You are on page 1of 25

“main” — 2006/3/9 — 15:58 — page 317 — #1

Volume 24, N. 3, pp. 317–341, 2005
Copyright © 2005 SBMAC
ISSN 0101-8205

Solving the unit commitment problem
of hydropower plants via Lagrangian Relaxation
and Sequential Quadratic Programming
1 LabPlan – Laboratório de Planejamento de Sistemas de Energia Elétrica

Campus Universitario, Trindade, Florianópolis – 88040-900, SC, Brasil
2 IMPA – Instituto de Matematica Pura e Aplicada
Estrada Dona Castorina 110, Jardim Botânico, Rio de Janeiro – 22460-320, RJ, Brasil
E-mail: / /

Abstract. We consider the optimal scheduling of hydropower plants in a hydrothermal interconnected system. This problem, of outmost importance for large-scale power systems with a
high proportion of hydraulic generation, requires a detailed description of the so-called hydro unit
production function. In our model, we relate the amount of generated hydropower to nonlinear
tailrace levels; we also take into account hydraulic losses, turbine-generator efficiencies, as well
as multiple 0-1 states associated with forbidden operation zones. Forbidden zones are crucial
to avoid nasty phenomena such as mechanical vibrations in the turbine, cavitation, and low efficiency levels. The minimization of operating costs subject to such detailed constraints results in a
large-scale mixed-integer nonlinear programming problem. By means of Lagrangian Relaxation,
the original problem is split into a sequence of smaller and easy-to-solve subproblems, coordinated by a dual master program. In order to deal better with the combinatorial aspect introduced
by the forbidden zones, we derive three different decomposition strategies, applicable to various
configurations of hydro plants (with few or many units, which can be identical or different). We
use a Sequential Quadratic Programming algorithm to solve nonlinear subproblems. We assess
our approach on a real-life hydroelectric configuration extracted from the south sub region of the
Brazilian hydrothermal power system.
#613/04. Received: 30/VIII/04. Accepted: 22/X/04.
Correspondence to: Erlon C. Finardi

“main” — 2006/3/9 — 15:58 — page 318 — #2



Mathematical subject classification: 90C90, 46N10, 47N10.
Key words: hydrothermal systems, Unit Commitment Problems, Lagrangian Relaxation, sequential quadratic programming.

1 Introduction
The optimal generation scheduling is an important daily activity for electric
power generation companies. The goal is to determine which units are to be
used in order to generate enough power to satisfy demand requirements and
various technological constraints, with minimum operating cost. In particular,
hydrothermal systems must consider the stream-flow equations for reservoirs.
These equations couple all the reservoir along a hydro-valley, because the amount
of outflow water1 released by one power plant affects water volumes in all the
plants downstream. Furthermore, water travel times and alternative uses of water,
such as irrigation or flood control, for example, must also be taken into account.
The optimal scheduling of hydropower plants is called the Hydro Unit Commitment (HUC) problem. To solve the HUC problem, a highly sophisticated modeling for the operation of hydro plants is required. Specifically, a hydropower
plant may be composed of several turbine-generator groups, referred to in this
work as “units”. The amount of power generated by one hydro unit depends on
the efficiency of both the turbine and the generator, as well as on the net head
and the unit turbined outflow. In turn, the net head is a nonlinear function of the
storage and of the reservoir outflow. The joint turbine-generator efficiency varies
with the net water head and the unit turbined outflow. In addition, the existence
of forbidden operation regions prevents the unit from generating power in a wide
and continuous range. These regions, modeled by 0-1 variables, aim at avoiding vibrating modes that may produce unwanted power oscillations, cavitation
phenomena, and low levels of efficiency. Thermal power plants have simpler
production functions, but they need start-up and shut-down times and they often
present nonlinear operating costs.
As a result, the hydrothermal unit commitment problem is a large-scale mixedinteger nonlinear programming problem which can only be effectively solved by
1 Turbined and spilt water flow.

Comp. Appl. Math., Vol. 24, N. 3, 2005

requiring less computation effort.. all derived from LR. (1) The net water head has the expression: hl j = f cm − f cj (Q.10]. and can be used in systems with both types of hydro plants. [1-5]. and forbidden operation zones. Comp. N. our contribution consists in a thorough analysis of three different decomposition schemes. turbine-generator efficiencies.8]. However. ph j . DA SILVA and CLAUDIA SAGASTIZÁBAL 319 applying decomposition techniques. nonlinear tailrace levels. is given in Section 4. 24. The third strategy combines the two other approaches. 810 10−3 h j hl j q j . FINARDI. The second strategy. η j : ph j = 9. which presents good convergence properties. Math. EDSON L. is applicable for plants with many units that are different and have many forbidden regions. s) − k j q 2j . We end in Section 6 with some concluding remarks. 2 Hydro generating units For a unit j. Lagrangian Relaxation (LR) is particularly suitable for this type of problems. The first strategy relies on a complete enumeration of all possible 0-1 operating states of the units composing a hydropower plant. and on joint turbine and generator efficiency. We solve them by a Sequential Quadratic Programming (SQP) method [7.“main” — 2006/3/9 — 15:58 — page 319 — #3 ERLON C. depends on the unit turbined outflow. hl j . 3. so far none of the works in the area has considered a modeling as comprehensive as ours. q j . Our work is organized as follows. Section 3 is devoted to the mathematical formulation of the optimal scheduling problem. with the three decomposition schemes. with representation of hydraulic losses. Vol. This approach is suitable for plants with a low number of identical units. and on the net water head. Many of the subproblems resulting from our decomposition schemes are Nonlinear Programs (NLP) of small size. the generated power. In Section 2 we give the hydro plants and units modeling. The solution strategy. In Section 5 we report numerical results on a hydrothermal system corresponding to Brazil’s southern electric sub region. Appl. in a quasi-Newton variant [9. With respect to the solution method. although some other methodologies have also been proposed [6]. expressed in [MW]. 2005 (2) .

