You are on page 1of 8

1

Energy and Spinning Reserve Payment Cost Co-


optimization
Xu Han, Peter B. Luh, Fellow, IEEE, Joseph H. Yan, Senior Member, IEEE, and Gary A. Stern

between the auction and settlement mechanism, and the


Abstract—In the current U.S. electricity market, supply bids consumer payments can be significantly higher than the
are selected to minimize the bid costs, but are then paid at the minimized bid costs. An alternative auction mechanism which
marginal price, leading to significantly higher payments than the minimizes the total payments directly has been discussed.
minimized bid costs. This gives rise to the “payment cost
Recent results demonstrate the significant consumer savings
minimization” mechanism, which minimizes payments directly.
We previously presented the “objective switching” method to under this “payment cost minimization” (PCM) setup with the
solve single-product payment cost minimization problems. In same set of supply bids [1]-[4], however, many other aspects
view that spinning reserve is typically co-optimized with energy, of the mechanism is still not clear.
this paper further develops the “objective switching” method to A thorough comparison of the two mechanisms hinges on
solve the co-optimization of energy and spinning reserve. The the availability of their solution methodologies. The bid cost
prices are defined in a linear way to be the highest bid price
minimization problems can be treated similarly to traditional
among the “participating units,” and congestion status variables
are introduced to describe congestion-dependent LMP for energy unit commitment problems and efficiently solved by standard
product. Similar to single-product payment cost minimization, branch-and-cut based solvers. While the methods for solving
the co-optimization problem has an irregular convex hull, which bid cost minimization abound, only a few have been reported
is difficult to obtain by the branch & cut. To overcome the above for payment cost minimization. An augmented Lagrangian
difficulty, the objective switching method is further developed. and surrogate optimization framework is presented in [5], and
Salient features of the improved method include a new way to
is extended in [6] to include transmission capacity constraints.
add performance cuts to guarantee the convergence, and
heuristics developed in the infeasibility minimization stage. The objective switching method is presented in [7] to solve
Numerical testing results of small examples and the 24-bus IEEE PCM problems for energy product in primal space.
Reliability Test System demonstrate the effectiveness and Similar to the energy product, ancillary services (e.g.,
efficiency of the method. regulation, spinning reserve and non-spinning reserve) are also
selected by auctions [8][9]. The ancillary service bids can be
Index Terms— Branch-and-cut, co-optimization of energy and either optimized after the energy auction, or co-optimized
spinning reserve, CPLEX, MIP, objective switching method,
payment cost minimization.
simultaneously with energy bids. In view that energy and
ancillary services are coupled through the unit capacity
I. INTRODUCTION constraints, the co-optimization is more difficult to solve, but
can lead to a more efficient allocation than optimizing them

I N the current U.S. electricity markets (e.g., the day-ahead


markets), supply bids are selected to minimize the total bid
costs, but are then paid at the uniform market clearing prices
individually. Most Independent System Operators (ISOs) are
currently co-optimizing the energy and ancillary service bids
under bid cost minimization setup. Few results have been
(MCPs) or congestion-dependent locational marginal prices reported for payment cost co-optimization yet.
(LMPs). This “bid cost minimization mechanism” (BCM) This paper solves energy and spinning reserve payment
maximizes social welfare if supply bids reflect production cost co-optimization problems by using the branch-and-cut
cost. However, this assumption may not be true since bidders based standard solver. Other ancillary services (regulation and
may strategically deviate from their true costs. As a result, non-spinning reserve) are not considered for simplicity. DC
minimizing bid costs may not be maximizing social welfare. flows are used to formulate the transmitted power and
Meanwhile, bid cost minimization leads to inconsistency transmission capacity constraints are considered. Literatures
on single-product optimization and co-optimization under the
two mechanisms are reviewed in Section II. The payment cost
This work was supported in part by the grant from Southern California Edison
and the National Science Foundation under grants ECS-0621936 and ECCS-
co-optimization model is presented in Section III, with the
1028870. new price definition. Defining prices for payment cost
Xu Han and Peter B. Luh are with the Department of Electrical and minimization is difficult, since the incremental payment for
Computer Engineering, University of Connecticut, Storrs, CT 06269-2157,
USA
1MW demand increase does not always equal the bid price of
Joseph H. Yan is with Bidding Strategy & Asset Optimization, Southern the marginal unit, but can be infinite when a new unit is
California Edison, Rosemead, CA 91770 USA. committed. In stead of using Lagrange Relaxation, the prices
Gary A. Stern is with Market Strategy and Resource Planning, Southern are defined directly in the primal space to be the highest bid
California Edison, Rosemead, CA 91770 USA.
price among the “participating units.” To describe congestion-

