194

Price-based unit commitment problem

Takayuki Shiina
$*$
Central Research Institute
of
Electric Power Industry,
2-11-1 IwadO-Kita, Komae, Tokyo 201-85:11 JAPAN
Abstract The unit commitment problem is determining the schedules for power generating units and the
generating level of each unit. The decisions concern which units to commit during each time period and at
what level to generate power to meet the electricity demand. The problem is atypical scheduling problem
in an electric power system. The electric power industry is undergoing restructuring and deregulation.
In this paper we develop a stochastic programming model which incorporates power trading. We need
to incorporate the uncertainties of electric power demand or electricity price into the unit commitment
problem. It is assumed that demand and price uncertainty can be represented by ascenario tree. We
propose a stochastic integer programming model in which the objective is to maximize expected prots. In
this model,
$\mathrm{o}\mathrm{n}/\mathrm{o}\mathrm{f}\mathrm{f}$
decisions for each generator are made at the rst stage. The approach to solving the
problem is based on Lagrangian relaxation and dynamic programming.
Keywords: Electric Power; Unit Commitment; Stochastic Programming; Lagrangian Relaxation
1 Introduction
The economic operation and planning of electric power generation occupy an important position in the
electric power industry. Wood and Wollenberg [17] oered a brief overview and applied many operations
research methods to real electric power problems. Electric power utilities have to maintain sucient capacity
to meet electricity demand during peak load periods. The unit commitment problem is determining the
schedules for power generating units and the generating level of each unit. The decisions concern which units
to commit during each time period, and at what level to generate power to meet the electricity demand.
The objective is to minimize the operational cost. This is the sum of the fuel and the start-up costs. The
problem is a typical scheduling problem in an electric power system.
Many types of optimization technique have been applied to the unit commitment problem. Delson and
Shahidehpour [6] illustrated how linear and integer programming were applied to power system engineer-
ing including generation scheduling, the allocation of reactive power supply and the planning of capital
investment in generation equipment. The paper by Sheble and Fahd [9] is a survey in this eld for the
period from the late 1960s to the early 1990s. They classied the solution techniques into exhaustive enu-
meration, priority list, dynamic programming, integer and mixed-integer programming,
$\mathrm{b}\mathrm{r}\mathrm{a}\dot{\mathrm{n}}$
ch-and-bound,
linear programming, network ow programming, Lagrangian relaxation. and expert
$\mathrm{s}\mathrm{y}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{m}\mathrm{s}/\mathrm{a}\mathrm{r}\mathrm{t}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{a}\mathrm{l}$
neural
networks. Among these approaches, the Lagrangian relaxation technique seems to be the most promising
because it decomposes the original problem into smaller subproblems. Muckstadt and Koenig [7] used this
approach by relaxing the demand constraints. Bard [1] used the Lagrangian relaxation to disaggregate the
problem into subproblems that were then solved by dynamic programming.
In these studies, the electricity demand at any period is known in advance. For many actual problems,
however, such an assumption is often unjustied. These data contain uncertainty and are represented as
random variables since the data represent information about the future. Takriti, Birge and Long [13] is
the rst paper that deals with the stochastic programming approach. Stochastic programming (Birge [3],
Birge and Louveaux [4], Shiina [10]
$)$
is a method for an optimization problem under uncertainty. Takriti,
Birge and Long [13] extended a technique used in the traditional deterministic unit commitment problem.
The uncertainty in demand is modeled by introducing a set of scenarios. The problem is decomposed
and solved using a Lagrangian relaxation type method, called a progressive hedging algorithm (Rockafellar
and Wets [8]
$)$
. For each scenario, the relaxed deterministic problem is then decomposed into a single
generator subproblem using Lagrangian relaxation. Each subproblem can be solved eciently by dynamic
programming. Takriti and Birge [15] generalized this approach and showed that the duality gap of the
relaxation is bounded by a certain constant. Shiina [11] developed a stochastic programming model in which
$\mathrm{o}\mathrm{n}/\mathrm{o}\mathrm{f}\mathrm{f}$
decisions for generators are made at the rst stage. Takriti and Birge [14] developed a technique for
rening the solution obtained by solving the Lagrangian relaxation problem. Their approach is to select
a schedule among the feasible solutions sought up to the latest iteration, and to combine them so that
the demand constraint can be met. The suggested model is a mixed-integer programming problem which is
solved by branch-and-bound method. Though the numerical results indicated improvements, the approach is
no more than heuristic. An ecient method for nding feasible schedules, and for rening them, is required.
$\mathrm{E}$
-mail address: [email protected]
1373 4, 2004 G/194-202
185
Shiina and Birge [12] propose a new algorithnt, based on the Dantzig-Wolfe decomposition (Dantzig-Wolfe
[5]
$)$
and the column generation approach (Barnhart et al. [2]), to solve the stochastic unit commitment
problem.
The electric power industry is undergoing restructuring and deregulation. Takriti, Krasenbrink and Wu
[16] presented a stochastic programming model that incorporates power trading and uncertainty in electricity
demand and spot prices. They modeled the spot market as two generating units which represent buying
and selling. In this model, the
$\mathrm{o}\mathrm{n}/\mathrm{o}\mathrm{f}\mathrm{f}$
decisions of generating units are dened as recourse variables. But
some types of generators such as coal red units involve time delay before the generators become available.
In this paper we develop a new stochastic programming model in which power trading is incorporated
and switching decisions for generators are made at the rst stage. We assume that there are hourly spot
markets for electricity so that buying or selling do not aect spot price. We need to incorporate the
uncertainty of electric power demand or electricity price into the unit commitment problem. It is assumed
that demand and price uncertainty can be represented by a scenario tree. We propose a stochastic integer
programming model in which the objective is to maximize expected prots. In this model, no articial units
are necessary to model the spot market. This problem is formulated as a multi-stage stochastic quadratic
integer programming problem because the fuel cost function is assumed to be a convex quadratic function.
The solution approach is based on Lagrangian relaxation method and dynamic programming. The problem
is decomposed into subproblems of single units. The feasible schedule is obtained by solving dynamic
programming on a scenario tree. To rene the solution obtained by dynamic programming, we solve an
economic dispatch problem in which the equality demand constraint is relaxed to the inequality constraint
with upper and lower limit. To solve this problem we develop an algorithm which combines the lambda
iteration method and golden section.
2 Uncertainty in Electricity Demand and Price
We assume the duration of the planning horizon in
$T$
periods. Since the electricity demand or spot market
price at any point in a period may be uncertain, we have to model the unit commitment problem as a
stochastic programming problem.
