en-
Figure 3c: DC-CP Method ergy system of mainland Spain are presented.
,

Number CPU CPU Total

Model iterations time time CPU
500 1000 1500 Phase I Ip.u.1 Ip.u.1 time
iter

Figure 3: Phase 1 for the large-scale case study

BD 1503 1.431 1 0.002 I 1.436
DC-CP 1280 0.995 1 0.002 I 1.000
The SG method shows slow convergence. The number of
iterations needed to solve Phase 1 is much larger than the Table 3: Number of iterations to solve the dual problem
number of iterations required by any of the other three and required CPU time for Phase 1 and Phase 2, in per
methods. In terms of CPU time the SG method is the unit (normalizing factor: 11088 seconds), for the large-scale
most inefficient method to solve Phase 1 and therefore case study
it will not be used in the realistic large-scale case study
presented below. If good initial values for the multipliers
are available and careful tune-up of the subgradient step
size is performed, the SG method may show better com- Acknow liedgment s
putational behaviour than the one shown in the previous The work reported in this paper has been partly supported
testing example by the Manasterto d e Educacadn y Cultura of Spain, project
Large-Scale Case Study DGICYT PB95-0472, by Red EMctraca d e Espafia, REF,
and by the Vicerrectorado d e lnvesdigncadn of the Uniuer-
Figures 3a, 3b and 3c show respectively the results when sidad d e Mdlaga, The authors are grateful t o Prof. A. de
the CP, BD and DC-CP methods are used to solve Phase la Calle for support and help.
1. These figures provide similar information than the one
provided by figures 2b, 2c and 2d respectively. It should be
noted that vertical axis scales are different for the plots of
Figure 3 . Figures 3a and 3b include the evolution of cost References
starting from iteration number 3. Table 2 presents the
cost of the dual problem (DPC) and the cost of the primal
reserve feasible solution (PPC), both in millions of pesetas [I] J. A. Muckstadt, S. A, Koening. An Applzcation of
Lagrangtan Relaxation to Scheduling in Power Gen-
(Pta), and the the size of the gap g('). Table 3 shows the eratzon Systems. Operation Research. Vol. 25, No 3,
number of iterations t o solve P h a s e 1 and the required
pp. 387-403. May/June 1977.
normalized CPU time. Total CPU time for the DC-CP
method is 11088 seconds. Most of this time (around 90 121 A. Merlin, P. Sandrin. A New Method for Unit Com-
% for the CP and DC-CP methods and 78 '3% for the BD mztment at Electrzczte' de f i a n c e . IEEE Transactions
method) is used to solve complex hydro subproblems. o n Power Apparatus arid Systems. Vol. PAS-102, No.
5 , pp. 1218-1225.May 1983.
F'rom the above results, it can be concluded that the
DC-CP method is a very competive method of solution [3] D. P. Rertsekas, G. S. Lauer, N. R. Sandell Jr., T . A.
for the dual problem. Posbergh. Optimal Short-Term Scheduling of Large
 95

[IS] J. B. Hiriart-Urruty, C. LemarCtchal. Convex Analysis

and Minimization Algorithms. Vol. I and 11. Springer-
Verlag. Berlin, 1996.
[17] P. B. Luh, D. Zhang, R. N. Tomastik. A n Algo-
rzthm for Solvzng the Dual Problem of Hydrothermal
Scheduling. IEEE Power Winter Meeting, Paper P E
601-PWRS-0-01-1997. New York, 1997.
[18] X. Guan, R. Baldick, W. H. E. Liu. Integrating Power
System Schedulang and Optamal Power Flow. Pro-
ceedings of the Twelfth Power Systems Computation
[6] V. M. Mendes, . F. M. Ferreira, P. Roldgo, Conference. PSCC’96, Vol. 11, pp. 717-723. Dresden,
Germany, August 1996.
[I91 B. T. Polyak. Introduction to Optimization. Opti-
mization Software, Inc. New York, 1987.
[20] C. A. Kaskavelis, M. C. Caramanis. Ejjicient La-
grangzan Relazataon Algorzthms for Real-Lzfe-Saze
[7] M. V. Rakik, Z. : Short Term Opera- Job-Shop Scheduling Problems. Unpublished manu-
tion and Power Ex( ing of Hydro-Thermal script. Department of Manufacturing Engineering,
Power Systems. I1 tions on Power Sys- Boston University, Boston MA 02215, 1996.
tems. Vol. 9, No. 1
Biography
N. JimBnez Redondo (S96) received the Ingeniero In-
dustrial Electric0 degree from the Universidad P. Comi-
laxation. IEEE oxis on Power Systems. Vol. llas, Madrid, Spain in 1991. She was a research assistant
a t the Universidad P. Comillas from January 1992 to Oc-
[9] W. J. Hobbs, G . n, S. Warner, G . B. Sheblk. An tober 1993. She is currently a lecturer at the Universidad
de MBlaga, where she is working toward her Ph.D. in oper-
ations planning of hydro-thermal electric energy systems.
Her research interests include operations and planning of
electric energy systems, and optimization theory.
A. J . Conejo (S86, M91) received the B.S. degree from
the Universidad P. Comillas, Madrid, Spain, in 1983,
the M.S. degree from MIT, Cambridge, Massachusetts,
-97, NO. 6, pp. 2154-2166. in 1987, and the Ph.D. degree from the Royal Institute
of Technology, Stockholm, Sweden, in 1990, all in Elec-
trical Engineering. He was a visiting engineer at MIT,
[ll] H. Brannlund, D. en, J. Bubenko. Short Term
Generation Sche with Security Constraints. Cambridge, Massachusetts, and a visiting lecturer at the
IEEE Transactions er Systems. Vol. 1, No. 3, Royal Institute of Technology, Stockholm, Sweden. He is
pp. 310-316. Octok currently a professor a t the Universidad de Castilla - La
Mancha, Ciudad Real, Spain. His research interests in-
clude control, operations, planning and economics of elec-
tric energy systems, as well as optimization theory and its
applications.
Thermal Coordina
Twelfth Power Sj

[15] F. Pellegrino, A.
Augmented Lagran
Commitment. Proc
tems Computatior
pp.730-739. Dresdc

View publication stats