978-1-4577-1002-5/11/$26.00 ©2011 IEEE


2

dependent energy prices, the congestion status variables are have been reported. An augmented Lagrangian and surrogate
introduced. Individual-product capacity constraints are added optimization framework is presented in [5] for solving
to describe the participating status of the unit. All the above payment cost minimization problems, and is extended in [6] to
constraints are formulated in linear forms to be solved by include transmission capacity constraints. Besides Lagrange
branch-and-cut based standard solver CPLEX MIP. Relaxation, the mixed integer programming techniques are
The efficiency for solving the linear co-optimization model also developed to improve the computational efficiency for
by the branch-and-cut method is still low. Our study on solving payment cost minimization problems. The objective
single-product payment cost minimization has shown the switching method is developed on top of CPLEX MIP and
irregular convex hull caused by constraints coupling market presented in [7] for solving single-product payment cost
minimization problems. The objective is switched to
prices and individual unit status variables [7]. The difficulty
minimizing the total infeasibilities of originally discrete
still exists in co-optimization problems, and is aggravated in
variables in the gradually reduced feasible regions and
view of the coupling of energy and spinning reserve through
quantifiable feasible solutions are obtained in a
unit capacity constraints, leading to a large number of computationally efficient manner. The payment cost co-
branching operations and the slow convergence of the optimization, however, is seldomly addressed.
algorithm. The “objective switching” method is further
developed to overcome the above difficulty. In the improved III. PROBLEM FORMULATION
method, performance cuts are added in a new way to
guarantee the convergence. Heuristics are used if the In this section, the co-optimization of energy and spinning
infeasibility cannot be minimized to 0 in given time to obtain a reserve under payment cost minimization is formulated for a
feasible solution. Details of the improved method are day-ahead market with given energy demand and spinning
presented in Section IV. Numerical testing results presented reserve requirement. DC power flow is used to model the
in Section V demonstrate effectiveness of the co-optimization transmitted power, and transmission capacity constraints are
model by using small examples, and demonstrate the considered. For simplicity, single-block bids are used.
efficiency of the method by using the IEEE Reliability Test Startup costs are assumed fully compensated. In the
System, with Monte Carlo simulations for 3 randomly selected following, the problem formulation is presented in subsection
load profiles and supply bids. III-A. The prices for energy and spinning reserve are defined
to be the highest bid price among the “participating” units, and
II. LITERATURE REVIEW the details are presented in subsection III-B. The allocation of
capacities is described in III-C. All the above constraints are
Many studies on the two auction mechanisms have been formulated in linear forms.
conducted. For the single-product bid cost minimization, the
Lagrange Relaxation and branch-and-cut techniques are often A. Problem formulation
used to solve the unit commitment problem in view of the Consider a transmission network with I buses indexed by i
mixed integer programming feature. The continuous = 1, 2,…, I, and L transmission lines indexed by l = 1, 2, …, L.
economic dispatch problem is often solved as a LP. The The transmission capacity for each line l is denoted by flmax.
branch-and-cut method represented by CPLEX MIP is now The transmitted power in line l at hour t (1 ≤ t ≤ 24) is denoted
widely used in ISOs for solving auction problems. In the by fl(t). The energy demand of bus i at hour t is given as
method, integrality constraints are first relaxed. Cuts are
PiDE(t). The spinning reserve requirement at hour t is given as
generated to obtain the convex hull (“convex hull” is used
PDS(t). For each bus i, there are Ki supply bids indexed by k =
henceforth for simplicity) of the original feasible solutions.
1, 2,…, Ki. For bid ik, the total generating level for energy
The simplex method then efficiently optimizes the relaxed LP
problem over the convex hull and generates an optimal and spinning reserve is limited within [pikmin, pikmax]. The
solution, which is also optimal to the original problem in view startup cost is denoted by Sik and is incurred if and only if the
of the problem linearity. Obtaining the convex hull itself is supply bid is turned “on” from “off” at hour t. The energy
NP hard, thus branching operations are needed to decompose price for bus i at hour t is denoted by LMPiE(t). The spinning
the problem as in the branch-and-bound method. reserve price for the system at hour t is denoted by MCPS(t).
The studies on multi-product bid cost optimization can be The cost to be minimized is the total MW payments for energy
divided into two categories: sequential optimization and and spinning reserve plus startup costs, subject to energy
simultaneous optimization. The sequential auction for demand and spinning reserve requirement constraints,
ancillary services that is used in the California ISO (CAISO) transmission capacity constraints, individual unit constraints,
is described in [10] and [11]. For simultaneous optimization, and constraints defining LMPiE and MCPS. Minimum
the co-optimization model for simultaneous auctions that is up/down times, and ramp rates are not considered for
used in the New York ISO (NYISO) and New England ISO simplicity.
(ISO-NE) are described in [12] and [13]. A detailed AC OPF- Objective function: The total payment cost is the sum of
based formulation for procuring, pricing, and settling energy MW payments for both energy and spinning reserve products
and ancillary service is presented in [14], with the economic and startup costs across the system over the 24 hour period,
dispatch problem solved as a LP. with the MW payment calculated as the product of prices and
While the methodologies for bid cost minimization abound, corresponding demands:
only a few results on single-product optimization problem
3