To model uncertainty, we dene the total demand for electricity and spot market price during period
$t$
as a random variable
$\tilde{d}_{t}(\geq 0)$
and
$\tilde{e}_{t}(\geq 0)$
, respectively. It is assumed that
$\tilde{d}_{t}$
and
$\tilde{e}_{t}$
are dened on a known
probability space and have a nite discrete distribution. Let
$d_{t}$
and
$e_{t}$ be realizations of random variable
$\tilde{d}_{t}$
and
$\tilde{e}_{t}$
. A sequence of the realization of electricity demand and spot market price
$(d_{1}, \ldots, d_{T}, e_{1}, \ldots, e\tau)$
is
called scenario. It is assumed that we have a set of
$S$
scenarios. We use a superscript
$s$
to denote an index of
scenario
$s$
. We associate a probability $p_{s}$
with each scenario
$d^{s}=$
(dSl, $\ldots$ ,
$d_{T}^{s}$
),
$e^{s}=(e_{1}^{s}, \ldots,e_{T}^{s})$
, $s=1$ , $\ldots$ ,
$S$
.
The scenarios are described using a scenario tree.
If two scenarios
$(d^{\mathit{8}1}, e^{s_{1}})$
,
$(ds_{2}, e\sim)s\circ$
,
$(s_{1}\neq s_{2})$
satisfy the following conditions
$(d_{1}^{s_{1}}$
, . . . ,
$d_{t}^{\mathit{8}1}$ $)$ $=$ $(d_{1}^{s}\underline, \ldots, d_{t}^{s_{2}})$
,
$(e_{1}^{s_{1}}$
, . . . ,
$e\mathrm{j}^{1}$ $)$ $=$ $(e_{1}^{s_{-}}$
, . . . ,
$e_{t}^{S\underline{\Phi}}$ $)$
for a period
$t$
, they are indistinguishable up to period
$t$
. The decisions made
for scenario
$(d^{\epsilon_{1}}, e^{s_{1}})$
up to period
$t$
must be the same as those made for
$(ds_{2}, e\sim)s\mathrm{o}$
up to period
$t$
. These
two scenarios are said to be included in the same bundle at period
$t$
. The set of indices for the scenarios
$\{$
1, $\ldots$ ,
$S\}$
at each period can be partitioned into disjoint subsets which represent scenario bundles. We
dene
$B(s, t)$
to be the bundle in which scenario
$s$
is a member at period
$t$
.
If $B(s, t)=B(s, t)$ and
$\mathrm{B}(\mathrm{s},\mathrm{t}t+1)\neq B(s, t+1)$
, $s<s,$ the period $t+1$ is a point when scenario
$s$
splits from the other scenario
$s$
. The scenario
$s$
is called a predecessor of scenario
$s$
. If there are multiple
predecessors for
$s$
, we dene the scenario with the lowest index as the predecessor of
$s$
. The predecessor of
scenario
$s$
is denoted by Pred(s). The period
$\mathrm{r}(\mathrm{s})$
is dened as the rst period in which a scenario
$s$
does
not share a bundle with another scenario
$s<s.$
For scenario 1, we dene $\tau(1)=1$ and Pred(l)
$=1.$
3 Stochastic Unit Commitment Problem
We assume that there are I generating units. The status of unit
$i$
at period
$t$
is represented by the 0-1
variable
$u_{\dot{1}t^{-}}^{s}$
Unit
$i$
is on at time period
$t$
under scenario
$s$
, if
$u_{\dot{|}t}^{s}=1,$
and o if
$u_{\dot{l}t}^{s}=0.$
The power
generating level of the unit
$i$
at period
$t$
under scenario
$s$
is represented by
$x_{\dot{1}t}^{s}(\geq 0)$
.
In this model, we assume that the amount of power which can be traded with the power pool is limited.
We denote the amount of power which can be bought from power pool at period
$t$
by
$\underline{d_{t}}$
, the amount which
can be sold to power pool at period
$t$
by
$\overline{d_{t}}$
.
$\sum_{i=1}^{I}xj_{t}$ $\geq d_{t}^{s}-\underline{d_{t}}$
, $t=1$ , $\ldots$ ,
$T$
, $s=1$ , $\ldots$ ,
$S$
(1)
1 mE
If
$\sum_{i=1}^{I}x_{it}^{s}.<d_{t}^{s}$
, the sum of the levels of generation is less than the demand. In this case, the demand
shortage is satised by buying from the power pool.
$\sum_{i=1}^{I}x\mathrm{j}_{t}\leq d_{t}^{s}+\overline{d_{t}}$
, $t=1$ , $\ldots$ ,
$T$
, $s=1$ , $\ldots$ ,
$S$
(2)
If
$\sum_{i=1}^{I}x_{it}^{s}>d_{t}^{s}$
, the sum of the levels of generation is greater than the demand. In this case, the surplus
of demand is sold to the power pool.
Rapid changes of temperature are not allowed for thermal units. When unit
$i$
is switched on, it must
continue to run for at least a certain period Lit These minimum up-time constraints are described in (3).
$u_{it}-u_{i,t-1}$
$\leq$ $u_{i\tau}$ , $\tau=t+1$, $\ldots$ , $\min\{t+L_{i}-\mathrm{I}, T\}$ , $t=2$ , $\ldots$ ,
$T$
(3)
Similarly, when unit
$i$
is switched o, it must continue to be o at least
$l_{i}$
periods. These constraints are
called minimum down-time constraints (4).
$u_{*,t-1}.-uit$
$\leq$ $1-u_{i\tau}$
, $\tau=t+1$ , $\ldots$ , $\min\{t+l_{i}-1, T\}$ , $t=2$,
$\ldots,$
$T$
(4)
Let
$[q_{\dot{1}}, Q_{i}]$
be an operating range of the generating unit
$i$
. That is,
$x_{il}^{\theta}$
has to satisfy the following
constraints (5).
$q_{i}u|.t\leq x_{it}^{s}\leq Q_{i}u_{it}$
,
$j$
$=1$
, $\ldots$ , $I,t=1$, $\ldots$ ,
$T$
, $s=1$ ,
$\ldots,S$
(5)
For two scenarios which are members of the same bundle, the decision variables must be the same.
$x_{it}^{s_{1}}=x_{i}^{s_{\frac{}{t}}}$
, $i=1$ , $\ldots$ ,
$I$
, $t=1$ , $\ldots$ ,
$T$
,
$/s_{1}$
,
$S.?\vee\in\{1$
, . . . , $S\}$ ,
$s_{1}\neq s_{2}$
, $B(s_{1}, t)=B(s_{-},, t)$
(6)
This type of constraint is called a nonanticipativity constraint, or a bundle constraint.
