Upload
myra-bruce
View
217
Download
1
Embed Size (px)
Citation preview
Capacity Credit for WindGeneration Adequacy Committee, NWPPC
Michael Milligan, Consultant
National Renewable Energy Laboratory Apr 26, 2007
Outline
• Why is capacity credit important?• Reliability-based method• LOLP, LOLE, and other important details• Representations of wind in models
– Single-pass methods– Sequential Monte Carlo methods
• Assessment of reliability-based methods• Simplified methods as approximations• Wind capacity credit in the U.S.• Desirable properties of the metric, and an
assessment of methods used in the US
Why is Capacity Credit Important?
• Capacity credit is a way of quantifying relative plant contributions to reliability in the context of overall system reliability
• Achieving acceptable system reliability is the goal
• Too much reliability provides small incremental benefit for large cost
Why is Capacity Credit Important?
• Capacity from any generator is not guaranteed when needed– Some generators are more likely to fail than others– Generators that contribute significant reliability
have more value (reduce outage risk)– …compared to generators that don’t contribute
significantly to reliability (not much impact on risk reduction)
• Reliability is expensive• Outages are expensive• Cost-benefit tradeoff
Why is Capacity Credit Important?
• What fraction of wind should count for planning margins?
• What other capacity is needed?
Uses for Capacity Credit
• PJM-style capacity market that involves capacity payment to generators (~$1/MWh)
• MAPP-style adequacy assessment to determine match between load and resources
• CA-style ranking of generators bidding to load-serving entities (LSE’s) for long term contracts
• Utility planning (PacifiCorp)• Power pool, balancing authority, reliability
region adequacy assessments and/or determining capacity shares to meet reliability
The System is Complex
Load
Distribution System
Capacity Credit: Generation Reliability
Reliability of generation power supply depends on how all components work together. Transmission/distribution also influences reliability, but not considered here.
Measuring Capacity Credit
• Capacity credit based on reliability analysis• Large and established literature• Effective load carrying capability (ELCC) is preferred
method• ELCC is data-driven, empirical approach based on
– Hourly load profile (year)– Actual generating unit data– Units with large forced outage rates have lower ELCC– Small units have smaller ELCC
• Other definitions of “capacity credit” have emerged, but often don’t recognize differential contributions to reliability among alternative power plants
Capacity Credit Calculations• Can be adapted to wind generators• Wind capacity credit depends on output
profile (hourly for at least one year):– Low when wind contributes small amount to
reliability– High when wind contributes large amount to
reliability– Depends on system and wind characteristics– Values can range from approximately 10%-40%,
depending on system and wind characteristics– Capacity credit outside this range are possible
– Use multiple years of data if available
How Does ELCC Work?
• Holds the system at constant annual risk level with/without wind
• Can be measured relative to a perfect unit or selected benchmark unit
• Utilizes reliability/production simulation model– Hourly loads– Generator characteristics– Wind generation pattern (hourly for >= 1 year)– Calculates hourly LOLP (loss of load probability)
• Conventional units ELCC is function of FOR
LOLP, LOLE, and All That Jazz…
• LOLP = Loss of Load Probability (0 <= LOLP <= 1)
• LOLE = Loss of Load Expectation– Sum of probabilities over a given time
period– Product of time and probability– Often expressed as either days/year,
days/10-years, hours/year
Sample Hourly LOLP
1.00E-10
1.00E-08
1.00E-06
1.00E-04
1.00E-02
1.00E+00
0 200 400 600 800
Top LOLP Hours
LO
LP
(L
og
Sca
le)
ELCC can be thought of as the amount by which the area underthe curve is reduced by wind, relative to a benchmark unit.
ELCC as a Function of FOR
0
20
40
60
80
100
120
10 20 30 40 50 60 70 80 90
Forced Outage Rate (FOR) %
EL
CC
as
% R
ate
d C
ap
ac
ity
Generic 100 MW PlantRelative to Gas Benchmark Unit
Can be approximated with unforced capacity (1-FOR)*Cap
Steps to Calculate Wind ELCC
• Develop benchmark system excluding wind and benchmark unit
• Adjust hourly loads to achieve target reliability level (1d/10y is common)
• Add wind and rerun model, noting annual reliability (this is the wind case)
• Remove wind and incrementally increase benchmark unit capacity until annual reliability matches the wind case
• The capacity that was added is the ELCC of the wind plant
0.04
0.06
0.08
0.10
0.12
0.14
LO
LE
da
ys/y
r
800 900 1000 1100 1200 1300 1400 Load (MW)
Wind Plant Capacity Credit ExampleReliability Curves With/Without Wind
1,132 ELCC With Wind1,087 ELCC Without Wind
Wind Plant ELCC = 45 MW
Example: ELCC Relative to Gas Benchmark Unit
7% Forced outage rate (FOR)
Reliability Curve
0.0000000
0.0000500
0.0001000
0.0001500
0.0002000
0.0002500
0.0003000
0.0003500
0 40 70 100 130 160 190 220 250 280
Gas Equivalent Capacity
LO
LE Gas Equivalent
Renewable
How is LOLP Calculated
• Reliability model– Hourly loads– Generation characteristics
• Capacity• Forced outage rates
• For each hour a capacity table is calculated that shows– Levels of generation and associated probabilities– Generation out of service and associated
probabilities
Representations of wind in models
• Question: What wind data do we feed into the LOLP convolution algorithm?