Therefore. in order to extend the lifetime of the unit and to avoid power oscillations. is usually represented by hill diagrams given by the factory. 2005 . . For example. for net head values smaller than the so-called nominal level (41. . specially in the Brazilian case. s 12 .5m). using a polynomial function: η j = ρ0 j + ρ1 j q j + ρ2 j hl j + ρ3 j hl j q j + ρ4 j q 2j + ρ5 j hl 2j . there is a limit of power limit imposed by the generator capabilities. . fcj also varies with the reservoir spillage.. fcm stands for the forebay level. the downstream level fcj(. Therefore. s. s 4 . f cj (Q. see Figure 1. depending on hl j and q j . (3) where the coefficients ρ0 j . The unit efficiency. . because the turbine could effectively reach power levels beyond 120 MW.5m. For some power plant configurations. N. 2 Since hl = f j   f q 7j . Q 12 . Math. whose huge reservoirs have typical regularization levels of a couple of years. and of order 7 in the variable q j . mainly due to the plant turbined outflow. where k j is a constant expressed in s2 /m5 [11].“main” — 2006/3/9 — 15:58 — page 320 — #4 320 SOLVING THE UNIT COMMITMENT PROBLEM Here. (4) For the Brazilian case. ph j = Comp. Figure 1 also displays some important operating constraints on the turbinegenerator group. Since at some power levels cavitation phenomena and nasty mechanic vibrations may appear. the term k j q 2j represents hydraulic losses resulting from friction of the water in penstock. from (4) we see that ph j is a polynomial2 of degree 12 on the variables Q and s. by (1). 24. For short term horizon problems. For example. In (2). s 8 and. Appl. On the other hand. it follows that η j = f q 4j . Q 8 . Figure 1 shows a forbidden operation region ranging from 70 to 90 MW. we consider fcm constant. for values higher than 41. such power levels are forbidden. thus. s) is represented by a fourth degree polynomial. We estimate it by interpolation.) varies abruptly in short times. 3. as in our case. By combining (1)–(3) we obtain our model for the hydro production function:  ph j = 9. given by the addition of the outflows of all the units composing the plant. ρ5 j have been computed beforehand. the turbine is unable to make the generator attain its nominal power (120 MW). Vol. 810 10−3 ρ0 j +ρ1 j q j +ρ2 j hl j +ρ3 j hl j q j +ρ4 j q 2j +ρ5 j hl 2j hl j q j .     q 2j . Q. By contrast. Q 4 . the forebay level remains practically constant. see [12].

“main” — 2006/3/9 — 15:58 — page 321 — #5 ERLON C. cit (. the thermal mix has I plants.90 Forbidden Generation Zone 0.80 0.. 3. Vol.86 0. 24.78 0. 340 320 300 0. At first sight our model may appear as “too complicated”. Frequently cit (. Appl. Comp.) represents the operating cost of the i-th thermal plant at time step t. see (8) below. N. FINARDI.88 0. 3 Problem formulation The objective function for the thermal-HUC problem has the expression: T X I X t=1 i=1 cit ( ptit ) + α .82 280 260 240 0. EDSON L. it is important to realize that only such a detailed description can accurately represent the diverse amounts of power generated by a unit at different operating states.[5]. however.94 220 200 180 0. (5) Here. Math.92 . DA SILVA and CLAUDIA SAGASTIZÁBAL 321 . and α stands for the system expected future cost at the end of the planning horizon. 2005 .80 Figure 1 – Hill diagram.84 0.) includes fixed costs as well as fuel costs related to start-up and nominal generation of thermal units [2]. the planning horizon is composed by T time steps. 0.

and τmr is the water travel time between reservoirs m and r .t−τmr + sm. thermal.“main” — 2006/3/9 — 15:58 — page 322 — #6 322 SOLVING THE UNIT COMMITMENT PROBLEM We formulate the thermal-HUC constraints by splitting them into three different subsets. 0 ≤ sr t ≤ srmax . 3. Vol.. 2005 (9) . <(r + is a set gathering all reservoirs upstream the r-th. (6) We use the index r for reservoirs. Each subset is characterized by a specific type of coupling. (8) • Penstock water balance equation per reservoir: Qr t = J (r ) X q jr t . Comp. see [13]. hydraulic. It is a piecewise affine function that depends on the final levels of stocked water. corresponding to the respective variables involved namely. v is the reservoir storage. 24. vr T : α = f (vr T ) . N. j=1 J (r ) is the number of generating units in reservoir r . and estimating the cost of using today water that might become necessary (and expensive) in the future. and maximum spillage srmax per reservoir: vrmin ≤ vr t ≤ vrmax . Appl.t−τmr − νr t = yr t . or different power plants in a given time step (space coupling). 3. such as units in the same power plant along different time steps (time coupling). y is the ) incremental inflow. • Maximum vrmax and minimum vrmin storage. Math. given by longer term planning models. We now proceed to give each constraint in detail.t+1 + Q r t + sr t − (r ) m∈<+  Q m. (7) • Expected future cost function. or both. CT and CHT .1 Constraints involving only hydraulic variables (CH ) • Stream-flow balance equation: X vr. CH .

Appl. 24. to alleviate notation. • Reservoir power balance: P Hr t = J (r ) X ph jr t (q jr t .“main” — 2006/3/9 — 15:58 — page 323 — #7 ERLON C. while CHUC represents the unit constraints. In this abstract formulation. q. given for each operating region of the unit: 8 jr X k=1 ph min jkr t z jkr t ≤ ph jr t (q jr t . The binary variable z jkr t is 1 if the j-th unit in reservoir r is operating in the k-th region at time step t. N. P H and V gather the respective variables. V ) ∩ CHUC (z. 8 jr X k=1 z jkr t ≤ 1 . Math. Q. modeling the reservoirs. s.. Q r t . (11) j=1 • Reserve constraints:   8 jr J (r ) X X   ≥ r hr t . The set CHH represents constraints given by (6)-(8). (9)-(13). sr t ) ≤ 8 jr X ph max jkr t z jkr t . q. (10) k=1 8 jr denotes the total number of non-forbidden regions of the j-th unit stand for the in reservoir r . and ph min. and it is set to 0 otherwise. DA SILVA and CLAUDIA SAGASTIZÁBAL 323 • Power limits. i.max jkr t minimum and maximum power limits. ph max j1r t z jkr t − P Hr t (12) j=1 k=1 r h r t is the minimum reserve of reservoir r at time step t. 1}. we write constraints CH above in the abstract form CH = CHH (Q. α = α(V ). where the vectors z. • Integrality constraints: z jkr t ∈ {0. EDSON L.. FINARDI. Q. s. P H ). s. Comp. 3. (13) In the sequel. Vol. k is the corresponding index. 2005 . sr t ) . Q r t .e.

