[IEEE 2014 11th International Conference on the European Energy Market (EEM) - Krakow, Poland...
Transcript of [IEEE 2014 11th International Conference on the European Energy Market (EEM) - Krakow, Poland...
Value of Price Dependent Bidding for ThermalPower Producers
Erik B. Rudlang, Carl Fredrik Tjeransen,Stein-Erik Fleten
Department of Industrial Economics and
Technology Management
Norwegian University of Sciences and Technology
Trondheim, Norway
Email: [email protected], [email protected],
Gro KlæboeDepartment of Electric Power Engineering
Norwegian University of
Sciences and Technology (NTNU)
Trondheim, Norway
Email: [email protected]
Abstract—The steady increase in the short-term trading ofelectricity through power exchanges has made the investigationof appropriate bidding strategies relevant. The goal of this paperis to quantify the difference between price-dependent and price-independent bidding strategies for thermal power producers. Thetwo bidding strategies are evaluated for trading in the German-Austrian day ahead market. Optimal bidding decisions are foundthrough two stochastic optimization models. The results indicatethat the price-dependent bids outperform the price-independentbids by utilizing the flexibility in the generators to increaseproduction in the case of price peaks and reduce productionfor price drops.
Index Terms—Bidding behavior, Day ahead markets, Electric-ity supply industry, Power markets, Stochastic programming
I. INTRODUCTION
Power producers participating in an electricity spot-market
may develop different types of bidding strategies. Two of these
are price-independent and price-dependent strategies. Price-
independent bids consist of one volume for each hour and
will be accepted regardless of the price in the market. Price-
dependent bids consist of a set of price-volume pairs which
make up a bidding curve for each hour. Price-independent bids
are promoted as a way to reduce the risk of unpredicted market
events that might leave the power producer unable to fulfill
the market obligation [1]. However, these bids are inflexible
and incapable of responding to price signals from the market.
Price-dependent bidding, on the other hand, has the potential
to better react to unforeseen market events, but might leave
the power producer unable to fulfill the market commitment
through self-scheduling. Determining optimal bidding deci-
sions is complex and depends on the success in both market
modeling, production planning and bid generation. For thermal
power producers who are restricted by temporal constraints in
the generation process, bidding is even more complex.
The Price-Based Unit Commitment Problem (PBUC) is
used to find the optimal unit commitment given a price signal,
and it is an important component in models that find bidding
strategies for power producers. In the papers [1]–[3] bids are
developed using the solution of the PBUC. In [2] bidding
curves are developed by relating the optimal quantity decisions
from the PBUC to the statistical properties of the price
estimator. In [1] the PBUC is solved a number of times for
different offsets in price to obtain bidding curves for a range of
prices. The PBUC is extended to include a multistage scenario
tree in [3] and the solution to the deterministic equivalent
problem is used as bidding curves.
The PBUC can take multiple forms depending on the market
structure that the power producer faces. In [4]–[6] the power
producer is assumed to be a price-taker and has no influence
over the market clearing price. An appropriate forecasting tool
is used to estimate the hourly electricity prices for the next
day, and the PBUC is solved based on the price expectation.
The PBUC has also been used where the power producer
holds market power. One example is [7] which presents
a formulation where the power producer can influence the
market clearing price through price quota curves.
The operation of an electricity market requires power pro-
ducers to submit bids before actual production to settle market
clearing prices and quantities. A more detailed description
of the role of power exchanges can be found in [8]. This
procedure can be modeled as a two-stage problem where
optimal bids are determined in the first stage on the basis of
unit commitment in the second stage. The PBUC can thus be
extended to a stochastic model where the first stage decisions
are made under uncertainty [6].
In [9] the bidding problem is formulated as a two-stage
problem where the spot market is explicitly modelled. The
power producer’s bids are decided in the first stage based on
the realized spot prices found by aggregating own and realized
bids of other power producers in the second stage. This
approach requires knowledge of competitors bidding functions
and can only be used where this information is available.
A two-stage stochastic formulation that assumes the power
producer to be a price-taker is found in [10], [11]. Second
stage production decisions are related to first stage bidding
decisions through a coupling constraint. By fixing a set of
price points, [11] is able to obtain a linear problem that can
be solved with standard MIP solvers.
This paper analyzes the differences between price-
dependent and price-independent bidding. Two stochastic opti-
mization models are presented in Section IV where the power
producer is assumed a price taker. These models can be used
to generate price-dependent and price-independent bids based
978-1-4799-6095-8/14/$31.00 ©2014 IEEE
on stochastic spot market prices. In Section V the numerical
results are presented. A discussion on the findings is found in
Section VI followed by a conclusion in Section VII.