The fuel cost function
$f_{i}(x_{it}^{\epsilon})$
is given by a convex quadratic function of
$x_{it}^{s}$
. This function relates to
the output of power generated by unit
$i$
and depends on the consumption of fuel. The fuel cost function
is regarded as convex quadratic, since the incremental fuel cost is a linear increasing function of
$x_{t}^{\epsilon}\dot{.}$
. The
start up cost function
$g_{i}(u_{i,t-1}, u_{it})$
satises the condition $g:(0,1)>0$, $g_{i}(0,0)=0$ , $g_{i}(1,0)=0,g_{i}(1,1)=0.$
The mathematical formulation of the stochastic unit commitment problem is described as follows.
$x_{it}^{s_{1}}=x_{i}^{s_{\frac{}{t}}}$
, $i=1$ , $\ldots$ ,
$I$
, $t=1$ , $\ldots$ ,
$T$
,
$\forall s_{1}$
, $s_{\vee}.?\in\{1, \ldots, S\}$
,
$s_{1}\neq s_{2}$
, $B(s_{1}, t)=B(s_{-},, t)$
(6)
This type of constraint is called a
nonanticipativity constraint, or abundle constraint.
The fuel cost function
$f_{i}(x_{it}^{\epsilon})$
is given by aconvex quadratic function of
$x_{it}^{s}$
. This function relates to
the output of power generated by unit
$i$
and depends on the consumption of fuel. The fuel cost function
is regarded as convex quadratic, since the incremental fuel cost is alinear increasing function of
$x_{t}^{\epsilon}\dot{.}$
. The
start up cost function
$g_{i}(u_{i,t-1}, u_{it})$
satises the condition $g:(0,1)>0$, $g_{i}(0,0)=0,g_{i}(1,0)=0,g_{i}(1,1)=0.$
The mathematical formulation of the stochastic unit commitment problem is described as follows.
(Stochastic Unit Commitment Problem)
$\min$
$\sum_{I}^{S}p_{s}\sum_{ts=1=1}^{T}\sum_{\dot{l}=1}^{I}+f_{i}(x_{t}^{\epsilon}.\cdot)u_{t}^{s}\dot{.}+g_{i}(u_{i,t-1}, u_{i,t})\}-\sum_{\epsilon=1}^{S}p_{s}\sum_{t=1}^{T}e_{t}^{s}(\sum_{\dot{|}=1}^{I}xi_{t}-d_{t}^{s})$
subject to
$\sum xj_{t}$ $\geq d_{t}^{*}-\underline{d_{t}}$
, $t=1$ , $\ldots$ ,
$T$
, $s=1$ , $\ldots$ ,
$S$
$i=1I$
$\dot{.}\sum_{=1}x_{it}^{\mathit{8}}\leq d_{t}^{s}+\overline{d_{l}}$
, $t=1$ ,
$\ldots,T$
, $s=1$ , $\ldots$ ,
$S$
$u_{iC}-u:,t-1\leq u_{i\tau}$ ,
$\tau$
$=t+1$ , $\ldots$ ,
rnin{t
$+L_{\dot{\iota}}-1,T$
},
$i=1$, ...,
$I$
, $t=2$, ...,
$T$
$u:,t-.1-u:t\leq 1-u_{i_{\mathcal{T}}}$ , $\tau=t+1$, $\ldots$ , $\min\{t+l:-1,T\}$,
$i=1$ , ...,
$I$
, $t=2$, $\ldots$ ,
$T$
$u:t\in\{0,1\},i=1$, $\ldots$ ,
$I$
, $t=1,$ . ..,
$T$
$q:u_{\dot{1}t}\leq x_{\dot{\iota}t}^{s}\leq$
Q:uit, $i=1$ , $\ldots$ ,
$I$
, $t=1$ , $\ldots$ ,
$T$
, $s=1$, $\ldots$ ,
$S$
$x_{*t}^{\epsilon_{1}}.=x_{\dot{1}t}^{\epsilon_{2}}$
, $i=1$, $\ldots$ ,
$I$
, $t=1$, $\ldots$ ,
$T$
,
$ls_{1},s_{2}\in\{1, \ldots, S\}$ ,
$s_{1}\neq s_{2}$
, $B(s_{1}, t)=B(s_{2},t)$
The problem results in a large scale mixed integer nonlinear programming problem that combines
$S$
deter-
ministic unit commitment problems. The objective is to maximize the expected operational prot over all
possible scenarios.
4 Solution Algorithm
4.1 Lagrangian Relaxation
First, we consider solving the stochastic unit commitment problem using a Lagrangian relaxation approach.
Instead of solving the problem directly, we solve the Lagrangian relaxation which results from relaxing the
137
demand constraints. Let
$\lambda_{t}^{s}$ $(\geq())$
,
$\mu_{t}^{s}(\geq 0)$
be Lagrange multipliers associated with constraints (1), (2),
respectively. The Lagrangian relaxation problem is shown as follows.
(Lagrangian Relaxation Problem)
$L( \lambda, \mu)=\min$
$\sum_{-,s--1}^{S},p_{s}\sum_{t=1}^{T}\sum_{i=1}^{I}\{f_{i}(x_{it}^{\mathit{8}}.)u_{it}^{s}+g_{i}(u_{i,t-1}, u_{i,t})\}-\sum_{s=1}^{\mathrm{q}}.p_{s}\sum_{t=1}^{T},t(s\sum_{i=1}^{I}x_{?.\mathrm{f}}^{\mathit{8}}.-d_{t}^{s})$
$- \sum_{s=1}^{S}.\sum_{t=1}^{T}\lambda_{t}^{s}(\sum_{i=1}^{\mathit{1}}x_{it}^{s}-d_{t}^{s}+\underline{d_{t}})-\sum_{s=1}^{S}\sum_{t=1}^{T}\mu_{t}^{s}(-\sum_{i=1}^{I}x;_{t} +d_{t}^{s}+\overline{d_{t}})$
subject to $u_{ii}-u_{i,t-1}\leq u_{i\tau}$ , $\tau=t+1$ , $\ldots$ , $\min\{t+L_{i}-1, T\}$ ,
$i=1$ , $\ldots$ ,
$I$
, $t=2$, $\ldots$ ,
$T$
$u_{i,t-1}-u_{it}\leq 1-u_{i\tau}$ ,
$\tau=t$
% 1, $\ldots$ , $\min\{t+l_{i}-1, T\}$ ,
$i=1$ , $\ldots$ ,
$I$
, $t=2$, $\ldots$ ,
$T$
$u_{\dot{\iota}t}\in\{0,1\}$
, $i=1$ , $\ldots$ ,
$I$
, $t=1$ , . $.=$ ,
$T$
$q_{*}.u_{it}\leq x_{it}^{s}\leq Q_{:}u_{it}$
, $i=1$, $\ldots$ ,
$I$
, $t=1$ , $\ldots$ ,
$T$
, $s=1$ , $\ldots$ ,
$S$
$x_{t}^{\epsilon_{1}}\dot{.}=x_{\dot{1}t}^{\mathrm{f}_{-}}.$
,,
$i=1$ , $\ldots$ , $I,t=1$, $\ldots$ ,
$T$
,
$\forall s_{1}$
,
$s_{2}\in\{1, \ldots, S\}$
,
$s_{1}\neq s_{2}$
,
$B(s_{1}, t)=B(s_{9}, t)\sim$
This relaxation decomposes the problem into smaller single generator subproblems. The objective func-
tion of Lagrangian relaxation problem
$\mathrm{L}(\mathrm{A}, \mu)$
can be rewritten as follows.