) and 1i (. where the state variable xit is equal to the number of time steps the unit has been up/down until time t. Appl. pt).. for each unit:  up  if 1 ≤ xit < ti  1 u it = 0 if − 1 ≥ xit > −tidown   0 or 1 otherwise . 0) + 1. xit ) ≤ ptit − pti. • Minimum up-time.t−1 ≤ 1i (u i. Here ptimin.t−1 . xit = ( max(xi. if u it = 1. Math.) are the maximum allowed variations of generation of the unit between two time steps. • Reserve constraints: ptimax u it − ptit ≥ r tit . Vol. 3. respectively. r tr t is the reserve of unit i at time step t. 2005 . if u it = 0. • Ramp constraints: δi (u i. δi (. Comp.max stand for the minimum and maximum power limits of unit i.2 SOLVING THE UNIT COMMITMENT PROBLEM Constraints involving only thermal variables (CT ) • Power limit for each unit: ptimin u it ≤ ptit ≤ ptimax u it . min(xi. N. xit ) . and downtime. constraints in the set CT correspond to CT (u.t−1 .t−1 . 0) − 1. and it is set to 0 otherwise. where u and pt are vectors gathering all binary and continuous thermal variables.t−1 . 24. The binary variable u it is 1 if the unit is operating at time step t. ti down .“main” — 2006/3/9 — 15:58 — page 324 — #8 324 3. In an abstract formulation. ti up .

max max . I ntelt max max 0 ≤ I ntlet ≤ I ntlet . N. is well known. from e(l) to l(e). where the vector I nt gathers the subsystem exchanges. Comp. Because of their solid theoretical background. P H. There are e subsystems interconnected with subsystem e. (I ntlet ): • Subsystems exchange limits. 0 ≤ I ntelt ≤ I ntelt . the set C H T is written as CHT ( pt. the exchange of energy at time t. i∈Ie r ∈Re (14) l∈e The interconnected hydrothermal system is divided into subsystems. FINARDI.“main” — 2006/3/9 — 15:58 — page 325 — #9 ERLON C. multipliers associated to demand constraints given by (14) are used to price energy. In our abstract notation. LR techniques appear in this area as the preferred solution method. is denoted by I ntlet (I ntelt respectively) when it goes from subsystem l to e (from e to l. respectively). The above description confirms the level of complexity of the optimization problem to be solved.3 325 Constraints involving both hydraulic and thermal variables (CHT ) • Satisfaction of demand. at time t. In particular. Det is the demand of subsystem and at time t. which finds new multipliers by making one iteration of a nonsmoth algorithm that maximizes the dual function. coupling constraints are relaxed via Lagrange multipliers whose corresponding dual problem is decomposable into simpler subproblems (called local subproblems). Accordingly. 3. indexed by e.. per time step and subsystem: X X X ptit + P Hr t + (I ntlet − I ntelt ) = Det . DA SILVA and CLAUDIA SAGASTIZÁBAL 3. all thermal units (reservoirs) of subsystem e are gathered in the index set Ie (Re ). Math. The divide to conquer approach of LR. 4 Solving the HUC problem The economic impact of the optimal scheduling of power plants is undeniable. also called price decomposition [9]. EDSON L. 2005 . I nt). 24. Appl. Vol. The coordination of subproblems is then done by a master program. Essentially. Finally. We now address the solution strategy adopted in this work.

q. Qa and sa replace Q and s in CHH .15] for an application to the thermal UC problem. s. V ) pt = pta. Variables pta and P H a are used in constraints CHT to replace pt and P H . see [14]. λ S and writing the corresponding dual problem3 : 3 From now on. the introduction of artificial variables to uncouple constraints appears as a good choice. u) ∩ CHT ( pt. N. ∩ CHUC (z. Q = Appl. q.t. With these additional variables. and also [5. we relax them by associating Lagrange multipliers λ P T . we apply a similar approach in this work. pta. pt and P H . I nt) ∩ CHH (Qa..Q. For this reason. sa. adapted to different unit configurations in the Brazilian hydrothermal system. λ P H . To achieve decomposition. Comp.s.Qa. In this matter. (15) is rewritten as follows: minimize u.P H. which duplicate. respectively. we introduce artificial variables pta and P H a. Hence.t. Math. respectively. s. the Euclidean inner product of two vectors. Vol.q. artificial variables Qa and sa duplicate Q and s. 24. P H ) P H = P H a. pt.I nt s. 4. An important criterion for deciding how to proceed is the resulting duality gap. In (16) the newly introduced artificial constraints hold the coupling of the problem. 3.1 First decomposition strategy – D1 In the abstract notation.: CT ( pt.q.I nt c( pt) + a(V ) s. V ) (15) ∩ CHUC (z. In addition.P H.P H a. Q. λ and v. P H a. and derive three different decomposition schemes. s. pt. (16) s = sa . λ Q . Q.s. the thermal HUC problem becomes: minimize u. P H ) .Q. which should be the smallest possible.: c( pt) + a(V ) CT ( pt.“main” — 2006/3/9 — 15:58 — page 326 — #10 326 SOLVING THE UNIT COMMITMENT PROBLEM There are many ways of relaxing coupling constraints.z.z. I nt) ∩ CHH (Q. will be denoted by λT v = 6i λi vi . u) ∩ CHT ( pta. 2005 . P H.