II. MEASURING THE VALUE OF PRICE-DEPENDENT
BIDDING
This paper proposes a new measure to aid the study of this
problem. The Value of Price-Dependent Bidding (VPDB) is
proposed to reflect the value of including the flexibility of
price-dependent bidding into the bidding model. The VPDB
is determined in a two-stage process. First the set of opti-
mal bidding decisions from the price-dependent model x∗PD
and the price-independent model x∗PI is found based on a
stochastic programming approach. These bidding decisions are
then fixed and the two optimization models are resolved based
on out-of-sample scenario data. The VPDB is the difference
between the optimal objective function value of the price-
dependent bidding problem and the optimal objective value
of the price-independent bidding problem (1) when fixing the
bidding decisions. To be able to evaluate the value of price-
dependent bidding between different scenarios, a relative value
of price-dependent bidding (RVPDB) is defined as the VPDB
divided by the optimal objective value of price-independent
bidding problem (2).
V PDB =V aluePD − V aluePI (1)
RV PDB =V PDB
V aluePI(2)
III. NOMENCLAUTURE
Indexess index of price scenarios
t index of time steps
i index of price points
j index of thermal units
l index of unit start up types
SetsS set of scenarios
T set of time steps for the planning period
TB set of time steps for the bidding period, TB ⊆ TT− planning horizon extended to the past
I set of price points
J set of thermal units
L set of unit temperature states L = {h,w, c}where h=hot, w=warm, c=cold
Parametersρst realized price at time step t in scenario spi bidding price at price point iπs probability of scenario scj marginal cost of unit jCl
j start up cost of start up type l in unit jCj commitment cost of unit j per hour
RLj ramping limit in unit j
UTj minimum up time in unit jDTj minimum down time in unit jPj maximum output of unit jPj minimum output of unit j
T lj number of time steps of start up
type l in unit jTmax number of time steps in the planning
horizon TR penalty for using the balance market
V aluePD value given price-dependent bidding
V aluePI value given price-independent bidding
Variablesxit volume bid at price point i in time step tyst volume committed at time step t in scenario szsjt volume produced in unit j at time step t
in scenario s+qst volume sold in the balancing market at
time step t in scenario s−qst volume bought in the balancing market at
time step t in scenario susjt binary variable that is equal to 1 if unit j
runs at time step t in scenario svsjt binary variable that is equal to 1 if unit j
shuts down at time step t in scenario sws
jt binary variable that is equal to 1 if unit jis started in time step t in scenario s
wlsjt binary variable that is equal to 1 if a type l
start up is initiated in time step t in unit jCprod aggregated variable costs
Cbal cost of using the balance market
IV. STOCHASTIC OPTIMIZATION MODEL
A. Price-dependent Bidding
The objective function (3) maximizes the power producers
profit, where profits equal the committed volume of electric-
ity at the realized electricity price minus the corresponding
operating costs and the cost for use of the balancing market.
max =
S∑s=1
πs
{T∑
t=1
ρstyst − Cprod − Cbal
}(3)
Operating Cost: The operating costs (4) consist of the
marginal production cost, the commitment cost, and the start-
up cost for the different start-up types.
Cprod =
T∑t=1
J∑j=1
cjzsjt +
T∑t=1
J∑j=1
Cjusjt +
L∑l=1
T∑t=1
J∑j=1
Cljw
lsjt
(4)
Bidding Curves: The bidding curves are piece-wise linear
convex functions where price-volume pairs determines each
line piece. The problem of finding both prices and quantities
results in a non-linear problem. To avoid non-linearities the
approach of [10] has been used where price points (pi)are fixed and the problem is solved for quantity variables
(xit). The interpolation between the price-volume points and
the realized price in each scenario (ρst ) gives the producers
committed volume (yst ) (5).
yst =
⎧⎪⎪⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎪⎪⎩
ρst−p1
p2−p1x2t +
p2−ρst
p2−p1x1t if p1 ≤ ρst < p2
...ρst−pi−1
pi−pi−1xit +
pi−ρst
pi−pi−1xi−1t if pi−1 ≤ ρst < pi
...ρst−pI−1
pI−pI−1xIt +
pI−ρst
pI−pI−1xI−1t if pI−1 ≤ ρst < pI
(5)
As required by the EPEX SPOT Operational Rules, Article
1.5.1 [12] the bid has to be monotonous. This implies that
with increasing prices the volume has to equal or exceed the
previous bid volume (6).
xit ≤ xi+1,t (6)
By fixing the set of price points as in [10], bidding decisions
might have to be made for prices that are not within the set
of realized prices (ρst ) for the respective time step. To reduce
the risk of infeasible commitments quantity variables for price
points outside the range of realized prices for each hour is
set equal the quantities assigned to the last price point with
information (7) (8).
xi+1,t = xit |max(ρst ) < pi (7)
xi,t = xi+1,t |min(ρst ) > pi (8)
Balancing Market: The model includes the use of a balanc-
ing market to satisfy second stage restrictions. The balancing
market is modeled as a penalty function (11) and upper limits
on the sales and purchases in the balancing market (9), (10).