• Simplest approach:– Treat wind as negative load– Requires hourly wind production estimate
that is time-synchronized with load
Wind as Load Modifier• Most reliability models don’t know about wind• …so use the load modifier approach with
wind modeled as hourly transaction, run-of-river hydro, etc.
• Advantages of the approach– Takes wind variability into account– 8,760 LOLP values based on actual data– Recognizes underlying correlation with load – Can repeat the analysis for several years, data
permitting– Caution: load forecasting algorithms can distort
the load profile for future years (Xcel/CO)
Wind as Load Modifier
• Dis-advantage of this approach– Treats the hourly wind generation as certain
(useful for post-hoc analysis)
• Other considerations– Useful for post-hoc analysis, such as PJM or other
capacity market with true-up– Variation in day-to-day loads less than day-to-day
variation in wind:• …so in effect we have many quasi-stochastic samples
pairing many wind scenarios with common daily load scenarios
Other Modeling Approaches
• X Typical day per– Week– Month– Need to get the annual energy correct– Advantage: retain some of the inherent variability– Dis-advantage: lose wind-load pairings,
understates wind variability (this is generally not good option)
• X Average output (flat)– Works reasonably well for energy, not for reliability
or capacity credit – not a good option
• Stochastic multiple-block model (next)• Sequential Monte Carlo (later)
Conventional Stochastic Approaches: Multiple-block Unit
• Wind as a multiple-block generator
• For each month, calculate several discrete generation levels and frequency of occurrence according to reliability model requirements
• Partition by hour of day
• Result: 24 distributions per month, each representing a given hour of the day
Conventional Stochastic Approaches: Multiple-block Unit
• Shortcomings– Tends to smooth wind generation relative to what
would be experienced in practice– Not available in all models or time-varying ability
may be limited/non-existent– Misses autocorrelation between successive hours’
generation but probabilistic results smoothes this
• Advantages– Computationally faster than Monte Carlo– Reasonable for planning– Can often be modeled with changing block
probabilities (forced outages) in each month
Conventional Stochastic Approaches: Monte Carlo
• Markov or state-transition matrix as implemented by the model– Shortcomings:
• Model may only allow for a single STM per year (not acceptable) X
• May lose wind/load systematic correlation
– Advantages: handles stochastic nature of the resource; picks up the hourly transition between wind output states
Sequential Monte Carlo
• Basic Approach– Build probabilistic model of wind resource or wind
generation– Repeatedly sample from the family of distributions– Run reliability model for each simulated year of wind data– Collect results
• Computationally expensive– May not adequately capture wind-load synergies– Can be difficult to obtain synthetic time series that
adequately represent the complex correlation and auto-correlation structure in the real wind generation patterns – requires advanced time-series modeling with complex stochastic properties
Sequential Monte Carlo (SMC)
• Does reliability model have capability?– If not, wind time-series can be realized
outside the reliability model– Typically requires use of state transition
matrix (STM)– A single STM for a year is not enough
• At least seasonal (depends on wind regime)• Monthly• Diurnal?
– Wind speed or wind power?
NREL Application of SMC
• 1999 Wind Energy Journal paper
• Created simulated wind speed based on monthly STM: each realization = 1 year
• Converted each realization to wind power
• Ran reliability model for each realization
Example One-Month STM
0.0 0.2 0.4 0.6 0.8 1.0
Rela
tive F
requency
0 6 12 18 24 Velocity(t-1) m/s
18 12
6
Velocity(t) m/s
Other Analysis
• 2nd-order Markov processes (time-of-day conditional STM)
• Comparison to Markov– Statistical properties of Markov matched
high-plains sites– Statistical properties of 2nd-order Markov
matched CA sites with significant diurnal component
Simpler Approaches to Capacity Credit
• Calculate the wind capacity factor over the top 10-20% of load hours over the year– Usually provides a slight under-estimate of
ELCC– Applied by Milligan & Parsons, NREL
• Wind Capacity factor over peak period– PJM 3-year average, HE 3-6 Jun-Aug– NYSERDA/GE Study Phase 2 – CA RPS Integration Cost Study/Multi-Year
Analysis
Example CalculationsCapacity Credit vs. Capacity Factor (Year 1)
0
0.1
0.2
0.3
0.4
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Top Loads (%)
Cap
acit
y F
acto
r/C
red
it
Capacity Factor Capacity Credit
Capacity Credit vs. Capacity Factor (Year 4)
0
0.1
0.2
0.3
0.4
0.5
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Top Loads (%)
Cap
acit
y F
acto
r/C
red
it
Capacity Factor Capacity Credit
Capacity Credit vs. Capacity Factor (Year 5)
0
0.1
0.2
0.3
0.4
0.5
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Top Loads (%)
Cap
acit
y F
acto
r/C
red
it
Capacity Factor Capacity Credit
Why Don’t Capacity-Factor Methods Match ELCC?