t. can be split as the sum of four terms. coupled along time steps. λTP H P H + λTQ Q + λTs s (22) CHUC (z. min z. sa. 2005 . but not along times steps.q. s. (18) D1T (λ) = min c( pt) + λTP T pt u. sa.: D1HUC (λ) = (20) CHT ( pta. Math. u) . DA SILVA and CLAUDIA SAGASTIZÁBAL maximize λ P T . P H a.[5]. EDSON L.P H.I nt  c( pt) + α(V ) + λTP T ( pt − pta) + λTP H (P H − P H a) + λTQ (Q − Qa) + λTS (s − sa) s. Problem (17) can be rewritten as follows: maximize λ=[λ P T . coupled both in time and space via the stream-flow constraints given by (6).: D1HH (λ) = min α(V ) − λTQ Qa − λTs sa Qa. P H ). pt CT ( pt.P H s. as in [1]. which can be solved by any (LP) commercial solver. In the LR approach. Vol. Even though (21) can be large-scale. but not along plants. V ) ..λs minimize u. pt. Subproblem (20) is a standard linear programming (LP) problem. Finally. FINARDI. I nt).λ P H . Appl. an LP solver can still solve it efficiently. Subproblem (19) is a nonlinear optimization problem with continuous and binary variables. I nt) ∩ CHH (Qa. V ) ∩ CHUC (z.  T  − λ P T pta + λTP H P H a (19) s.s.t.t.Q.: (21) CHH (Qa.z. s.I nt s. q. Q.λ Q . the primal problem (15) is replaced by the dual problem (18) whose objective function D1(λ).: 327  (17) CT ( pt.P H a.Q.P H a. Subproblem (21) is also an LP problem. pta. subproblem (22) is a nonlinear mixed-integer optimization problem.λ P H . corresponding to subproblems (19)-(22).q.t.“main” — 2006/3/9 — 15:58 — page 327 — #11 ERLON C.λ Q .λ S ] where: D1(λ) := D1T (λ) + D1HT (λ) + D1HH (λ) + D1HUC (λ) .: D1HT (λ) = min pta. coupled along plants. 24.s. Q. uncoupled both in time and Comp. N. P H ) . It can be solved by a classic Dynamic Programming method. q. u) ∩ CHT ( pta. P H a.t. 3.

s] which appear in CHUC . Q. is a mixed-integer NLP problem. a complete enumeration of modes seems a good strategy.. In this case. The total number of possible operating modes is given by the product of all combinations of the operating modes of all the units composing the plant.Q. but J (r ) + 1 and. P H ) ∩ CHUCb (z.t. for each time step and for a given power plant.: c( pt) + α(V ) CT ( pt. Each combination of a unit is a configuration where the corresponding binary variables are fixed to one of the feasible values. s) ∩ CHUCres (z. q. CHUCb contains (10) and (13) and CHUCres corresponds to the reserve constraint (12). we eliminate the coupling between the binary variable z and the continuous variables [q. and several operating modes. Appl. Generally. For these configurations. there are power plants with many different types of units.P H. with binary variables corresponding to different operating modes in the plant. Sometimes. we rewrite (15) as follows: minimize u. whose size is dependent on J (r ). 24. thus. the problem becomes a nonlinear program. V ) (23) ∩ CHUCa (q. Each sub-subproblem in (22). We now introduce an alternative decomposition scheme. Vol. respectively) are. Besides the artificial variables used in (16).I nt s. we use pha to replace the hydro production Comp. adapted to such situations. P H.q. Once the binary values are fixed. 3. Q. Therefore. and each unit has a single operating zone. Now the set CHUCa gathers constraints given by (9) and (11). 2005 .s.z. N.2 Second decomposition strategy – D2 In order to avoid the enumerative process required to solve subproblem (22). for a given reservoir and time step. however. This subproblem corresponds to the commitment of hydro units. Q. s. The higher J (r ) and 8 jr (the number of units in the reservoir and of operating zones. pt. hydropower plants have identical units. the total number of modes is no longer 2 J (r ) .“main” — 2006/3/9 — 15:58 — page 328 — #12 328 SOLVING THE UNIT COMMITMENT PROBLEM units. u) ∩ CHT ( pt. s. an enumeration procedure may become too expensive from the computational point of view. the bigger computational effort will be required to solve (22). I nt) ∩ CHH (Q. P H ) . 4. Math.

λ Q . V ) ∩ CHUCa (q.s.P H a. P H ) . The remaining terms are: D2HUCa (λ) = min λTRes P H + λTph ph(q.λ Res ] D2(λ) := D1T (λ) + D1HT (λ) + D1HH (λ) + D2HUCa (λ) + D2HUCb (λ) + λTRes r h where D1T (λ). s.t. λ S . D1HH (λ) are the dual functions from (19). Math. P H a.: + λTQ (Q − Qa) + λTS (s − sa) + λTph [ ph(q. N. Q. Q = Qa. Q.P H. Q.   D2HUCb (λ) = min − λTph pha + λTRes ph max z (25) s. Here. (20) and (21).λ P H . pta.z. V ) ∩ CHUCa (q.q. pt. EDSON L. P H ) ∩ CHUCb (z. Hence. Separability in (24) has now the expression: maximize λ=[λ P T .P H CHUCa (q. I nt) ∩ CHH (Qa. ph max and r h are vectors corresponding to a unit’s maximum capacity and to reserve levels for the plant. P H = P H a.: z.: CHUCb (z. pha) . P H a. λ ph and λ Res and writing the dual problem: maximize λ P T . s. λ P H . and rewrite (23) as: minimize u.Qa.I nt c( pt) + α(V ) + λTP T ( pt − pta) + λTP H (P H − P H a) s. s) in the set CHUCb . respectively.t. P H ) ∩ CHUCb (z. λ Q . Appl.“main” — 2006/3/9 — 15:58 — page 329 — #13 ERLON C.Q. pta.Q.Q.I nt s.s. sa.λ ph . pha s. FINARDI. s = sa. Now not only the artificial constraints.P H a. but also the constraint set CHUCres keep the problem coupled.z.λ P H . pt. u) ∩ CHT ( pta. pha) ∩ CHUCres (z.t. Comp. 24. sa. Q. Q.λs  minimize u. pha). D1HT (λ).sa. u) ∩ CHT ( pta.s. I nt) ∩ CHH (Qa.t. DA SILVA and CLAUDIA SAGASTIZÁBAL 329 function ph( s) = pha .λ Q .P H. s) + λTQ Q + λTs s q. Q.λ S .: c( pt) + a(V ) CT ( pt. Q. ph(q. we relax these constraints by introducing multipliers λ P T . s) − pha]  + λTRes [P H − ph max z + r h] (24) CT ( pt. P H ) pt = pta.Qa. 3. 2005 (26) . Vol. s..