T I
DE
T
S by solving the discrete unit commitment problem. Levels of
J ≡ ∑ ∑ LMP i ( t ) ⋅ Pi ( t ) + ∑ MCP ( t ) ⋅ P DS ( t )
t =1 i =1 t =1 the selected bids are then determined by solving a continuous
T I Ki economical dispatch problem and the prices are obtained in an
+ ∑ ∑ ∑ uik (t ) ⋅ Sik . (1) ex post manner. The prices are defined to be ∂LED/∂PD, where
t =1i =1 k =1
LED denotes the Lagrangian of economic dispatch and PD
Energy demand constraints: The total energy generation denotes the demand level. The above definition, however, is
across the system should equal the total energy demand in
not good for defining prices in payment cost minimization. In
each hour: view that the total payments are minimized in PCM, if PD
I Ki I
DE
∑ ∑ pikE (t ) = ∑ Pi (t ), ∀t. (2) increases 1 MW, the incremental total payments does not
i =1 k =1 i =1 always equal the bid price of the marginal unit, but can be
Spinning reserve requirement constraints: The total infinite at some “corner points” when a new block or unit is
spinning reserve level across the system should equal the total committed.
spinning reserve requirement in each hour: In single-product economic dispatch, minimizing total
I Ki
DS payments and minimizing total bid costs with fixed unit
∑ ∑ pikS (t ) = P (t ), ∀t. (3) commitment can lead to the same dispatch results under
i =1 k =1
certain conditions, and the total payments can be substituted
Startup cost constraint: The turn on/off decision variable
by total bid costs to define prices by using the Lagrangian of
uik(t) in the objective function equals 1 if and only if the
economic dispatch [6][7]. However, such substitution cannot
supply bid is turned “on” from “off” at hour t:
be done for co-optimization problems, since minimizing total
uik (t ) ≥ xik (t ) − xik (t − 1), ∀i, ∀k , ∀t . (4)
payments and total bid costs lead to different dispatch results
Table 1 shows the above relationship. Only when xik(t-1) [1]. The fact triggers the new way to define prices under
equals 0 and xik(t) equals 1, uik(t) is fixed at 1, otherwise payment cost minimization setup.
equation (4) will be redundant. New price definition:
In this model, the LMPiE for energy at bus i is defined to be
TABLE I
RELATIONSHIP BETWEEN X AND U
the highest bid price among the units selected to “participate
xik(t) xik(t-1) uik(t) in the energy market” and serving bus i. A binary variable xikE
0 0 uik(t) ≥ 0 (redudent) is introduced to denote the “participating status” of unit ik.
0 1 uik(t) ≥ -1 (redudent) xikE = 1 if and only if unit ik is selected for the energy market.
1 0 uik(t) ≥ 1 (fixed at 1) It can be easily seen that only the most expensive selected unit
1 1 uik(t) ≥ 0 (redudent) at each bus can be a “price setting” unit:
LMPi E (t ) ≥ cikE ⋅ xikE (t ), ∀i, ∀k , ∀t. (8)
Generation capacity constraints: For unit ik, the total
generating level for energy and spinning reserve is limited by The price at bus i can be either set by a local unit at bus i, or
its minimum and maximum limits if it is committed. can be some other remote unit located at other buses. If a
Otherwise, the total generating level is 0: local unit ik at bus i is setting LMPiE, the above equation is
active, and LMPiE = cikE. If a unit jh located at some other bus
xik (t ) pik min ≤ pikE (t ) + pikS ≤ xik (t ) pik max , ∀i, ∀k , ∀t. (5)
j is setting the LMP for bus i, the above inequality still holds
DC power flow equations: The flow fl(t) in line l can be but will not be active.
expressed as the linear combination of net nodal injection of A remote unit jh can set the price for bus i only if the power
energy [15] at hour t: can be delivered from bus j to bus i without violating
I ⎡ ⎛ Ki ⎞⎤ transmission capacity constraints, i.e., unit jh can serve bus i.
f l (t ) = ∑ ⎢a i l ⋅ ⎜⎜ ∑ pikE − Pi DE (t ) ⎟⎟⎥, ∀l , ∀t. (6)
i =1 ⎣ ⎝ k =1 ⎠⎦ Because of congestion, it is sometimes difficult to identify the
units serving a bus. In stead of dealing with the whole system,
The coefficient ail known as power transfer distribution
this model describes the LMPiE differences between any two
factor (PTDF) denotes the sensitivity of the transmitted power
adjacent buses.
in line l with respect to the net injection at bus i, and is
Suppose that line l is connecting two adjacent buses i and j,
determined by the network structure, transmission line
and the positive direction is from bus i to bus j. Two
reactance, and reference bus selection.
congestion status variables xl+ and xl- are introduced for line l.
Transmission capacity constraints: The flow in line l cannot
xl+ = 1 if and only if there is congestion in line l along the
exceed the transmission capacity in any hour, i.e.,
positive direction, i.e., fl = flmax. xl- has a similar meaning. The
− f l max ≤ f l (t ) ≤ f l max , ∀l , ∀t. (7) relationship between congestion status variable (xl+ and xl-)
The transmission capacity limits for both directions are set and the flow fl can be described as:
to be the same for simplicity. 1 − xl+ (t ) ≤ [ f l max − f l (t )] ⋅ M , ∀l , ∀t. (9)
B. LMPiE and MCPS definition f l max − f l (t ) ≤ [1 − xl+ (t )] ⋅ M , ∀l , ∀t. (10)
Dilemma for defining prices in PCM co-optimization:
In bid cost minimization, ISOs typically solve the auction 1 − xl− (t ) ≤ [ f l max + f l (t )] ⋅ M , ∀l , ∀t. (11)
problem by using a two-step procedure. Bids are first selected f l max + f l (t ) ≤ [1 − xl− (t )] ⋅ M , ∀l , ∀t. (12)
4