$\mathrm{L}(\mathrm{A}, \mu)=\min$ $\sum_{i=1}^{I}\sum_{t=1s}^{T}\sum_{=1}^{S}\mathit{1}s\{f_{i}(x_{it}^{s})u_{t}^{s}\dot{.}-e_{t}^{\epsilon}x_{it}^{s}\}$ $- \lambda_{t}^{s}x_{it}^{s}+\mu_{t}^{s}x_{it}^{s}]+\sum_{i=1}^{I}\sum_{t=1}^{T}g.\cdot(u_{i,t-1},u_{*,t}.)$
$s$ $T$
$+ \sum$
5
$\{p_{stt}e^{s}d^{s}-\lambda_{t}^{s}(-d_{t}^{s}+d_{t}I-\mu_{t}^{s}(d_{t}^{s}+\overline{d_{t}})\}$
(7)
$s=1t=1$
The last term of objective function (7) is a constant. So the function (7) is separable in each unit.
The Lagrangian relaxation problem can be solved by calculating dynamic programming on the scenario
tree. First , we solve the following generation level decision problem to obtain optimal
$\hat{x}$
:t
$t=\tau(s)$ , $\ldots$ ,
$T$
, $s=$
$1$
, . .
$\cdot$
.

$S$
. The problem is a convex quadratic programming problem which can be solved easily.
(Generation Level Decision Problem for Unit
$i$
at Period
$t$
under Scenario
$s$
)
$\min$
$\sum_{s^{J}\in B(s,t)}p_{s^{JJ}}\{f_{i}(x_{it}^{s})-e_{t}^{s}x_{i\ell}^{s}\}-(\lambda_{t}^{\epsilon}-\mu_{t}^{s})x_{il}^{s}$
subject to
$q:\leq x_{it}^{s}\leq Q_{i}$
Let
$\hat{x}$
:
$t$ , $t=$ (1), $\ldots$ , $T,s=1$ , $\ldots$ ,
$S$
be solutions of the generation level decision problem. By setting
$x_{it}^{s}\prime\prime=\hat{x}\mathit{7}_{t}$
, $t=1,$ . . . ,
$T,\forall s\prime\prime\in B(s, t)$
, the scenario bundle constraints can be satised.
Then the binary variables
$u_{it}^{s}$
, $t=1$ , $\ldots$ ,
$T$
, $s=1$, $\ldots$ ,
$S$
are determined. The calculation of dynamic
programming is done by the following recursive equations. A unit
$i$
must be in one of
$L_{:}+l_{:}$
states. The
rst
$L_{i}$
states mean that the unit
$i$
is on, and the last
$l_{i}$
states mean that the unit
$i$
is o. Let Ci(t,
$k$
) be
the optimal cost of unit
$i$
under the scenario
$s$
from stage
$t$
to the end of the horizon, if unit 2 is in state
$k$
at stage
$t$
. The recursive equations are dened as shown in (8).
$C_{*}.(t, k)=$
$\{$
$C_{\dot{*}}(t+1, k+1)+ \sum_{\epsilon\in\{\epsilon^{l}\tau(\epsilon)\leq t\}},(\sum_{\in\epsilonB(\epsilon,t)}p_{\epsilon}\prime\prime)\{f:(\hat{x}_{\dot{\iota}t}^{\mathrm{g}})-e_{t}^{\epsilon}\hat{x}_{\dot{l}t}^{s}\}-(\lambda_{t}^{\epsilon}-\mu_{t}^{s})\hat{x}_{\dot{l}t}^{\epsilon}$
if
$1\leq k<L:$
$\min[C\dot{.}(t+1, k)+$ $\sum$ $( \sum p_{\mathit{8}}\prime\prime)\{f_{i}(\hat{x}_{t}^{\epsilon}\dot{.})-e_{t}^{\epsilon}\hat{x}_{it}^{\theta}\}-(\lambda_{t}^{\epsilon}-\mu_{t}^{\epsilon})\hat{x}_{t}^{l}.\cdot$
$c_{:}(t+1, k+1)+ \cdot\sum^{\theta\in\{s\cdot\tau(s^{l})\leq t\}}(\sum p_{s}\prime\prime)s\in B(\epsilon,t)\{f_{:}(\hat{x}\mathit{7}_{t})-e_{t}^{\epsilon}\hat{x}_{it}^{\epsilon}\}-(\lambda_{t}^{\mathit{8}}-\mu_{t}^{\ell})\hat{x}i_{t}l$
if
$k$ $=L_{i}$
$\epsilon\in\{s:\tau(s)\leq t\}$ $\epsilon\in B(s,t.)$
$c_{:}(t+1, k+1)$ if
$L_{i}<k<L_{i}+l\dot{.}$
$\min\{C_{\dot{l}}(t+1, k),c_{:}(t+1,1)+g_{\dot{1}}(0,1)\}$ if
$k=L_{*}$
. $\dotplus l_{i}$
(8)
In the recursive equation of dynamic programming (8) at period
$t$
, the probability of scenario
$s(\tau(s)\leq t)$
is replaced by
$\sum_{s^{l}\in B(\epsilon,t)}.p_{s^{l}}$ . This procedure makes possible an exact calculation of dynamic programming.
We dene the optimal $u_{it},i=1$ , $\ldots$ ,
$I$
, $t=1$ , $\ldots$ ,
$T$
obtained from (8) by
$\text{\^{u}}_{i}t,i=1$
, $\ldots$ ,
$I$
,
$t$ $=1$
, $\ldots$ ,
$\mathrm{A}$
199
Then we maximize the Lagrangian dual function because the function is concave.