Q. q.. D1HH (λ) are those in (19). s.I nt s. P H ) ∩ CrHUCa (q. sa. (20) and (21). which can be solved by enumeration of the operating zones. 4. to plants where (22) and (25)-(26) are applied: minimize u.λ Q . After relaxation. The dual function D1HUC (λ) from (22) only applies for reservoirs Comp. P H a. Q. Accordingly. s.t. V ) ∈R(0) ∈R(4) (z. for each time step and power plant. s.t.Qa.“main” — 2006/3/9 — 15:58 — page 330 — #14 330 SOLVING THE UNIT COMMITMENT PROBLEM Subproblem (25) is an NLP problem (with only continuous variables). 2005 .3 Third decomposition strategy – D3 Our last decomposition combines D1 and D2. pt.I nt s. Q. Q. Q. the dual problem of (27) is: maximize λ=[λ P T . but splitting the reservoir index sets: minimize u. s.: c( pt) + a(V ) CT ( pt. while D2 is applied for plants for which D1 would be too expensive. ph(q.z. r ∈ 0(4) . P H ) ∩ CrHUCa (q. I nt) ∩ CHH (Q. P H ) . Q. Vol. ∈R(4) ∈R(4) ∩ CrHUCb (z. I nt) ∩ CHH ( respectively. Appl.s. P H.λ ph .q.P H.: c( pt) + a(V ) CT ( pt. (27) s = sa.P H a.P H.λ S .λ P H . P H ) ∩ CrHUC ∈R(4) ∈R(4) ∩ CrHUCb (z. P H ) pt = pta. q. N. u) ∩ CHT ( pta. s) = pha. 3. q. we employ D1 for those plants with a reduced number of units and operating zones.λ Res ] D3(λ) := D1T (λ) + D1HT (λ) + D1HH (λ) + D1HUC (λ) + D2HUCa (λ) (28) + D2HUCb (λ) The dual functions D1T (λ). Subproblem (26) is a mixed-integer linear program on variables corresponding to a single unit. V ) ∈R(0) ∈R(4) ∩ CrHUC (z. We split the hydro plants index set into R(0) and R(4). Q. corresponding.Q. Math. P H ) P H = P H A. 24. D1HT (λ). respectively.z.s. Q = Qa. s.q.Q. s) ∩ CrHUCres (z. q. pt. We proceed like for D1 and D2. s) ∩ CrHUCres (z. u) ∩ CHT ( pt. pta.

20(s k )T M k s k . N. (s k )T M k s k (s k )T r k where: z k = ∇x L k+1 − ∇x L k .: ∇ce (x k )T p k + ce (x k ) = 0. In order to avoid the calculation of secondorder derivatives. The remaining dual functions in (28) apply to reservoirs r ∈ R(4). appended with a Powell correction [17]: M k+1 = M k − M k s k (s k )T M k r k (z k )T + . the objective and equality and inequality constraint functions at a point x k . 3.4 Nonlinear programming subproblems A crucial issue for an effective application of LR is the fast resolution of subproblems defining the dual function for a fixed multiplier λ. FINARDI. DA SILVA and CLAUDIA SAGASTIZÁBAL 331 with index r ∈ R(0). 4. and to preserve positive definiteness of the sequence of quasiNewton matrices. r k = θ k z k + (1 − θ k )M k s k . 2005 . EDSON L. L k . f (x k ). respectively. 8(s k )T M k s k   if (s k )T z k < 0.5( p)T M k p p s. 24.t. and are given by subproblems (25)-(26). ce (x k ) and ci (x k ) represent. Appl.. 20(s k )T M k s k   θk = 0. s k = x k+1 − x k . σ ) = φ(x k ) + σ k c(x k )# ∞ . we use a BFGS (Broyden-Fletcher-Goldfarb-Shanno) [16].“main” — 2006/3/9 — 15:58 — page 331 — #15 ERLON C. More precisely. each iteration k of the algorithm generates a direction pk by solving the quadratic programming problem: minimize ∇ f (x k )T p + 0. Math. Here. Comp. (s k )T M k s k − (s k )T z k Globalization of the method is achieved by performing a line search on the following function4 : φ(x k . Vol. formula. The matrix M k estimates the Lagrangian Hessian for the NLP. (29) ∇ci (x k )T p k + ci (x k ) ≤ 0. we implemented a Sequential Quadratic Programming (SQP) quasi-Newton method.  1 if (s k )T z k ≥ 0. (30) 4 The symbol # is used to denote only active constraints at x k . Since many of such subproblems are nonlinear programs.

We update the parameter σ k accordingly. The corrected direction pk .2 MW. More precisely. Vol. 3. 1/2[ and α is the positive stepsize. Appl. We estimate it by the upper bound from (31): 1k := ∇ f (x k )T p k − σ k c(x k )# ∞ . ω ∈]0. N.σ k ) solves the original NLP for all σ k ≥ σ k . 24.“main” — 2006/3/9 — 15:58 — page 332 — #16 332 SOLVING THE UNIT COMMITMENT PROBLEM known as Han’s merit function [9]. such as p k . as shown in [9]. Since φ(x k . p k ] ≤ ∇ f (x k )T p k − σ k c(x k )# ∞ (31) ≤ ( p k )T M k p k + (ζ k )T c(x k ) − σ k c(x k )# ∞ . 5 Numerical Results We assess the three decomposition schemes on a real-life hydroelectric configuration extracted from the Southern region of the Brazilian hydrothermal power system. σ k ). 1k should be the exact value of the directional derivative. (32) here. More precisely. For any stationary point of (30). it can be shown that the estimate in (31) gives a descent direction for φ if M k is positive definite and σ k is updated in order to satisfy: σ k ≥ ζ k ∞ + δ..1 below for more details. see 5. Comp. Math.129. Ideally. We exit the line search when the Armijo [18] condition is satisfied: φ k (x k + αp k ) ≤ φ k (x k ) + ωα1k . there is a finite positive value σ k such that an unconstrained minimum of φ(x k . ensures that asymptotically there is enough constraint reduction. Finally. This phenomenon may impair the superlinear local convergence rate by rejecting unit stepsizes when close to a solution.σ k ) is an exact penalty function. we add an extra term to p k in order to avoid Maratos effect [19]. 2005 .5 Figure 2 reports the data for the 5 This amount corresponds to about 49% of the total hydraulic capacity of Brazil. we consider a system with 121 generating units whose maximum installed capacity is 31. δ > 0. the directional derivative of the merit function along the direction satisfies the relation: D[φ(x k . where ζ is the Lagrange multiplier associated with constraints in (29).