To promote feasible first stage decisions, both the sale and
purchase in the balancing market is related to a penalty. The
penalty is a fixed deviation (R) from the electricity price.
+qst ≤∑j∈J
Pj (9)
−qst ≤∑j∈J
Pj (10)
Cbal(+qst ,− qst ) = (ρst +R)−qst − (ρst −R)+qst (11)
Thermal Constraints: Thermal power plants have technical
constraints that couples the production in different time peri-
ods together. The formulation in [4] is used to describe these
technical constraints. The maximum and minimum output of
power for each unit is restricted in (12) and (13).
zsjt ≤ Pjusjt (12)
zsjt ≥ Pjusjt (13)
The maximum change in production in one unit from one
time step to the next is limited by the ramp-up (14) and ramp-
down (15) restrictions.
zsjt ≤ zsjt−1 +RLj (14)
zsjt ≥ zsjt−1 −RLj (15)
The limitation on minimum up and down time is enforced
by (16) and (17).
t∑τ=t−UTj+1
wsjτ ≤ us
jt (16)
t∑τ=t−DTj+1
vsjτ ≤ 1− usjt (17)
Balance Constraints: Equation (18) ensures that production,
commitment and use of the balancing market are in balance.
∑j∈J
zsjt = yts − q+st + q−s
t (18)
Logical Constraints: Equation (19) defines the start- and
stop variables. If a unit runs in one time step (usjt = 1) and
is off in the previous time step (usjt−1 = 0) the start variable
(wsjt) is set to equal 1. The stop variable (vsjt) will in the
same manner be set to 1 if the opposite is true. Equation (20)
specifies that a unit cannot start and stop in the same time
period.
wsjt − vsjt = us
jt − usjt−1 (19)
wsjt + vsjt ≤ 1 (20)
The cost of a start-up increases with the time the plant has
been shut down. This is modeled with different start-up types:
hot, warm and cold. To assign the correct start-up type to the
start-up variable, a time interval is specified that determines
each start-up type (21) [4]. The start-up variable is further
limited in (22) by the fact that there can only be one start-up
type in any time period on any generator.
wlsjt ≤
t−T lj∑
τ=t−Tl
j+1
vsjt (21)
wsjt =
∑l∈L
wlsjt (22)
Initialization Constraints: The state of the system at the
start of the planning period influences the bidding problem.
Instead of a predefined initial state the starting conditions
are model by a wrap around formulation similar to what is
done in [13]. For the production (zsjt) and commitment (usjt)
variables this is done by connecting the initial period (t = 0)
to the last period (t = Tmax) (23) ,(24). For the start (vsjt) and
stop variables (wsjt) a larger time frame is needed to determine
the start-up type and the minimum up-time and down-time
(16), (17). This is formulated in (25), (26).
zsj,0 = zsj,Tmax(23)
usj,0 = us
j,Tmax(24)
vsj,t−Tmax= vsj,t (25)
wsj,t−Tmax
= wsj,t (26)
B. Price-independent Bidding
The price-independent bidding problem is a stochastic two-
stage problem similar to the price-dependent bidding problem.
In the price-dependent model price points were used to give
volumes at different prices. In the formulation of the price-
independent model, price points are not needed since the bid
only consists of one volume per hour. The whole set I can
therefore be left out of the model. The bid variable will thus
only be dependent on time step t. The omission of price points
leaves restrictions (5), (6), (7) and (8) redundant. However, to
keep the coupling between the volume bid (xt) for each hour
and the committed volume (yst ), equation (27) is needed.
xt = yst (27)
V. NUMERICAL RESULTS
A time series model for the German-Austrian spot market
was developed and prices for Wednesday 2 October 2013
was simulated using this model (Appendix A). 5000 scenarios
were generated and then reduced through a scenario reduction
algorithm [14] to obtain a workable set of scenarios that also
ensured convergence of the model.
The portfolio of generators is based on the portfolio used in
[4] and consists of two lignite-based generators, two combined
cycle gas turbines (CCGT) and one open cycle gas turbine
(OCGT). The generator costs and thermal properties can be
found in Tables I and II.