(Lagrangian Dual Problem)
$\max L(\lambda, \mu)=$
$\sum p_{s}\sum\sum\{f_{i}(\hat{x}_{it}^{s})\hat{u}_{it}^{s}+g_{i}(\hat{u}_{i,t-1}, u_{i,t}^{\mathrm{A}})\}-STIE$ $p_{\mathit{8}} \sum e_{t}^{s}(\sum\hat{x}_{it}^{e}-d_{t}^{s})I$
$s=1- \sum_{s=1}^{S}\sum_{t=1}^{t-}\lambda_{t}^{s}(\sum_{i=1}^{I}\hat{x}_{it}^{\epsilon}-d_{t}^{s}+\underline{d_{t}})-\sum_{\epsilon=1}^{S}\sum_{t=1}^{T}\mu_{t}^{s}(-\sum_{i=1}^{I}\hat{x}\tau^{-1}i=1s=17_{t}t=1i=1+d_{t}^{s}+\overline{d_{t}})$
subject to
$\lambda\geq 0,\mu\geq 0$
The optimal obj ective value of the Lagrangian relaxation problem provides the lower bound of the original
objective value if the solution satises the constraints (1) and (2). Then we wish to obtain the largest possible
lower bound. However the function is not dierentiable, so we use the subgradient optimization technique.
Let
$\lambda^{0}$
,
$\mu^{0}$
be any initial multiplier. The Lagrangian multipliers are rened for $l=1,2$, $\ldots$ , as shown in (9),
where
$\alpha\iota$
denotes a step size at iteration
$l$
.
$\lambda^{t+1}=\lambda^{t}+\alpha\iota\xi^{l}$
,
$\mu^{t+1}=\mu^{t}+\alpha\iota\overline{\xi^{l}}$
(9)
A vector
$(\xi^{t}, \xi)$
is called subgradient of
$L(\lambda, \mu)$
at
$(\lambda\iota, \mu)\iota$
if inequality (10) holds.
$L(\lambda, \mu)\leq L(\lambda^{t}, \mu^{t})+(\lambda^{\mathrm{T}}-\lambda_{t}^{\mathrm{T}})\xi^{t}+(\mu^{\mathrm{T}}-\mu_{l}^{\mathrm{T}})\overline{\xi^{t}}$
(10)
In the Lagrangian dual problem,
$(\xi^{t}, \xi^{t})$
is given as follows.
$\underline{\xi^{t}}=\sum_{\dot{l}=1}^{I}\hat{x}_{it}^{s}+d_{t}^{s}-\underline{d_{t}},\overline{\xi^{t}}=\sum_{i=1}^{I}" i_{t}-d_{t}^{s}-\overline{d_{t}}$
(11)
The nessesary and sucient condition for the convergence of the sequence
$(\lambda^{tt}, \mu)$
to obtain the optimal
solution is shown as follows.
$\alpha_{t}||\xi^{t}||arrow$
Oand
$\sum_{t}\alpha_{l}||(\underline{\xi^{t}},\overline{\xi^{t}})||arrow 0$
(12)
A geometric convergence rate can be achieved if we set
$\alpha_{l}$
as (13), where
$L^{*}$
denotes an optimal value of
$L(\lambda, \mu)$
.
$\alpha\iota=\frac{L^{*}-L(\lambda^{t})}{||\xi^{t}||^{\mathrm{o}}\sim}$
(13)
But we cannot know the value of
$L^{*}$
in advance. Instead, we adopt a heuristic for selecting the step length.
$\alpha_{l}=\frac{\theta^{t}(UB-L(\lambda^{t}))}{||\xi^{t}||^{2}}$
(14)
In this expression,
$UB$
denotes an upper bound of
$\mathrm{L}(\mathrm{A}, \mu)$
and
$\theta^{t}$
is chosen between 0 and 2.
This procedure terminates if the dierence between the upper bound and lower bound is relatively small.
For further details, the reader should refer to the paper by Takriti, Krasenbrink and Wu [16].
4.2 Economic dispatch problem with power trading
The optimal solution obtained from solving the Lagrangian relaxation problem may not give a primal feasible
solution. After solving the Lagrangian relaxation problem by dynamic programming, the set of solutions
$\hat{x}$
:t
\^ut:t, $i=1$, $\ldots$ ,
$I$
, $s=1$, $\ldots$ ,
$S$
, $t=$ t(s), $\ldots$ ,
$T$
is obtained. To modify the level of power generation
$\hat{x}i_{t}$
,
we solve the following economic dispatch problem with power trading for $s=1,$ $\ldots$ , $S,t=$ t(s), $\ldots$ ,
$T$
.
(
$\mathrm{E}\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{o}_{\min\sum f_{\dot{\iota}}(x_{it}^{\epsilon})\hat{u}_{il}^{\epsilon}-e_{\mathrm{t}}^{\epsilon}(\sum x_{\dot{1}t}^{\theta}-d_{t}^{\epsilon})}\mathrm{m}\mathrm{i}\mathrm{c}\mathrm{D}\mathrm{i}\mathrm{s}\mathrm{p}\mathrm{a}\mathrm{t}\mathrm{c}\mathrm{h}\mathrm{P}\mathrm{r}\mathrm{o}\mathrm{b}1\mathrm{e}\mathrm{m}\mathrm{w}\mathrm{i}\mathrm{t}\mathrm{h}\mathrm{P}\mathrm{o}\mathrm{w}\mathrm{e}\mathrm{r}II$
Trading at Period
$t$
under Scenario
$s$
)
subject to
$\sum,i_{t}\mathit{2}i_{-,I}^{-}1$
$d_{t}^{s}-d$
j
$t\dot{.}=1=1$
, $\ldots$ ,
$T$
, $s=1$ , $\ldots$ ,
$S$
$. \sum_{i=1}^{I}|--1xit$
$\leq d_{t}^{s}+\overline{d_{t}}$
, $t=1$, $\ldots$ ,
$T$
, $s=1$, $\ldots$ , 5
$q_{i}\hat{u}_{it}\leq x_{t}^{s}\dot{.}\leq Q_{i}\hat{u}_{\dot{\iota}t}$
, $i=1$, $\ldots$ , 1, $t=1$ , $\ldots$ ,
$T$
, $s=1,$ $\ldots$ ,
$S$
189
Lagrangian relaxation
. Step 0. Given Lagrangian multipliers
)
0, $\mu^{0}(>0)$
.
. Step 1. Decompose Lagrangian dual into smaller single-generator subproblems by relaxing con-
straints (1) and (2).