#4. DA SILVA and CLAUDIA SAGASTIZÁBAL 333 Figure 2 – Hydroelectric configuration. #6 and #14 have production functions independent of spillage. λ P Hr t .“main” — 2006/3/9 — 15:58 — page 333 — #17 ERLON C. we refer to [5] for this subject. Instead. The values for λ P Hr t are chosen based on generation costs associated with typical demand curves. 3.. For this configuration. where power plants numbered #3. #14. has only 2 units. We do not address here the dual solution in detail. The planning horizon of two days is discretized in hourly time steps. 9 in [9]. FINARDI. N. EDSON L. with higher values for peak times with high demand6 . while the smallest plant. yielding T = 48. 2 correspond to water travel times.e. Comp. 6 These costs presented values between 0 and 45 $/MW.. 2005 . while the inflows were considered null. the biggest power plants are #7 and #16. 24. system. with 20 units each one. i. we fixed Lagrange multipliers for each reservoir and time step. see Ch. Math. Values between brackets in Fig. Initial reservoir volumes were taken at 50% of the usable volumes. Appl. Vol. expressed in hours. and use a proximal quasi-Newton variant of a bundle method to optimize the remaining multipliers.

2005 . Figures 3 and 4 show the values for Q and Qa obtained at the last iteration of both D1UCH and D2UCH for Sobradinho #18 power plant. subproblem (22) has 1536 variables. Finally. The size of subproblem D3UCH depends on the decomposition scheme.168. Comp. to λ P Hr t . in addition.1 solver. In (25)-(26) subproblem D2UCH has (2 × R + Rv + ngmi x) × T = 8208 variables. one continuous). However. with only two variables (one integer. Math. after 175 iterations.. corresponding respectively to (22). Since T = 48. (25)-(26) and (28). The values for Q exhibit a different behaviour. For D1UCH we found an optimal value of $ –17. where ngmi x = 121 denotes the total number of units in the mix.230. see [20]. Since the dual value in D2UCH is smaller than the one from D1UCH . as well as other LPs are solved using ILOG CPLEX 7.e. Appl.“main” — 2006/3/9 — 15:58 — page 334 — #18 334 SOLVING THE UNIT COMMITMENT PROBLEM We implemented the three dual subproblems D1UCH . and two constraints. for peaks of demand. 6288 NLP problems. 3. closer to the price profile. R = 18 and Rv = 14.. The LP given by D1HH (21).0. for D2UCH there are R × T = 864 NLP and T × ngmi x = 5808 easy mixed-integer LP problems7 . Subproblem D1UCH has (R + Rv) × T variables. D2UCH . It can be seen in the figures that Qa takes mostly two values: 0 and 4278 m3 /s (maximum value). i. i.339.. N. For D1UCH there are. and dual values give lower bounds for the primal optimal value. we can conclude that primal variables associated with D1UCH are better than those associated with D2UCH . For D2UCH the optimal value found was $ –17. after 325 iterations in 50 minutes. where Rv denotes the number of reservoirs with production function depending on spillage. Vol. 8 The difference between Q and Qa. we now consider in more detail some selected primal variables. From the comparison of D1UCH and D2UCH we see that the relaxed primal constraint8 is more violated for D2UCH . 24. For the dual solution we use N1CV2 code. that took 180 minutes of CPU times in a Pentium III 550 MHz computer with 128 Mb of RAM memory.450. We also show the value of λ P Hr t (price) used to solve the subproblems. For these time steps. This bang-bang behaviour is explained by the linear nature of subproblem (21). To further assess the previous remark. infeasibility becomes smaller for time steps with bigger λ P Hr t . for some specific reservoirs.0.e. a problem that contains a higher number of relaxed constraints. D3UCH . primal points obtained with D2UCH are good approximations to those 7 These are indeed easy to solve problems.

For lower prices. requiring a computational effort that can be up to 3 times bigger. from D1UCH . Math. Appl. FINARDI. EDSON L. 24. both problems give primal points that are even more infeasible. the worse values being associated with D2UCH . In general. The results observed for Sobradinho power plant are typical for all the plants in the mix. 3. Vol. 2005 . with variations in the computed primal infeasibility. DA SILVA and CLAUDIA SAGASTIZÁBAL Price Q Qa 40 4000 35 30 3000 25 Price ($/MW) Turbined Outflow (m3/s) 335 20 2000 15 10 1000 5 0 1 7 13 19 Stages (h) 25 31 37 43 0 Figure 3 – Sobradinho Power Plant #18 D1UCH . we observed that for reservoirs downstream. that tend to operate with outflows near Comp.. N.“main” — 2006/3/9 — 15:58 — page 335 — #19 ERLON C. Price 4000 Q Qa 40 35 Turbined Outflow (m3/s) 30 3000 Price ($/MW) 25 20 2000 15 10 1000 0 5 1 7 13 19 Stages (h) 25 31 37 43 0 Figure 4 – Sobradinho Power Plant #18 D2UCH .

Price 10000 Q Qa 40 35 30 6000 25 4000 15 Price ($/MW) Turbined Outflow (m3/s) 8000 20 10 2000 0 5 1 7 13 19 Stages (h) 25 31 37 43 0 Figure 5 – Ilha Solteira Power Plant #7 D1UCH . 3. Price Q Qa 40 35 8000 30 25 6000 Price ($/MW) Turbined Outflow (m3/s) 10000 20 4000 15 10 2000 0 5 1 7 13 19 Stages (h) 25 31 37 43 0 Figure 6 – Ilha Solteira Power Plant #7 D2UCH . Vol.. We conclude from our experiments that. Appl.5 higher than D2UCH . Math. However. since average hourly CPU times for D1UCH are 3.“main” — 2006/3/9 — 15:58 — page 336 — #20 336 SOLVING THE UNIT COMMITMENT PROBLEM to the nominal values. subproblem D1UCH is a better option. 24. 2005 . the enumerative process required by Comp. We report this behaviour in Figures 5 and 6. primal points obtained from D2UCH were close to those from D1UCH . with the results for Ilha Solteira – 7 power plant. in terms of primal solutions. N.