where M is a large positive number. Equation (19) shows that if unit ik is selected for the energy
It can be seen from the equation (9) that if fl = flmax, then market, then pikE > 0, and xikE equals 1. If unit ik is not selected
there is congestion in line l along the positive direction, and for providing energy, then pikE will be 0 and xikE will be fixed
the right hand side is 0. As a result, the congestion status at 0. Equation (20) can be explained in a similar way.
variable xl+ will be fixed at 1. Similarly, in equation (10), if The optimization problem is minimizing (1) subject to (2)-
xl+ is 1, then the right hand side will be 0 and fl will be fixed at (20). All the constraints are formulated in linear forms, and a
flmax. So xl+ = 1 is a necessary and sufficient condition for the linear co-optimization model is obtained.
congestion in line l along the positive direction and xl+ can be
used to denote the congestion status. Equation (11) and (12) IV. SOLUTION METHODOLOGY
can be interpreted in a similar way.
The above congestion status variables are used to describe Even though formulated linearly, the co-optimization under
the LMP difference between adjacent buses. When there is no payment cost minimization is still much more difficult to solve
congestion along the positive direction and xl+ = 0, any cheap in CPLEX MIP than bid cost co-optimization. The reason is
units serving bus i can also serve bus j, and LMPjE cannot be explained in IV-A by the convex hull analysis. The distinct
higher than LMPiE. When there is congestion along the structure of payment cost minimization leads to an irregular
positive direction and xl+ = 1, the cheap power at bus i cannot convex hull which is difficult to obtain, leading to a large
number of branching operations and the low efficiency. This
be delivered to bus j, and LMPjE can be higher than LMPiE.
difficulty is aggravated in the co-optimization in view of the
The above relationship can be expressed by equation (13):
coupling of energy and spinning reserve through unit capacity
LMPj (t ) − LMPi (t ) ≤ xl+ (t ) ⋅ M , ∀i, ∀j, ∀t. (13) constraints. The “objective switching method” is further
where M is a large positive number. developed to overcome the above difficulties. The objective is
Similarly, equation (14) can be formulated for the negative switched to minimizing total infeasibilities of originally
direction: discrete decision variables in gradually reduced feasible
regions bounded by performance cuts, and quantifiable
LMPi (t ) − LMPj (t ) ≤ xl− (t ) ⋅ M , ∀i, ∀j, ∀t. (14) feasible solutions are obtained. To guarantee the convergence
To define the price MCPS for the spinning reserve, another of the algorithm, performance cuts are added in an improved
participating status variable xikS is introduced, with the way. Heuristics are developed to obtain feasible solutions in
meaning similar to xikE. The uniform price MCPS for the the infeasibility minimization stage. Details of the improved
spinning reserve product is defined as: method are presented in subsection IV-B.
MCP S (t ) ≥ cikS ⋅ xikS (t ), ∀i, ∀k , ∀t. (15) A. The convex hull analysis for payment cost co-optimization
The following non-transmission 1-hour 2-bid co-
C. Allocation of the generating capacities optimization problem is used to analyze the convex hull of
In payment cost minimization, the prices are defined before payment cost co-optimization problems. The bid parameters
the auction rather obtained in an ex post manner as in bid cost are given in table II. The energy demand is 40MW, and the
minimization. Whether a unit can affect the price of one spinning reserve requirement is 20MW.
product depends not only on its on/off commitment status xik,
TABLE II
but also on whether it is selected to participate the market or PARAMETERS OF TWO BIDS
not, i.e., the participating status xikE and xikS. The coupling pmin pmax Startup cost E bid price SR bid price
relationship between commitment status variables and (MW) (MW) ($) ($/MW) ($/MW)
participating status variables can be described as: Unit 1 10 80 0 65 65
Unit 2 10 70 0 30 30
xikE (t ) ≤ xik (t ), ∀i, ∀k , ∀t , (16)
xikS (t ) ≤ xik (t ), ∀i, ∀k , ∀t , (17) The total payments for the two products are minimized.
LMPiE for energy are substituted by MCPE for this non-
xikE (t ) + xikS (t ) ≥ xik (t ), ∀i, ∀k , ∀t. (18) transmission problem. The problem is:
Equation (16) and (17) shows that if unit ik is not committed min: 40MCPE+20MCPS, subject to
and xik(t) = 0, then the unit cannot be selected for any market, Energy demand: p1E + p2E = 40,
i.e., a unit can participate in a market only when it is online. Spinning reserve requirement: p1S + p2S = 20,
Equation (18) shows that once unit ik is committed online, it Generation capacity: 10 x1 ≤ p1E + p1S ≤ 80 x1,
must at least participate in one market, i.e., either selected to 10 x2 ≤ p2E + p2S ≤ 70 x2,
provide energy or selected to provide spinning reserve. Product capacity: δ⋅x1E ≤ p1E ≤ 80x1E,
Since the price for each product is only related to the units δ⋅x1S ≤ p1S ≤ 80x1S,
participating in the market, two individual-product capacity δ⋅x2E ≤ p2E ≤ 70x2E,
constraints are introduced: δ⋅x2S ≤ p2S ≤ 70x2S,
xikE (t ) ⋅ δ ≤ p ikE (t ) ≤ xikE (t ) p ik max , ∀i, ∀k , ∀t, (19) MCP constraints: MCPE ≥ 65x1E,
MCPE ≥ 30x2E,
xikS (t ) ⋅ δ ≤ pikS (t ) ≤ xikS (t ) pik max , ∀i, ∀k , ∀t, (20) MCPS ≥ 65x1S,
where δ is a very small positive number. MCPS ≥ 30x2S,
5