. Step 2. Solve generation level decision problem to obtain
$ij_{t}$
, $s=1$ , $\ldots$ ,
$S$
, $t=$ P(a)
$)$
$\ldots$ ,
$T$
. Then
solve dynamic programming problem by recursive equations (8) to obtain
$\text{\^{u}}_{i}t$
, $i=1$, $\ldots$ , $I,t=$
$\tau(s)$
,
$\ldots,T$
.
. Step 3. Update Lagrange multiplier by subgradient optimization (9). Go to step 1.
Figure 1: Algorithm of Lagrangian relaxation method
If the value of
$\sum_{i=1}^{I}x_{it}^{s}$
is xed and the constraints
$\sum_{i=1}^{\mathit{1}}x_{il}^{s}\geq d_{t}^{\mathit{8}}-\underline{d_{t}}$
,
$\sum_{i=1}^{I}x_{it}^{\epsilon}\leq d_{t}^{s}+\overline{d_{t}}$
do not exist,
the problem can be solved by the lambda iteration method (Wood and Wollenberg [17]). The lambda
iteration method, which is based on the method of indeterminate coecents, seeks the optimal value of
undetermined multipliers using binary search. We develop a solution method which combines the lambda
iteration method and the golden section. In our approach, the economic power dispatch problem with power
trading is formulated as a parametric optimization problem. Then the golden section is applied to select
an optimal parameter. The lambda iteration method is also modiied in order to treat the constraints
$\sum_{i=1}^{I}x_{it}^{s}\geq d_{t}^{\epsilon}-\underline{d_{t}}$
, $\sum_{=1}$

$x_{it}^{s}\leq d_{t}^{s}+\overline{d_{t}}$
.
(Parametric Optimization Problem:
$\mathrm{P}(\alpha)$
)
$z(\alpha)=rni\mathrm{n}$
$\sum_{I}^{I}f_{i}(x_{it}^{s})\hat{u}_{it}^{s}-e_{t}^{\epsilon}(\sum_{ii=1=1}^{I}x_{it}^{g}-d_{t}^{\epsilon})$
subject to
$\sum_{i=1}l_{it}s=\alpha$
$q_{i}\hat{u}_{it}\leq x_{\mathrm{i}t}^{s}\leq Q_{i}\hat{\mathrm{e}\iota}_{it}$
, $i=1$ , $\ldots$ ,
$I$
We solve the parametric optimization problem in the following range of parameter
$\alpha$
.
$\max\{\sum_{i=1}^{I}q_{i}\hat{u}_{it}, d_{t}^{s}-\underline{dt}\}\leq\alpha\leq\min\{\sum_{i=1}^{I}Q_{i}\hat{u}_{it}, d_{tt}^{\theta}+\overline{d}\}$
The algorithm to solve economic dispatch problem with power trading is shown in Figure 2.
Algorithm to solve Economic Dispatch Problem with Power Trading by Golden Section
. Step 0. Set
$\underline{\alpha}=\max\{\sum_{i=1}^{I}q_{i}\hat{u}_{it}$
,
$d_{t}^{s}-d\mathrm{J},$
$\overline{\alpha}=\min\{\sum_{i=1}^{J}Q_{i}\hat{u}_{it}, d_{t}^{s}+\overline{d_{t}}\}$
,
$\epsilon>0$
and $k=0.$
. Step 1. Set
$\alpha_{1}=\underline{\alpha}+F_{1}$ $(\overline{\alpha}- 0)$
,
$\alpha_{2}=\underline{\alpha}+F_{)}..(\overline{\alpha}-\underline{\alpha})$
, where
$F_{1}= \frac{3-\sqrt{5}}{2}$
,
$F_{2}=L\mathrm{s}_{2}-\underline{1}$
.
. Step 2. If
$\overline{\alpha}-\underline{\alpha}<\epsilon$
, then stop.
. Step 1. Set
$\alpha_{1}=\underline{\alpha}+F_{1}(\overline{\alpha}-\underline{\alpha})$
,
$\alpha_{2}=\underline{\alpha}+F_{)}..(\overline{\alpha}-\underline{\alpha})$
, where
$F_{1}= \frac{3-\sqrt{5}}{2}$
,
$F_{2}=L\mathrm{s}_{2}-\underline{1}$
.
. Step 2.
$\mathrm{I}\mathrm{f}\overline{\alpha}-\underline{\alpha}<\epsilon$
, then stop.
. Step 3. If $z(1)<z(\alpha_{2})$ , then set
$\overline{\alpha}=\alpha_{2}$
, $\alpha_{\underline{7}}=\alpha_{1}$
and
$\alpha_{1}=\underline{\alpha}+F_{1}(\overline{\alpha}-\underline{\alpha})$
. If
$z(\alpha_{1})\geq z$
(2).
then set
$\underline{\alpha}=\alpha_{1}$ , $\alpha_{1}=\alpha_{\sim}$
and
$\alpha_{2}=\underline{\alpha}+$
F2
$(\mathrm{a}-\underline{\alpha})$
. Go to step 2.
Figure 2: Algorithm to solve Economic Dispatch Problem with Power Trading
In step 3 of the algorithm shown in Figure 2, parametric optimization problem
$\mathrm{P}(\mathrm{a})$
is solved. To solve
problem
$\mathrm{P}(\alpha)$
, the lambda iteration method is modied. We briey summarize the outline of our method. We
consider the following economic dispatch problem in which the constraints
$q_{i}\hat{u}_{it}\leq x_{it}^{s}\leq Q_{i}\hat{u}_{i\mathrm{f}}$
, $i=1,$ . .. ,
$I$
are relaxed.
200
(Relaxed economic Dispatch Problem)
$\min$ $\sum_{i=1}^{I}f_{i}(x_{it}^{s})\hat{u}_{it}^{s}-e_{t}^{s}(\sum_{i=1}^{I}x_{it}^{\epsilon}- \mathrm{c}/;)$
subject to
$\phi=\sum_{i=1}^{I}x$
:
-
a
$=0$
To solve the relaxed economic dispatch problem, the indeterminate coecents method is applied. The
constraint function
$\phi$
is multiplied by an undetermined multiplier
$\pi$
and added to the objective function.
This function is known as the Lagrange function denoted by
$L$
.
$L= \sum_{i=1}^{I}f_{i}(x_{it}^{s})\hat{u}i_{t}$ $-e_{t}^{s}( \sum_{\dot{l}=1}^{I}x_{\dot{1}t}^{S}-d_{t}^{s})-\pi(\sum_{i=1}^{I}x_{\dot{*}t}^{\epsilon}-\alpha)$
The nessesary condition for an extreme value of the objective function is shown as follows.