In order to recover primal feasibility. Even though D3UCH 10 needs to solve 27. i.11%. the previously obtained values for D1UCH and D2UCH .“main” — 2006/3/9 — 15:58 — page 337 — #21 ERLON C. DA SILVA and CLAUDIA SAGASTIZÁBAL 337 D1UCH should not be used for power plants with complex configurations. EDSON L. that satisfies the constraints.077 220 70 2592 4560 Table 1 – Numerical results. in terms of CPU times the gain was of 61. In Table 1 we report the main results for the third decomposition scheme. general-purpose combinatorial optimization heuristics are not suitable.359. while the difference between D3UCH and D1UCH is $ 19. 24. where for R 4 = {7. 2005 .e. We also give.5 smaller. we use (25)-(26) instead.22]. 16} we applied the scheme corresponding to (25)-(26).. The computational effort required for solving subproblems involving the big plants #7 and #16 (each one having 20 identical units . about 5.230 175 180 1536 6288 D2UCH –17. Dual Problem Cost ($) Iteration Time (minutes) Variables Number of PNL9 D1UCH –17..450. FINARDI. D3UCH .e.50% NLP problems less than D1UCH . was not employed.. requiring the enumerative process.938. Another important matter shown by Table 1 concerns dual optimal values: the (absolute value) difference between D2UCH and D1UCH is of $ 110.847. such as #7 and #16. for comparison purposes. 3. for the Brazilian case.339. An augmented Lagrangian technique seems in this case better. Comp. N.168 325 50 8208 864 D3UCH –17. After solving the dual problem. Vol. In particular. Appl. see for example [21. Such processes are often based on heuristics depending on the particular problem structure. Math. it is necessary to adjust the primal points corresponding to the dual optimum in order to find a primal point that is feasible. is clear in Table 1. 10 Ilha Solteira #7 and Tucuruí #16 power plants are the only ones where (22). i. a purification-like phase should be executed afterwards. We refer to [5] for a description of this technique in a similar context.

Math. we use the Fortran code PLCBAS. both the penstock stream-flow equations and power limit constraints).1 SOLVING THE UNIT COMMITMENT PROBLEM Sequential Quadratic Programming Algorithm In our work. Armijo condition (32).7 in [9]. and the “arc” search ends. if needed. it may no longer yield a descent direction for the merit function (30). NLP problems have ng + 2 variables and 2ng + 2 constraints. 2005 . It is called ‘second-order correction’ because c(x k + p k ) =O(| p k |2 ). Once the direction is computed. as well as maximum spillage levels. where ng is the number of generating units in the considered configuration (in our example ng ≤ 20).e. for instance considering active constraints as [23]. where Ak is a right inverse for the Jacobian matrix whose rows are the transposed gradients of the constraint functions at x k . or introducing slack variables in a trust-region SQP. we introduced additional constraints on the direction obtained from (29). in our implementation starting points only satisfy constraints (6) and (10) (i. we use the arc 0 < α| → x k + αp k − α 2 Ak c(x k + p k ) . 15.. SQP algorithms do not need starting feasible points.“main” — 2006/3/9 — 15:58 — page 338 — #22 338 5. Such bounds are related to each unit maximum turbined outflow. as in [24]. written with p k replaced by the corrected step p k . is eventually satisfied whenever pk is a descent direction.. 3. see Prop. Since p k is tangent to the arc at α = 0. 24. A remedy proposed by Maratos in [19] is to move towards the feasible region by using a second order model of the constraints. aimed at satisfying physical operating bounds. Vol. instead of searching the stepsize along the line x k + αpk . a correction strategy to avoid Maratos effect. N. We mention that more sophisticated alternatives would be possible. Although this modification preserves fast convergence. an active set QP solver described in [25]. In addition. Appl. The additional number of constraint evaluation required by our correction is compensated by Comp. A more simple strategy is to use the corrected step p k as follows: p k = p k − Ak c(x k + p k ) . The correction can be thought of as a constraint restoration step from x k + p k . but not the reserve constraint (12). the algorithm checks validity of (32) and then performs. For this reason. to prevent the starting quadratic program (29) to be infeasible. To solve each quadratic program (29).

as well as multiple operating zones. with inferior CPU times than those employed by Easy!. Emergency exits. with eps = 1.“main” — 2006/3/9 — 15:58 — page 339 — #23 ERLON C. available for academic applications from the Optimization Group at IME. Comp. For the whole Brazilian system. Vol. at each NLP solution. 6 Concluding Remarks We address in this work the problem of optimal commitment of hydraulic generating units in a hydrothermal power system. DA SILVA and CLAUDIA SAGASTIZÁBAL 339 the robustness and efficiency gained by the method. hydraulic losses. 3. the enumerative process appeared preferable for most power plants. the second order correction was used in average 30% of the iterations. 24. we gave an LR decomposition approach using variable duplication to uncouple difficult constraints. we did not take advantage of all resources available in Easy! solver11 . in terms of CPU times and obtained primal points. Brazil [26]. Math. With respect to the solution methodology. otherwise a finite difference approximation needs to be estimated at each simulation. namely the modeling and solution strategy for the hydraulic problem. We gave a detailed modeling for the production function of each hydro unit. Two main topics were discussed. Our implementation focused on the hydraulic subproblems. For our configuration. FINARDI. In our runs. which takes into account the effect of variable efficiency rates.. We obtained identical results. We assessed the decomposition method by implementing our approach and testing it on a real-life hydraulic configuration. This solver uses the augmented Lagrangian method described in [27]. 2005 . after 300 iterations of 600 simulations were also implemented. University of Campinas. Appl. we compare its performances the NLP solver Easy!. our own implementation was tailored for the structure of our specific problem. extracted from the Brazilian system. 0×10−8 . double in size to our test. EDSON L. The observed average values for achieving convergence were of 8 iterations and 10 simulations. Furthermore. N. We analyzed in detail the practical applicability of three decomposition schemes. However. tailrace levels. We use the following stopping test: opt_test = ∇ L k ∞ + c(x k )# ∞ ≤ eps. In order to assess our self-made algorithm. this approach might become too 11 Easy! Solver is more efficient when analytic derivatives are available for the simulations.