The price points in the optimization model was selected
to define intervals that equally divide the range of prices in
the scenarios into 10 sorted groups. This number of price
points was found sufficient to represent the price variation
using the approach of [15]. Running the two optimization
models presented in Section V resulted in optimal quantities
for these price points for all 24 time steps. The bidding curves
for both the price-dependent and price-independent model can
be viewed in Figures 1 and 2.
TABLE IGENERATOR COSTS [4], PARAMETERS cj , Cj , Cl
j
Nr Type Marginal Cost Commitment Start Cost [EUR][EUR/MWh] Cost [EUR/h] Hot Warm Cold
1 lignite 29 1894 46600 64007 872172 lignite 31 1644 58165 79892 1088623 CCGT 55 3367 16012 24832 424724 CCGT 55 3839 19766 30476 518965 OCGT 85 965 2568 2568 2568
TABLE IIGENERATOR THERMAL PROPERTIES [4], PARAMETERS Pj , Pj , RLj , T l
j ,
UTj , DTj
Nr Power: [MW] MaxRamp Time from stop to: Min. time: [h]Max. Min. [MW/min] Warm [h] Cold [h] Start-Stop Stop-Start
1 274 160 2 5 12 8 42 342 180 2 5 12 8 43 378 200 24 5 12 4 34 476 250 24 5 12 4 35 152 63 8 5 12 1 1
Time [h]
0
5
10
15
20
Price [EUR/MW
h]
−50
0
50
100
150
200
Power [M
W]
0
500
1000
1500
Fig. 1. Bidding curves for the price-dependent model
To evaluate the quality of the bidding decisions, the bid-
ding curves were tested in an out-of-sample test. The seven
consecutive Wednesdays from EPEX SPOT starting October
2nd were used as input. The test was performed by fixing
the bidding decisions, and running the optimization models
with the new scenario set. Table III shows the result from this
analysis. The results illustrate the value of price-dependent
bidding where the price-dependent bids performs better than
the price-independent bids for all scenarios based on a penalty
of 0.5 EUR/MWh for the use of the balancing market.
VI. DISCUSSION
The optimal bids generated in this paper were based on
simulations of the behavior of the German-Austrian spot mar-
ket and on a specific portfolio of generators. In the German-
Austrian wholesale market prices have fallen as a result
Time [h]
0
5
10
15
20
Price [EUR/MW
h]
−50
0
50
100
150
200
Power [M
W]
0
500
1000
1500
Fig. 2. Bidding curves for the price-independent model
TABLE IIIPROFIT, VPDB AND RVPDB WITH SEVEN OUT-OF-SAMPLE SCENARIOS
FOR FIXED BIDDING DECISIONS FROM BOTH OPTIMIZATION MODELS,WITH PENALTY OF 0.5 EUR/MWH FOR THE USE OF THE BALANCING
MARKET
Scenario Price-dep. Price-indep. VPDB RVPDBbidding bidding
1 27233.4 26746.3 487.1 1.82 %2 250750.0 250650.0 100.0 0.04 %3 214453.0 214283.0 170.0 0.08 %4 -6627.2 -7131.0 503.8 7.06 %5 -6727.1 -7131.0 403.9 5.66 %6 -30379.0 -30871.0 492.0 1.59 %7 173380.0 173284.0 96.0 0.06 %
Mean 88869.0 88547.2 321.8 2.33 %
of the feed-in-tariff of renewable energy, this has degraded
margins for thermal power producers, and made the prospects
of turning off generators or reduce the output of generators
relevant. Price-dependent bids are able to reduce the impact of
unpredicted price drops and therefore provide superior bidding
decisions. For markets with higher prices and less volatility
improved margins might make price-independent bids suffi-
cient as generators might run continuously at maximum power.
The low prices in the market resulted in dispatch of only
the low-cost and less flexible generators. Thus, the price-
dependent and price-independent bids were based on very
similar dispatch solutions. For different market conditions the
optimal dispatch might be significantly different in the two
models. This paper shows however, that there will always exist
price-dependent bids that out-perform the price-independent
bids. These price-dependent bids are simply obtained by fixing
the unit commitment of the price-dependent model to the
unit commitment of the price-independent model and generate
price-dependent bids based on the flexibility in the committed
units.