$\frac{\partial L}{\partial x_{it}^{s}}$ $=$
$f^{1}.\cdot(x_{\dot{1}t}^{S})\hat{u}:t-e_{t}^{\epsilon}-\pi=0$
, $i=1,$ . . . ,
$I$
(15)
$\frac{\partial L}{\partial\pi}$
$=$ $\phi=\sum_{\dot{|}=1}^{I}x\mathit{7}_{t}-\alpha=0$
(16)
The lambda iteration method seeks an optimal solution by obtaining the optimal
$\pi$
using the binary search.
In our algorithm, the lambda iteration method is modied so as to satisfy the constraints
$q_{i}\hat{u}_{it}\leq x_{it}^{\epsilon}\leq$
QiUit, $i=1$, $\ldots$ ,
$I$
. The algorithm is shown in Figure 3.
Modied Algorithm of Lambda Iteration
. Step 0. Suppose
$Eti_{=1}q_{i}\hat{u}_{it}-$
$x<0$ and
$\sum j_{=1}Q_{i}\hat{u}_{it}-\alpha>0.$
Solve
$f_{i}(q_{i})\hat{u}_{it}-e_{t}^{s}-\underline{\pi}=0$
and
$f_{i}^{l}(Q_{i})\hat{u}_{it}-e_{t}^{\epsilon}-\overline{\pi}=0$
to obtain
$\underline{\pi}$
and
$\overline{\pi}$
.
. Step 1. Let
$\hat{x}i_{t}$
be the solution of the equation
$f_{i}(x_{il}^{s})\hat{u}_{it}-e_{t}^{\epsilon}-\underline{\underline{\pi}+\overline{\pi}}2=0$
If
$\hat{x}\mathrm{j}_{t}$
$<q:,$ then
$\hat{x}\mathit{7}_{t}$
$=q:.$
Else if
$\hat{x}\mathrm{j}_{t}>Q_{:}$
, then
$\hat{x}j_{t}$ $=Q_{*}$
..
. Step 2. If
$\sum_{=1}$

$\hat{x}$
7t
- $\alpha<0,$
then
$\overline{\pi}=\underline{\pi}+\pm_{2}^{-}\pi\pi$
. Otherwise, if
$\sum_{i=1}^{\mathit{1}}\hat{x}_{t}^{s}-\alpha>0,$
then
$\underline{\pi}=\underline{\pi}+\pm_{2}^{-}\pi\pi$
.
Go to step 1.
Figure 3: Modied Algorithm of Lambda Iteration
5 Numerical Experiments
This section demonstrates how our algorithm works. The column generation method for the stochastic unit
commitment problem was implemented using
$\mathrm{C}$
language on SUN Enterprise
$420\mathrm{R}$
(
$450\mathrm{M}\mathrm{H}\mathrm{z}$
UltraSPARCII).
The whole framework of the algorithm was coded in C.
We applied our solution method to the problems based on an electric power utility in Japan. Some of
the data was modied for the sake of condentiality. The test problems considered in this section consist of
$I=10,20,30$ units, $T=168$ periods and $S=16$ scenarios. We have predicted demand for electric power
$\overline{d}_{t}$
for
$t=1$ , $\ldots$ ,
$T$
. The demand data
$\overline{d}_{t}$
was also multiplied by 7/10. The scenarios are generated from the
predicted power demand by increasing or reducing predicted demand
$\overline{d}_{t}$
, as shown in Table 1.
We assume that hourly spot price
$e_{t}^{\epsilon}$
follows the next equation (17).
$e_{t}^{s}= \sum_{j=1}^{I}f^{J}$
$.(\begin{array}{l}\mathrm{i}q.+Q2\end{array})$
$/I$
$+ \frac{d_{t}^{\epsilon}-\overline{d}_{t}}{1000I}$
(17)
201
Table 1: Demand increases and decreases for dierent scenarios
Table 2 presents the results of our test problems. The rst column provides the number of generators.
For example,
$I=10\mathrm{x}3$
indicates that there are 10 sets of generators each of which contains 3 identical
generating units. The columns labeled Primal provide the objective values which are modied from the
solution of Lagrangian relaxation by the economic dispatch algorithm shown in Figure 2. The columns
labeled Dual provides the value of Lagrangian. The number of iteration of Lagrangian relaxation is set
to 1000, and the number of iteration of golden section and lambda iteration (Figure 3) is limited to 20 and
30, respectively.
We compare the conventional unit commitment model, in which we minimize the total cost, with our
model which we developed by incorporating power trading. In the conventional model, power trading is not
taken into consideration. The column labeled Gain shows the prot generated from considering power
trading.
Table 2: Computational Results
$I=$
Sets of Conventional Unit commitment Gain Gain
generators unit commitment with power trading
$\mathrm{x}$
identical Primal Dual Gap(%) Primal Dual Gap(%) (%)
units
$10=10\cross 1$
3486224 3398633 2.51 3402920 3366875 1.05 83304 2.39
$20=10\mathrm{x}2$
6973285 6802345 2.45 6803181 6752377 0.75 170104 2.44
$30=10\mathrm{x}$ $3$
10458953 10208589 2.39 10186560 10142858 0.43 272393 2.60
Then, we consider the value of the stochastic solution. Let
$\overline{x}$
i
$t$ ,
$\overline{u}_{\mathrm{i}}t$
be the solution of the unit commitment
problem which is obtained by replacing all random variables with their expected values. This solution
is referred to expected-value solution. We appliy this expected-value solution to all available scenarios
and the expected value of the objective function at
$(\overline{x},\overline{u})$
can be computed. The dierence between the
expected objective value of the expected solution and the optimal objective value of the original stochastic
programming problem is called the value of the stochastic solution, and is denoted as VSS. The value of the
stochastic solution is shown in Table 3.
Table 3: Value of stochastic solution
$I=$
Expected objective value Expected objective value VSS VSS
$\frac{\mathrm{u}\mathrm{s}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{e}\mathrm{d}- \mathrm{v}\mathrm{a}1\mathrm{u}\mathrm{e}\mathrm{s}\mathrm{o}1\mathrm{u}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{u}\mathrm{s}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{c}\mathrm{h}\mathrm{a}\mathrm{s}\mathrm{t}\mathrm{i}\mathrm{c}\mathrm{s}\mathrm{o}1\mathrm{u}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}(\%)}{10=10\mathrm{x}134422073402920392871.14\%}$
$20$ $=$
10x 2 6884416 6803181 81235 1.18%
$30=10\mathrm{x}3$
10326624 10186560 140064 1.36%
202
6 Concluding Remarks
We proposed a stochastic integer programming model in which the objective is to maximize expected prots.