• MW on peak carries more reliability benefit than near-peak
• MW contribution declines with LOLP rank
Sample Hourly LOLP
1.00E-10
1.00E-08
1.00E-06
1.00E-04
1.00E-02
1.00E+00
0 200 400 600 800
Top LOLP Hours
LO
LP
(L
og
Sca
le)
Why Don’t Capacity-Factor Methods Match ELCC?
• Risk may not be perfectly correlated with load– Scheduled maintenance during shoulder
months may increase risk relative to peak– Dispatchable hydro can lower on-peak risk,
increase near-peak risk– Other on-peak transactions that aren’t
available during near-peak hours– Other factors that shift risk profile
Hydro and No-Hydro LOLP
0
10
20
30
40
50
60
70
0 20 40 60 80 100
Top Load Hour
Ab
s P
ct D
iff
Abs Pct Diff Mean Abs Pct Dif
31000
33000
35000
37000
39000
41000
43000
0 50 100 150 200 250
Rank
Lo
ad
(M
W)
Ranked by load
Ranked by LOLP
Ranked by load net hydro and interchange
Hydro and Interchange Affect The Risk Profile and ELCC
Other Approximation Methods
• Reliability-based approaches– Garver’s approximation
mimics risk curve– CA RPS Integration Cost
Analysis Phase III Appendix uses actual risk curve, applies simpler method on the risk curve (does not work well with hydro)
Garver’s Approximation
R’ = Exp{-[(P-L)/m]}
• P = annual peak load, L = load for the hour in question. R’ is the risk approximation (LOLP), measured in relative terms (peak hour risk = 1).
• Garver’s constant, m is calculated with a reliability model
Garver’s Approximation• Construct a spreadsheet that calculates R’ for
the top loads. Then modify the values of L by subtracting the wind generation in that hour.
• Calculate LOLE approximation for (a) no-wind case and (b) wind case by summing the hours. Use all hours for which no-wind risk exceeds some tolerance – probably around 500 hours. Compare to gas plant or other benchmark, de-rated by its forced outage rate.
• Could expand method to include dispatchable hydro and transactions
PacifiCorp’s Proposed Approach
• Z-method
• Current IRP work
• Benchmarked against ELCC
• Uses an approach to estimate ELCC based on cumulative probability distribution of hourly surplus generation during peak period
Selected Capacity Results
CA RPS Integration Cost ResultsRelative to Gas Reference Unit
0%
5%
10%
15%
20%
25%
30%
35%
40%
45%
2002 2003 2004
EL
CC
% o
f R
ate
d C
ap
ac
ity
Wind (Northern Cal)
Wind (San Gorgonio)
Wind (Tehachapi)
0
5
10
15
20
25
30
2002 2003 2004 3-yr Avg
ELCC
CapFac
Northern California
0
5
10
15
20
25
30
35
40
2002 2003 2004 3-yr Avg
ELCC
CapFac
San Gorgonio
ELCC Without Hydro, Exchange Compared to Capacity Factor Jun-Sep 12p-6p
0
5
10
15
20
25
30
35
2002 2003 2004 3-yr Avg
ELCC
CapFac
Tehachapi
CA RPS: Impact of Hydro, Transactions
Multi-year Capacity Value (Excludes Hydro, Interchange)
05
101520253035
Wind (NorthernCal)
Wind (SanGorgonio)
Wind(Tehachapi)
Wind Resource Area
Per
cen
t o
f R
ated
C
apac
ity
ELCC
Peak CapFac
Multi-year Capacity Value (Includes Hydro, Interchange)
05
101520253035
Wind (NorthernCal)
Wind (SanGorgonio)
Wind(Tehachapi)
Wind Resource Area
Per
cen
t o
f R
ated
C
apac
ity
ELCC
Peak CapFac
3-year Averages
Other Selected Wind Capacity Values
0
5
10
15
20
25
30
35
40
NYSERDA/GE MN/DOC/Xcel(1)
MN/DOC/Xcel(2)
MN/DOC/Xcel(3)
CO Green PacifiCorp
EL
CC
as
% R
ate
d
(1) Existing wind (2) Potential new wind (3) Potential w/Monte Carlo Analysis
Wind Capacity Credit in the U.S.