Coupling constraints: x1E + x1S ≥ x1, x1E ≤ x1, x1S ≤ x1, which is not fractional in each dimension, and the branching
x2E + x2S ≥ x2, x2E ≤ x2, x2S ≤ x2. operation can be reduced. Heuristics are developed when the
For the purpose of visualization, x1E, x2E, MCPE are plotted infeasibility cannot be minimized to 0 in a given time limit to
for the “energy part” convex hull of the co-optimization obtain a feasible solution.
problem as shown in Fig. 1. The method can be divided into two stages. In stage 1, the
simplex method is used as in the traditional branch-and-cut
MCPE F Cut method to minimize the total payments for the continuous
G
I relaxation of the original co-optimization problem. Feasible
H E points can be quickly obtained by using the relaxation induced
Gradient neighborhood search (RINS) method embedded in CPLEX
Continuous
Convex D relaxation MIP, and a performance cut can be added according to the
hull best feasible solution. To prevent ending up with the same
feasible point in the infeasibility minimization in stage 2, the
Feasible Fractional cut is shifted by a positive number ε to exclude the feasible
optimum A C optimum
point obtained thus far out of the reduced region. The value of
ε can be changed with the convergence of the algorithm.
B In stage 2, the infeasibility of a participating status variable
0 1
MCPE
constraints
J x 1E is defined as the difference between its value and the nearest
integer as shown in Fig. 2.
1
x2E
The piece-wise linear functions in Fig. 2 can be converted
to linear forms by using the special ordered set (SOS)
Fig. 1. The “energy part” convex hull of co-optimization technique [16]. The infeasibility minimization problem is
formed by minimizing the total infeasibility of xikE and xikS. In
In the figure, x1Eand x2E can take any value within [0, 1] view of the coupling constraints, once xikE and xikS are
after relaxing the integrality constraints. In view that at least obtained, calculating xik is trivial. As a result, the infeasibility
one bid must be selected to satisfy the energy demand, a cut of commitment status xik is not include in the infeasibility
can be added to tighten the continuous relaxation:
minimization problem. The infeasibility minimization
x1E + x2E ≥ 1. (21)
problem is:
In view that MCPE is related to every xE, the MCP
constraints intersect each other at different angles. Together
with the cut (21), the feasible region ABCDEFGH is formed,
T I Ki
[
min J , with J ≡ ∑ ∑ ∑ inf ikE (t ) + inf ikS (t ) ,
t =1 i =1 k =1
] (22)

with a fractional optimum B. The convex hull of this problem subject to the constraints in the original problem and SOS
ADEFGH when plotted by these three dimensions cannot be constraints for linear conversion. It can be seen from Fig.2
obtained by further adding cuts. Branching will be performed that the infeasibilities for participating variables are defined to
for each dimension of the fractional continuous optimum B, be non-negative, thus when the total infeasibilities are
and the total branching operations grow exponentially when minimized to 0, a feasible solution with quantifiable quality is
the problem size increases, leading to the low efficiency. obtained. Because of the way for adding performance cuts,
When using x1S, x2S, MCPS to plot the “spinning reserve” part the new feasible solution will be better than the one used for
of the problem, a similar difficulty exists.
adding performance cut in the previous stage. This new
The two products are coupled through the generation
feasible solution will in turn be used to add performance cut
capacity constraints, and the coupling constraints on xik, xikE,
following the method presented in stage 1 and the feasible
xikS. Because of the high dimensionality, plotting a three
dimensional figure to show the difficulties caused by these region can be further reduced. The switching of objective
couplings is difficult. However, it is clear that the branching functions can be repeated iteratively and the solution quality
operations needed for the co-optimization problem are at least can be improved.
doubled comparing with a single-product optimization
problem with the same number of units. When the problem InfikE InfikS
size is large, the above difficulties caused by the irregular 1 1
convex hull can lead to a “curse of dimensionality” for co-
optimization problems.
B. The objective switching method
The objective switching method is further developed to 0 0.5 1 xikE 0 0.5 1 xikS
overcome the above difficulties. The performance cuts are (a) (b)
added in an improved way to guarantee that a better feasible Fig. 2. Definition of the infeasibility for xikE (a) and xikS (b)
solution can be obtained in the following infeasibility
Sometimes, minimizing the total infeasibility costs to zero
minimization. After adding performance cuts, the objective is
switched to minimizing total infeasibility of originally discrete can be difficult in view that the infeasibility minimization
decision variables xikE and xikS. The in feasibility problem is still a MIP. In such cases, a heuristic will be used
minimization problem has a different continuous optimum to round the “slightly” infeasible solution to the nearest integer
6

