Title: Offer Construction for Generators with Intertemporal Constraints via
1Offer Construction for Generators with
Inter-temporal Constraints via Markovian DP and
Decision Analysis Grant Read, Paul Stewart Ross
James and Deb Chattopadhyay
2Presentation Outline
- Deregulated Electricity Market Offering Process
- Intertemporal Constraints
- Literature
- Two-Phase DP Algorithm Concept
- DA/DP Concept
- Conclusions
2
3Electricity Markets
Rest-of-Market Supply Curve
Demand Curve
Total Supply Curve
P
P
offer
Q
Q
3
4Player perspective
Note, we are only optimising a single-player
reaction to the assumed residual demand curve,
not finding a multi-player equilibrium
Price (/MWh)
Generator offer
Residual Demand Curve
P
P
Generator marginal cost
Q
Quantity (MW)
Q
Generator profit
4
5Intertemporal Considerations
- Market outcome expectations and correlations
- Reactions of rival market participants
- Linkages within a Company
- Water/Fuel storage limits
- Reservoir Balancing
- Inflow correlations
- Unit operational rules (ramping,
startup/shutdown etc) - We will focus on the impact of
- Market price correlations
- For an energy limited (hydro or thermal) plant
5
6Intuition
- Inter-temporal interactions make a difference
- Some may be managed by adjusting the shape of the
(monotone) offer curve - Provided an increase in this periods price
always implies an increase in desired output - Others may have to be managed by dynamically
moving the (vertical) offer curve - Eg if an increase in this periods price may
imply a decrease in desired output, so as to
save fuel/water for even higher prices implied
for later
6
7Optimal offer curves with correlation
- Optimal offer curves for various hours of the day
- Derived by plotting price/quantity pairs
- Assuming higher price now implies a higher price
path for the rest of the day
8Literature
- Philpott, Zakeri, Pritchard, etc
- Optimal offers for a Market Distribution
Function (roughly) a set of possible residual
demand curves
- Rajaraman Alvarado (2003). Optimal Bidding
Strategy in Electricity Markets Under Uncertain
Energy and Reserve Prices - Two-Level Dynamic Program
- Lower Level Determining the optimal offer for
the given state - Upper Level Moving back through horizon,
constructing a value curve (defined over
reservoir level) for each previous market
outcome. - Two-Dimensional (Markov) Dynamic Program for
upper level - Reservoir Level, Previous Market Outcome
-
8
9Upper Level Markov DP
1
1
1
Market States
T
T-2
T-1
2
2
2
9
10Lower Level DP
Price (/MWh)
Quantity (MW)
(Similar to Philpott et al, noting that there may
be many possible residual demand curves
corresponding to each market state)
10
11Our Goals
- Given that Alvarado et al have developed this
basic approach, our goals are - To improve its computational efficiency by
efficiently separating the upper/lower DP levels
in a two-phase approach - To extend it to cover more general uncertainty
structures - (To test our intuition with respect to optimal
offer patterns)
11
12Two-Phase DP Approach
- This offering process needs to be done once
every half hour, and so it is important that the
computational complexity of what needs to be done
on the spot, in real-time is minimised.
- Two-Phase DP that we have developed recognises
that the construction of the offers at a given
market state is very repetitive and inefficient.
- Uses a concept we call marginal cost patching,
to enable the production of all the offers to be
brought out of real-time into a Pre-Processing
Phase
- The Real-Time Phase that remains is highly
efficient and can deal with highly detailed
representations of a generators situation in a
reasonable amount of time.
12
13Marginal Cost Patching
- The optimal offer price at any particular
quantity level depends only on the local marginal
cost level. It is therefore independent of the
marginal cost levels occurring at higher/lower
quantities - So segments of the optimal offers for constant
marginal costs can be patched together to provide
the optimal offer for a marginal cost curve that
is stepped - For example, the optimal offer for a firm with a
marginal cost curve that steps from MC1 to MC2 at
the quantity BP will be the combination of - The section of the offer from 0 to BP under the
assumption that marginal cost is equal to MC1
over the entire range, and - The section of the offer from BP to Qmax under
the assumption that marginal cost is equal to MC2
over the entire range
13
14Two-Phase DP Approach
- The actual marginal cost structure (eg water
value curve) is unknown at the start of the
analysis - It is determined internally by the DP,
- So we can not pre-compute optimal offer curves
- But Marginal Cost Patching enables us to bring
the lower level DP out of real time, by - Producing a set of offers for a set of fixed MC
levels in advance - Then, in the real time DP, for each period and
market state - We know the end-of period expected marginal water
value curve (by backwards recursion) so - We can quickly patching together segments of
fixed-MC offers to produce an offer curve
matching that marginal water value curve.
15Pre-Processing Diagram
Optimal offer curves for various assumed MC levels
Set of possible residual demand curves for this
market state
15
16Real-Time Algorithm
1
1
1
Market States
T
t-2
t-1
2
2
2
16
17Patching Diagram
Patched offer
(MC determined by backward recursion in real-time
DP)
17
18RT Phase Finding Values
(These price quantity pairs then provide a pdf
for outcomes for the period in the real time DP
recursion)
18
19Results Computation Time
- Results over 320 Test Instances covering a wide
range of the problem domain. Largest problem
considered - 40 Periods
- 300 Reservoir Levels
- 40 Dispatch Levels
- 100 Possible RD Curves per Period
- 110 Fixed MC Levels
80,324 seconds (22.3 hours)
4578 seconds (1.3 hours)
99 seconds
19
20Results Solution Quality
5.0
1.7
- These differences
- Are defined in terms of deviation from
(deterministic) optimality, as determined by ex
post simulation on a dataset which contains
some non-Markov correlations - Arise from differences in the degree of
approximation employed. The RA algorithm could
be made more accurate, but at even higher
computational cost
20
21Extension DA/DP
- Not all uncertainty is well described by a Markov
Chain - Often it will become clear at some point in time
that a new state of the world has arisen, and
will remain for some time. For example - A major breakdown
- A change in weather
- A different competitor strategy
- Our DA/DP approach models this structure using
- A DA decision Tree linking macro-states on a
coarse time scale - A set of Markov DPs optimising behaviour within
each macro-sate on a finer time scale
22DA/DP Structure (example)
22
23Results
- Despite added complexity, computational time
actually reduces, for the same TOTAL number of
micro-states - By about 25 in real time, and much more in
pre-processing - because the number of possible state transitions
etc reduces - But the real issues are
- Does this structure exist in the real world?
- How much do we gain by modelling it?
- And this obviously depends on the
strength of the structure - For our test problem set, the gain was actually
quite marginal (1.6 error gtgt 1.3 on average,
or 6.9gtgt5.6 worst case) - Still, it is better and significantly quicker
24Conclusions
- MC patching can be used to create a TWO-PHASE DP
- This separation greatly improves computational
efficiency, so more complex problems can be
considered - Solution quality is also superior to the original
RA algorithm - Generalisation to the DA/DP structure can
- Further reduce computation time and
- Further improve solution quality
- . Provided the real world actually exhibits this
structure - (Solutions exhibit the expected patterns with
respect to offer curve dynamics) - This methodology is quite workable for real-time
application to a single reservoir/stockpile
situation - Various extensions and variations are covered in
the thesis - (See http//www.mang.canterbury.ac.nz/people/ste
wart.shtml ) - But a multi-reservoir model is the obvious next
step