• California: Uses ELCC in RPS Integration Study
• PJM: Uses a proxy for ELCC by taking 4-hour period during summer peak months and calculating wind capacity factor – may do ELCC study in future
• Study for Minnesota Department of Commerce/Xcel Energy used ELCC with sequential Monte Carlo (Enernex)
Wind Capacity Credit in the U.S.
• ELCC method used by GE in study in study of New York for NYSERDA (capacity factor during 1:00-5:00 Summer months)
• NYISO 4-hour test – revisiting approach• ISONE: (1-capacity factor), may revise/Cape Wind• Colorado PUC and Xcel/CO agreed on ELCC method
in mid-1990’s• Xcel/CO 10-year ELCC study: 12.5% but load profile
distortion puts results in question• Rocky Mountain Area Transmission Study (RMATS)
used a fixed 20% of rated capacity• PacifiCorp IRP used ELCC and Sequential Monte
Carlo (20%) in past IRP; Z-method proposed in most recent IRP
Mid-continent Area Power Pool (MAPP)
• Monthly method• Based on actual wind power data
– Up to 10 years if available
• Monthly time window of 4 hours including peak hour
• Capacity credit is the median value of (up to) 10 years of wind data in the window, repeated for each month
Southwest Power Pool (SPP)• Monthly: select top
10% of load hours• For those hours,
calculate the 85th percentile of wind generation (i.e. the level of generation that would be exceeded 85% of the time during the time period)
The 85th Percentile Myth
• Why doesn’t this work?– Percentiles are unstable statistics– Implies that capacity supplied less often than 85%
doesn’t count– Conventional unit outage during peak period would
receive 0 capacity value for that month for almost 7 years using this method
– Eliminates capacity value of any generator with outage rate > 15%
– System can be reliable (but expensive) even if some units deliver < 85% of the time
Can 1d/10y Be Achieved with Units with FOR>15%? (Yes, but $)
Number of 100 MW Units to Achieve 1d/10y Reliability Target @ Different
Forced Outage Rates
0
200
400
600
800
1,000
1,200
0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
Forced Outage Rate for 100 MW Units
No
. 1
00
MW
Un
its
54 Units 61 Units
Properties of a Good Capacity Metric
• Consistent in the way all generators are treated– Horizontal consistency: two generators with identical
(similar) reliability properties should be treated identically (similarly)
– Vertical consistency : a generator with higher reliability/ability to deliver on peak should have higher capacity value than generator that is not able to consistently deliver
• Based on accepted reliability theory and practice
Properties of a Good Capacity Metric
• Reflect the risk-reduction contribution of any generator
• Capture the importance of load shape• Reflect delivery pattern relative to load
shape• Mathematically consistent• Data driven• Simple
Assessment of Methods
• ELCC– Reliability model and stochastic treatment of wind
needs refinement (NREL’s sliding window)– Sequential Monte Carlo needs refinement– Interplay between resources– Not transparent– But is consistent method
• MAPP– Arbitrary window– Not specifically linked to reliability theory– Median is unstable statistic
Assessment of Methods
• PJM-like method (NY/GE, CA RPS)– May miss hours of high-risk that ELCC captures,
especially if used by utilities with high summer and winter peak, but interchange, hydro schedules may compensate…leading to good match w/ELCC
• SPP– Good start in selecting top 10% load hours– Median is unstable statistic– 85th percentile discounts available capacity– Diverges from the large body of reliability theory,
analysis, and application to wind
ELCC Properties
Horizontally consistent? Y
Vertically consistent? YReflect risk-reduction contribution of any generator YCapture load shape YReflect delivery pattern relative to load shape YMathematically consistent YData driven YTransparent ?Simple ?
Trade-offs
• Simple method – May over-simplify– May introduce arbitrary assumptions– May not be stable metric
• ELCC – data intensive– not necessarily transparent– sensitive to many other conditions, may not
be stable
Recommended Methods
• Multi-year ELCC is best approach• Approximation methods often work well, but
should be benchmarked against ELCC as a calibration
• Any simplified method should be based on data, benchmarked with ELCC, and accepted practice
• Consistent with conventional generation• Recognize system reliability and individual’s
contribution
Summary• ELCC can be data intensive but based
on solid analytics• Approximation methods can be
applied, more transparent, easier to calculate
• …but may not be based on solid analytics and/or inconsistent
• Xth percentile/median methods aren’t stable nor are they based on reliability methods
• More work required on Sequential Monte Carlo
• (see M. Milligan and K. Porter, “The Capacity Value of Wind in the United States: Methods and Implementation,”, Electricity Journal, Electricity Journal, Vol. 19, Issue 2, March 2006. pp 91-99. Elsevier, Inc.)