Li. X. August (1996). like the one given in our third decomposition scheme.J. 24.. Numerical Optimization. Lemaréchal and C.B. IEEE Transactions on Power Systems. PART I: Problem Formulation and General Solution Procedure. [2] C. [11] A. Sagastizábal. February (1999). Diniz. Appl. IEEE Trans. The authors acknowledge with grateful thanks the contributions made by Maria Elvira Maceira and Albert Cordeiro Geber de Melo from CEPEL. [9] J. A. 2001. 2nd Edition. Tolle. Math. (1996). Power Systems. Guan and R. Scheduling Hydro Power Systems with Restricted Operating Zones and Discharges Ramping Constraints. Maceira and C. Wright. Fahd. 14(1). A Flexible Approach to Short-Term Hydro-Thermal Coordination. Sagastizábal. Bundle Relaxation and Primal Recovery in Unit Commitment Problems – The Brazilian Case. Hsu. Ruzic.A Li. Hydro Unit Commitment in Hydro-Thermal Optimization. Annals of Operations Research Special – Volume on OR models for Energy Policy. may be preferable. Theoretical and Practical Aspects. Scheduling Hydrothermal Power Systems with Cascaded and Head-Dependent Reservoirs. REFERENCES [1] S. Universitext. IEEE Transactions on Power Systems. Sequential Quadratic Programming. Optimal Dispatch of Generating Units of the Itaipú Hydroelectric Plant. Comp.L. Vol.W. Belmont.J. Guan. E. C. J. August (1999). MA. Arce. 2005 . Springer-Verlag. Vuckovic. Soares. Netherlands. 17(1).T. T. [7] P. Unit Commitment Literature Synopsis. Gilbert. IEEE Transactions on Power Systems. Ni. C. Numerical Optimization. Berlin (2002). Springer Series in Operations Research (1999). (1999).F. [6] G. Research supported by Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) and Centro de Pesquisas de Energia Elétrica (CEPEL). [3] X. 12(2).E. Johnson. Svoboda. May (1997). [8] J. Nonlinear Programming. N. [10] D.“main” — 2006/3/9 — 15:58 — page 340 — #24 340 SOLVING THE UNIT COMMITMENT PROBLEM expensive and a combined strategy. Tseng and R.B. Sheble and G. Kluwer Academic Publishers. Acta Numerica. Rajakovic and A. IEEE Transactions on Power Systems. Boggs and J. Li. Nocedal and S. 14(3). February (2002). Bertsekas. Acknowledgments. 1–51. [5] A.P. 3. Athena Scientific.C. A. [4] E. M. IEEE Transactions on Power Systems. 9(1). Belloni. A. Planning and Management. Svoboda and C. February (1994).N. N. Bonnans. Ohishi and S. 11(3).P.

Dennis and R. SIAM. Friedlander. 263–282. RAIRO-Recherche Opérationnelle/Operations Research. Thesis.ime. [20] C. 27 (1993). August (1995). Applied Mathematics and Optimization 30 (1994). A. Heredia and N. Minimization of Functions Having Lipschitz Continuous First Partial Derivates. A Trust-Region Strategy for Nonlinear Equality Constrained Optimization. DA SILVA and CLAUDIA SAGASTIZÁBAL 341 [12] F. Finardi.A. Berlin (1978). 2005 . 399–427.D. [18] L. Celis. Pacific Journal of Mathematics. With Applications. Pellegrino. Parallel Processing Applied To the Planning of Hydrothermal Systems. Comp. Blondel and A.D.D. Nabona. Galiana. J. Ph. Exact Penalty Function Algorithms for Finite Dimensional and Control Optimization. France (2002).A.L. Maratos. Renaud and C. Use of the Code N1CV2.E. Mathematical Programming.“main” — 2006/3/9 — 15:58 — page 341 — #25 ERLON C. [24] M. Décomposition Par Les Prix et Lagrangien Augmenté: Une Association Efficace Pour le Traitement des Problèmes de Production.J. Renaud. IEEE Transactions on Parallel and Distributed Systems. 10(3). Martinez and S. FINARDI. Sagastizábal. Lemaréchal and C. IEEE Trans. 14 (2003). Zhuang and F. da Silva and E. 71–82 (1985). Ph. Sagastizábal. A Geometric Study of Duality Gaps. Numerical Analysis. Casas and C. 235–266. [26] Internet address: http://www. Power Systems. [19] N.M. 395–402. Tapia. 144–157.R. 4 (1991). [14] C. A Computationally Efficient Feasible Sequential Quadratic Programming Algorithm. Thesis. Bundle Methods Applied to the Unit Commitment Problem. 16(1-3) (1966). F. Bulletin EDF. [27] A. USA (1998). Lawrence. 11 (1976). [23] C.htm. EDSON L. IEEE Transactions on Power Systems. 3(2). 23–43.D. Superlinearly Convergent Variable Metric Algorithms for General Nonlinear Programming Problems. Série C.∼martinez/software. Appl. University of Maryland. (1996). 24. May (1988). Mathematical Programming. Lemaréchal. 721–729. Math. 90(3) (2001).C. Towards a More Rigorous and Practical Unit Commitment by Lagrangian Relaxation. Han. [16] S. Armijo.P. Renaud. Lemaréchal and A. Optimum Short-Term Hydrothermal Scheduling with Spinning Reserve Through Network Flows. Numerical Optimization.T. Pola. [25] E. Vol. A Fast Algorithm for Nonlinearly Constrained Optimization Calculations. [15] C. [22] F. 401–426.. [21] H. Santos.unicamp.J. System Modeling and Optimization. [13] E. J. [17] M. An algorithm for indefinite quadratic programming based on a partial Cholesky factorization. Powell. University of London (1978). A New Trust Region Algorithm for Bound Constrained Minimization. Springer-Verlag. N. Inria. 3.