point. However, whether this integer point belongs to the at bus 1 and 40MW at bus 2. The total spinning reserve
original feasible set needs to be checked. If all the constraints requirement is 45MW. Bus 2 is selected as the reference bus.
are satisfied, the integer point is a new feasible solution and The positive direction for line l is from bus 1 to bus 2, and the
will be used to add new performance cut. If it is not a feasible line capacity is to be defined in each of the following cases.
solution of the original problem, then no feasible solution can Without considering transmission capacity constraints, unit 11
be obtained in the reduced feasible region in the given time and unit 21 will be selected since they are cheap with zero
limit. In such cases, the best feasible solution obtained will be startup cost. The problem is how to select one unit from unit
selected as the near-optimal solution. Other stopping criteria 12 and unit 22 to meet the total energy demand and spinning
can also be used in practice, e.g., the gap between the best reserve requirement, and how to allocate the generating
capacities for the committed units to minimize the total
feasible solution obtained and the continuous optimum is
payments of energy and spinning reserve. The non-congestion
smaller than a given threshold or a given CPU time is reached.
case 1 and congestion case 2 will be discussed.
The flow chart of the objective switching method is shown
in Figure 3. TABLE III
PARAMETERS OF THE FOUR UNITS
Begin pmin pmax Startup Energy Price SR Price
(MW) (MW) cost ($) ($/MW) ($/MW)
Solve the continuous relaxation by Unit 11 0 45 0 10 5
the simplex method Unit 12 0 40 200 30 15
Unit 21 0 45 0 20 10
Find feasible points by heuristics Stage 1,
Unit 22 0 12 20 100 50
minimizing
Add performance cut payment
costs
Switch the objective to minimizing Unit 11 Unit 21
infeasibility costs fl

Solve the new MIP by the branch- X=0.01


and-cut method Unit 12 Bus 1 Bus 2 Unit 22
Use the
Infeasibility = Yes integer point
0? as a new
Stage 2, Fig. 4. The network structure for Example 1
feasible
minimizing
point
No infeasibility Case 1: flmax = 30MW. The problem is solved by the objective
Round the solution to the nearest costs
integer point switching method and the testing result is listed in the first part
of table IV. Unit 11 and unit 21 are cheap bids with 0 startup
cost and they will be first selected at 45MW. In view that the
Satisfy all Yes
constraints?
total energy demand (20 + 40 = 60MW) in this example is
larger than the total spinning reserve requirement 45MW and
No the energy bids are more expensive, the cheap capacity
Select the best feasible point
Stop provided by unit 11 and unit 21 will be first allocated to the
obtained as the final solution
Fig. 3. Flow chart of the objective switching method energy market to achieve a lower energy price and total
energy payments. Unit 21 will set the energy price LMP1 =
LMP2 = 20$/MW. The remaining 30MW capacity from unit
V. NUMERICAL RESULTS 21 will be allocated to the spinning reserve market. To meet
the remaining 15MW spinning reserve requirement, unit 12
The improved objective switching method has been run on and unit 22 are compared. If unit 12 is selected, then the MCP
an Intel Xeon dual 1.6-GHz server with 8G memory. Two for spinning reserve will be 15$/MW, and the total payment
different size co-optimization examples are tested. Example 1 will be 20 × (20 + 40) + 15 × 45 + 200 = 2075. If unit 22 is
uses a 1-hour example to show the impact of transmission selected, then the MCP for spinning reserve will be 50$/MW,
congestion on the price of each product, and compares the and the total payment will be 20 × (20 + 40) + 50 × 45 + 20 =
optimal results obtained by the objective switching method 3470. As a result, unit 12 will be committed and the total
with the ones obtained by exhaustive search to verify the payment is $2075.
correctness of the model and the effectiveness of the price Case 2: flmax is reduced to 20MW to create this congestion case.
definition. Example 2 then tests the algorithm on the modified The problem is solved by the objective switching method and
IEEE 24-bus Reliability Test System to compare the the testing result is listed in the second part of table IV.
performance of the objective switching method with branch- Because of the congestion, only 20MW can be delivered from
and-cut method in terms of consumer payment and CPU time. bus 1 to bus 2 in this case. Thus unit 11 reduces its generation
Example 1: Consider a 4-bid 1-hour co-optimization problem. level to 40, and unit 21 increases the generation level to 20.
The network structure is shown in Fig. 4. The parameters of Since unit 11 and unit 21 are cheap, their remaining capacities
generators are given in table III. Each supplier submits energy (5 + 25 = 30MW) are selected for the spinning reserve market.
bid and spinning reserve bid. The energy demands are 20MW To meet the remaining 15MW spinning reserve requirement,
7