VII. CONCLUSION
As prices have fallen and volatility risen in the German-
Austrian spot market thermal power producers trading on
EPEX SPOT is presented with a new reality. Prices might
fall below marginal cost of generators and even fall below
zero for a number of hours in the day-ahead market. This
might make the prospects of turning off generators for a
number of hours viable. The results from this paper show
that by submitting price-dependent bids, the power producer
is able to reduce losses in scenarios with low prices and
improve gains in scenarios with higher prices compared to
price-independent bids. Thermal power producers trading on
EPEX SPOT might incorporate some of the ideas presented in
this paper to construct price-dependent bids that will improve
their returns in the market.
APPENDIX
PRICE FORECASTING MODEL
An ARMA time series model is used to forcast the price
for 2 October 2013. This model is based on price data for
the German-Austrin spot market from 1 January 2013 to 1
October 2013.
(1− φ1B1 − φ2B
2 − φ24B24) log(yt − α)
= (1− θ24B24)εt (28)
TABLE IVESTIMATED PARAMETERS FOR THE ARMA PRICE MODEL
φ1B1 φ2B2 φ24B24 θ24B24 α
1.0454 -0.1391 0.9927 -0.9042 5.2312s.e. 0.0122 0.0122 0.0017 0.0072 0.0344
ACKNOWLEDGMENT
The authors would like to thank Christian Skar for valuable
feedback and FICO for access to their Xpress solver.
REFERENCES
[1] T. J. Larsen, “Daily scheduling of thermal power production in aderegulated electricity market,” Ph.D. dissertation, Dept. Elect. PowerEng., Faculty of Elect. Eng. and Telecommun., NTNU, Trondheim,2001.
[2] A. J. Conejo, F. J. Nogales, and J. M. Arroyo, “Price-taker biddingstrategy under price uncertainty,” IEEE Trans. Power Syst., vol. 17, no. 4,pp. 1081–1088, 2002.
[3] M. A. Plazas, A. J. Conejo, and F. J. Prieto, “Multimarket optimalbidding for a power producer,” IEEE Trans. Power Syst., vol. 20, no. 4,pp. 2041–2050, 2005.
[4] C. K. Simoglou, P. N. Biskas, and A. G. Bakirtzis, “Optimal self-scheduling of a thermal producer in short-term electricity markets byMILP,” IEEE Trans. Power Syst., vol. 25, no. 4, pp. 1965–1977, 2010.
[5] B. K. Pokharel, G. Shrestha, T. T. Lie, and S.-E. Fleten, “Price basedunit commitment for gencos in deregulated markets,” in Proc. PowerEngineering Society General Meeting. IEEE, 2005, pp. 428–433.
[6] S. W. Wallace and S.-E. Fleten, “Stochastic programming modelsin energy,” in Handbooks in Operations Research and ManagementScience, A. Ruszczynski and A. Shapiro, Eds. Elsevier, 2003, vol.Volume 10, pp. 637–677.
[7] S. de la Torre, J. M. Arroyo, A. J. Conejo, and J. Contreras, “Price makerself-scheduling in a pool-based electricity market: a mixed-integer LPapproach,” IEEE Trans. Power Syst., vol. 17, no. 4, pp. 1037–1042,2002.
[8] M. M. Roggenkamp and F. Boisseleau, Regulation of Power Exchangesin Europe. Antwerpen: Intersentia, 2005, vol. 2.
[9] M. P. Nowak, R. Schultz, and M. Westphalen, “A stochastic integerprogramming model for incorporating day-ahead trading of electricityinto hydro-thermal unit commitment,” Optimization and Engineering,vol. 6, no. 2, pp. 163–176, 2005.
[10] S.-E. Fleten and E. Pettersen, “Constructing bidding curves for a price-taking retailer in the Norwegian electricity market,” IEEE Trans. PowerSyst., vol. 20, no. 2, pp. 701–708, 2005.
[11] S.-E. Fleten and T. K. Kristoffersen, “Stochastic programming for opti-mizing bidding strategies of a nordic hydropower producer,” EuropeanJournal of Operational Research, vol. 181, no. 2, pp. 916–928, 2007.
[12] EPEX SPOT SE, “EPEX Spot operational rules,” Sep 2013.[13] A. B. Philpott, M. Craddock, and H. Waterer, “Hydro-electric unit com-
mitment subject to uncertain demand,” European Journal of OperationalResearch, vol. 125, no. 2, pp. 410–424, 2000.
[14] H. Heitsch and W. Romisch, “Scenario reduction algorithms in stochasticprogramming,” Computational Optimization and Applications, vol. 24,no. 2-3, pp. 187–206, 2003.
[15] N. Lohndorf, D. Wozabal, and S. Minner, “Optimizing trading decisionsfor hydro storage systems using approximate dual dynamic program-ming,” Operations Research, vol. 61, no. 4, pp. 810–823, 2013.