This problem is formulated as a multi-stage stochastic quadratic integer programming problem because the
fuel cost function is assumed to be a convex quadratic function. The approach to solving the problem is
based on Lagrangian relaxation method and dynamic progra mming. The feasible schedule is obtained by
solving dynamic programming on a scenario tree. To rene the solution obtained by dyna nic program ming.
we solve an economic dispatch problem in which the equality demand constraint is relaxed to the inequality
constraint with upper and lower limit. To solve this problem we develop an algorithm which combines the
lambda iteration method and golden section. More research is necessary to make a scenario set which reects
real demand or spot price. As for application to real power systems, the coordination of the operation of
hydroelectric generation plants is left as a future problem.
References
[1] J. F. Bard. Short-term scheduling of thermal-electric generators using Lagrangian relaxations. Opera-
tions Research 36(1988),756-766.
[2] C. Barnhart, E. L. Johnson, G. L. Nemhauser, M. W. P. Savelsbergh, P. H. Vance. Branch-and-price:
column generation for solving huge integer programs. Operations Research 46(1998), 316-329.
[3] J. R. Birge. Stochastic programming computation and applications. INFORMS Journal on Computing
$9(1997)$ , 111-133.
[4] J. R. Birge, F. Louveaux. Introduction to Stochastic Programming, Springer-Verlag, 1997.
[5] G. B. Dantzig, P. Wolfe. Decomposition principle for linear programs. Operations Research
$8(1960)$ ,
101-111.
[6] J. K. Delson, S. M. Shahidehpour. Linear programming applications to power system economics, plan-
ning and operations. IEEE Transactions on Power Systems $7(1992)$ , 1155-1163.
[7] J. A. Muckstadt, S. A. Koenig. An application of Lagrangian relaxation to scheduling in power-
generation systems. Operations Research, 25(1977), 387-403.
[8] R. T. Rockafellar, R. J. -B. Wets. Scenarios and policy aggregation in optimization under uncertainty.
Mathematics of Operations Research, 16(1991), 119-147.
[9] G. B. Sheble, G. N. Fahd. Unit commitment literature synopsis. IEEE Transactions on Power Systems,
11(1994), 128-135.
[10] T. Shiina. Stochastic programming (in Japanese). In: Kubo
$\mathrm{M}$
, Tamura
$\mathrm{A}$
, Matsui
$\mathrm{T}(\mathrm{E}\mathrm{d}\mathrm{s}.)$
. Oyo S\^uri-
Keikaku Handbook. Asakura Syoten, 2002. p.710-769.
[11] T. Shiina. Stochastic programming model for unit commitment problem (in Japanese). Transactions of
the Japan Society for Industrial and Applied Mathematics, 13(2003), 181-190.
[12] T. Shiina, J. R. Birge. Stochastic unit commitment problem. International Transactions in Operational
Research, 11(2004), 1932.
[13] S. Takriti, J. R. Birge, E. Long. A stochastic model for the unit commitment problem. IEEE Transac-
tions on Power Systems 11(1996), 1497-1508.
[14] S. Takriti, J. R. Birge. Using integer programming to rene Lagrangian-based unit commitment solu-
tions. IEEE Transactions on Power Systems, 15(2000), 1501-1505.
[15] S. Takriti, J. R. Birge. Lagrangian solution techniques and bounds for loosely coupled mixed-integer
stochastic programs. Operations Research, 48(2000), 91-98.
[16] S. Takriti, B. Krasenbrink, L. S.-Y. Wu. Incorporating fuel constraints and electricity spot prices into
the stochastic unit commitment problem. Operations Research, 48(2000), 268-280.
[17] A. J. Wood, B. J. Wollenberg. Power Generation, Operation and Control second edition. John Wiley
& Sons, 1996.
[5] G. B. Dantzig, P. Wolfe. Decomposition principle for linear programs. Operations Research
$8(1960)$ ,
101-111.
[6] J. K. Delson, S. M. Shahidehpour. Linear programming applications to power system economics, plan-
ning and operations. IEEE
Transactions
on Power Systems $7(1992)$ , 1155-1163.
[7] J. A. Muckstadt, S. A. Koenig. An application of Lagrangian relaxation to scheduling in power-
generation systems. Operations Research, 25(1977), 387-403.
[8] R. T. Rocfellar, R. J.-B. Wets. Scenarios and policy
$\mathrm{a}\Re \mathrm{r}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}_{\dot{1}}\mathrm{o}\mathrm{n}$
in optimization under uncertainty.
Mathematics of Operations Research, 16(1991), 119-147.
[9] G. B. Sheble, G. N. Fahd. Unit commitment literature synopsis. IEEE Transactions on Power Systems,
11(1994), 128-135.
[10] T. Shiina. Stochastic programming (in Japanese). In: Kubo
$\mathrm{M}$
, Tamura
$\mathrm{A}$
, Matsui
$\mathrm{T}(\mathrm{E}\mathrm{d}\mathrm{s}.)$
. Oy\^o S\^uri-
Keikaku Handbook. Asakura Syoten, 2002. p.710-769.
[11] T. Shiina. Stochastic programming model for unit commitment problem (in Japanese). Transactions of
the Japan Society for Industrial and Applied Mathematics, 13(2003), 181-190.
[12] T. Shiina, J. R. Birge. Stochastic unit commitment problem. International Transactions in Operational
Research, 11(2004), 19-32.
[13] S. Takriti, J. R. Birge, E. Long. Astochastic model for the unit commitment problem. IEEE
$\mathrm{q}$
}
$\mathrm{a}\mathrm{n}\mathrm{s}\mathrm{a}\mathrm{c}rightarrow$
tions on Power Systems 11(1996), 1497-1508.
[14] S. Takriti, J. R. Birge. Using integer programming to rene Lagrangian-based unit commitment solu-
tions. IEEE Transactions on Power Systems, 15(2000), 1501-1505.
[15] S. Takriti, J. R. Birge. Lagrangian solution techniques and bounds for loosely coupled mixed-integer
stochastic programs. Operations Research, 48(2000), 91-98.
[16] S. Takriti, B.
$\mathrm{K}\mathrm{r}\mathrm{a}\mathrm{s}\mathrm{e}\mathrm{n}\mathrm{b}\mathrm{r}\mathrm{i}\mathrm{n}\mathrm{k}$
, L. S.-Y. Wu. Incorporating fuel constraints and electricity spot prices into
the stochastic unit commitment problem. Operations Research, 48(2000), 268-280.
[17] A. J. Wood, B. J. Wollenberg. Power Generation, Operation and Control second edition. John Wiley
&Sons, 1996.