unit 12 and 22 are compared. If unit 12 is selected, then the Those cheap units can thus serve local areas more efficiently
MCP for spinning reserve will be 15$/MW, and the total than in the non-congestion case, leading to lower LMPiE for
payment will be 10 × 20 + 20 × 40 + 15 × 45 + 200 = 1875. If some bus and reduced total payments.
unit 22 is selected, then the MCP for spinning reserve will be
50$/MW, and the total payment will be 10 × 20 + 20 × 40 + TABLE V
RESULTS OBTAINED BY THE TWO METHODS
50 × 45 + 20 = 3270. So unit 12 will be committed and the
total payment will be $1875. In both of the above two cases, Non-congestion case Congestion case
the results obtained by the objective switching method are Branch- Objective Branch- Objective
consistent with the ones obtained by exhaustive search. and-cut switching and-cut switching
Consumer 1224032.35 1139482.49 NA 1103928.96
payments ($)
TABLE IV
TESTING RESULTS FOR NON-CONGESTED AND CONGESTED CASES CPU time (s) 2000s 369s 2000s 584s
Non-congested case Congested case
pikE pikS pikE pikS VI. CONCLUSION
p11 (MW) 45 0 40 5
p12 (MW) 0 15 0 15 The co-optimization of energy and spinning reserve under
p21 (MW) 15 30 20 25 payment cost minimization is discussed in this paper. The DC
P22 (MW) 0 0 0 0 flow is used to formulate the transmitted power and
LMP1E ($) 20 10 transmission capacity constraints are considered. The price
LMP2E ($) 20 20 for each product is defined in a linear way to be the highest
MCPS($) 15 15 bid price among the “participating units.” To formulate the
fl(MW) 25 20
congestion-dependent energy prices, the congestion status
Payment ($) 2075 1875
variables are introduced. The efficiency for solving co-
optimization problems by branch-and-cut in CPLEX MIP is
Example 2: Consider a 24-hour co-optimization problem with low because the co-optimization faces the same difficulty
24 buses connected by 34 transmission lines based on the caused by the irregular convex hull as single-product
IEEE Reliability Test System of 1996 (RTS-96). The original optimization does, and the situation is aggravated in view of
RTS-96 system is modified here for testing co-optimization the coupling of the two products. The objective switching
problems. The spinning reserve requirement is set to be 5% method is further developed to overcome the above difficulty.
of the total energy demands. The network information, such This method provided one way to improve the performance of
as reactance and capacities of transmission lines, remains the branch-and-cut in standard solver CPLEX MIP. It can be used
same as the original system. The whole system is divided into not only for payment cost co-optimization, but also for solving
two zones as described in [6], with 46.74% of system load but other difficult mixed integer programming problems.
only 20.09% of total generation capacity located in zone I.
The flows are therefore from zone II to zone I. The capacities VII. REFERENCES
of the five lines connecting zone I and zone II are assumed to [1] G. A. Stern, J. H. Yan, P. B. Luh, and W. Blankson, “What Objective
be enough for the non-congestion case, and are reduced to Function Should Be Used for Simultaneous Optimal Auctions in the
ISO/RTO Electricity Market,” Power Engineering Society General
100MW in the congestion case. Other lines are assumed to Meeting, IEEE, 2006.
have enough transmission capacities. The supply bid [2] J. H. Yan, “Bid Cost Minimization versus Payment Cost Minimization
parameters are randomly generated with Gaussian distribution. in the ISO/RTO Markets” Power Engineering Society General Meeting,
IEEE, 2007.
Three scenarios are tested, including the yearly peak-load day. [3] J. H. Yan, G. A. Stern, P. B. Luh and F. Zhao, “Payment Cost
Through out the testing, the default setting of the branch- Minimization vs. Bid Cost Minimization for ISO Markets,” IEEE Power
and-cut method in CPLEX MIP is used. The time limit for the & Energy Magazine, Vol. 6, No. 2, 2008, pp. 24-36.
infeasibility minimization stage is restricted to less than 180s. [4] F. Zhao, P. B. Luh, Y. Zhao, J. H. Yan, G. A. Stern and S. C. Chang,
“Bid Cost Minimization vs. Payment Cost Minimization: A Game
The CPU time limit for the whole process is 2000s. The co- Theoretic Study of Electricity Markets,” Proceedings of the IEEE Power
optimization problem is solved in CPLEX MIP by the branch- Engineering Society 2007 General Meeting, Tampa, Florida, June 2007.
and-cut method and the improved objective switching method, [5] P. B. Luh, W. E. Blankson, Y. Chen, J. H. Yan, G. A. Stern, S. C. Chang
and F. Zhao, "Payment Cost Minimization Auction for the Deregulated
and the results are given in table V. It can be seen that for the Electricity Markets Using Surrogate Optimization," IEEE Transactions
non-congestion case, the best feasible solution can be obtained on Power Systems, Vol. 21, No. 2, 2006, pp. 568-578.
within 2000s by the branch-and-cut method is $1224032. The [6] F. Zhao, P. B. Luh, J. H. Yan, G. A. Stern and S. C. Chang, “Payment
Minimization Auction for Deregulated Electricity Markets with
objective switching method can obtain a better solution Transmission Capacity Constraints,” IEEE Transactions on Power
$1139482 within a much shorter CPU time (369s) than the Systems, Vol. 23, No. 2, May 2008, pp. 532-544.
branch-and-cut method. For the congestion case, the branch- [7] X. Han, P. B. Luh, J. H. Yan and G. A. Stern, “Payment cost
minimization with transmission capacity constraints and losses using the
and-cut method cannot find any feasible solution within the objective switching method,” Proceedings of the IEEE Power
2000s CPU time limit. The objective switching method can Engineering Society 2010 General Meeting, Minneapolis, Minnesota,
obtain a feasible solution $1103928 in 584s. The consumer July 2010.
[8] ISO New England, “ASM Regulation Market Settlement,”
payment in the congestion case is a little lower than that of the http://www.iso-
non-congestion case. The reason is that some cheap power ne.com/support/training/courses/asmph1/05_asm_reg_mkt_settlement.p
cannot be delivered to serve other buses in the congestion case. df, 2005.
8

[9] ISO New York, “10-Minute Reserve Market,”


http://www.nyiso.com/public/webdocs/committees/bic/meeting_material
s/2000-05-18/5_10_min_res_market.pdf, 2000.
[10] H. Singh and A. D. Papalexopoulos, “Competitive Procurement of
Ancillary Services by an Independent System Operator,” IEEE Trans.
Power Systems, vol. 14, No. 2, pp. 498-504, May. 1999.
[11] Y. Liu, Z. Alaywan, M. Rothleder, S. Liu and M. Assadian, “A Rational
Buyer’s Algorithm Used for Ancillary Service Procurement,” presented
at 2000 IEEE/PES Winter Meeting, Singapore, Jan. 23-27, 2000
[12] A. I. Cohen, V. Brandwajn and S. K. Chang, “Security Constrained Unit
Commitment for Open Markets,” Proc. 21st International Conference on
Power Industry Computer Applications, pp. 39-44.
[13] K. W. Cheung, P. Shamsollahi, and D. Sun, “Energy and Ancillary
Service Dispatch for the Interim ISO New England Electricity Market,”
Proc. 21st International Conference on Power Industry Computer
Applications, pp. 47-53.
[14] T. Wu, M.Rothleder, Z. Alaywan and A. D. Papalexopoulos, “Pricing
Energy and Ancillary Services in Integrated Market Systems by an
Optimal Power Flow,” IEEE Transactions on Power Systems, Vol. 19,
No. 1, 2004, pp. 339-347.
[15] A. J. Wood, and B. F. Wollenberg, Power generation operation &
control, 2nd edition. New York: John Wiley & Sons, Inc., 2003.
[16] IBM ILOG, “ILOG CPLEX 11.0 user’s
manual,”http://www.lingnan.net/lab/uploadfile/200864184419679.pdf,
2007.

VIII. BIOGRAPHIES
Xu Han received the B.S. degree in automation the M.S. degree in control
theory and control engineering from the University of Science and
Technology Beijing, China, in 2005 and 2008. He is currently pursuing the
Ph.D. degree in electrical and computer engineering at the University of
Connecticut, Storrs. His research interests include optimization, operations,
and economics of electricity markets.

Peter B. Luh (M’80–SM’91–F’95) received his Ph.D. degree in Applied


Mathematics from Harvard University in 1980, and has been with the
University of Connecticut since then. Currently he is the SNET Professor of
Communications & Information Technologies. His interests include design of
auction methods for electricity markets; electricity load and price forecasting
with demand management; control and optimization of boilers and energy
systems; optimized resource management and coordination for sustainable,
green, and safe buildings; planning, scheduling, and coordination of design,
manufacturing, and service activities; decision-making under uncertain,
distributed, or antagonistic environments; and mathematical optimization for
large-scale problems. He is a Fellow of IEEE, Vice President of Publication
Activities for the IEEE Robotics and Automation Society, the founding
Editor-in-Chief of the new IEEE Transactions on Automation Science and
Engineering (2003-2008), and was the Editor-in-Chief of IEEE Transactions
on Robotics and Automation (1999-2003).

Joseph H. Yan (M’02–SM’06) received the Ph.D. degree in electrical


engineering from the University of Connecticut, Storrs, in 1996. He is the
manager in the Department of Bidding Strategy and Asset Optimization of
Southern California Edison, Rosemead, CA. In the past ten years, he worked
in areas of the electricity system operations, wholesale energy market analysis
for both regulated and non-regulated affiliates, market monitoring and market
design in California, and California Electricity Refund case of the 2000–2001
crisis. His research interest includes operation research, optimization, unit
commitment/scheduling and transaction evaluation, and optimal simultaneous
auction in deregulated ISO/RTO markets.

Gary A. Stern received the Ph.D. in economics from the University of


California at San Diego in 1984. He is the Director of Market Strategy and
Resource Planning for Southern California Edison, Rosemead, CA. He reports
to the Senior Vice President of the Power Production Business Unit and
manages a division responsible for resource planning, capacity and energy
market design, and monitoring the wholesale electricity market in California.
He is currently leading a broad internal team in developing strategies to
implement California’s recent Greenhouse Gas legislation. He is also
managing the design and the development of a capacity market for California,
partnering with various other stakeholders.

You might also like