University of New South Wales Sydney, Australia, February 3, 2012

Wind Power Forecasting in Electricity Markets

Audun Botterud*, Zhi Zhou, Jianhui WangArgonne National Laboratory, USA*abotterud@anl.gov

Ricardo Bessa, Hrvoje Keko, Jean Sumaili, Vladimiro MirandaINESC Porto, Portugal

Project website: http://www.dis.anl.gov/projects/windpowerforecasting.html

Outline

Background– Brief Intro to Argonne National Laboratory– Wind energy in the United States

New Statistical Approaches to Wind Power Forecasting– Point forecasting– Uncertainty forecasting

Forecasting in Operational Decisions– System Operation: Unit Commitment and Dispatch– Wind Power Trading under Uncertainty

Concluding Remarks

2

Argonne is America's First National Laboratory and one of the World's Premier Research Centers Founded in 1943, designated a

national laboratory in 1946

Part of the U.S. Department ofEnergy (DOE) laboratory complex– 17 DOE National Laboratories

Managed by UChicago Argonne, LLC– About 3,200 full-time employees

– 4,000 facility users

– About $600M budget

– Main site: 1500-acre site inIllinois, southwest of Chicago

Broad research anddevelopment portfolio

Numerous sponsors in government and private sector

3

Argonne: Science-based Solutions to Global Challenges

Energy production, conversion, storage and use

National Security

Environmental Sustainability

Use‐inspired science and engineering…

… Discovery and transformational science and engineering

Major User Facilities S&T Programs4

5

The Decision Information Sciences Division Develops State-of-the-Art Energy Analysis Models

DEVELOPS decision support tools for energy systems analysis, power systems analysis, and environmental analysis that are:Useful, Usable, and Used APPLIES models to

– Conduct country/region/state/city-specificstudies for domestic and foreign clients

– Consult clients and lending agencies onspecific investments or energy project loans

TRANSFERS software tools by– Conducting training programs

• Energy demand forecasting, energy and electric system analysis, analysis of environmental impacts (first training course in 1978)

• International technical cooperation projects to provide technical support funded by World Bank/GEF, regional lending banks USDOE, USAID, IAEA, etc.

– Software licensing and distribution

Work on renewable energy include hydro, wind, and solar5

Argonne Campus

Bldg 221

Advanced Photon Source

6

7

The Best Land-Based Wind Resources in the United States Are in the Great Plains and Upper Midwest

8

Problem is: Not a Lot of People Live Where the Resource is

U.S. Population Density by County (July 1, 2009)

9

Over 160,000 Miles (450,000 Curcuit-Miles) of Transmission Lines To Move Power; But Will Need to be Upgraded for Large-Scale Renewables

9

Share of Wind Power in Selected Countries, 2010

10

LBNL: 2010 Wind Technologies Market Report.

Growth in U.S. Wind Power Capacity

11

Source: AWEA

Installed Wind Power Capacity in U.S. States

12

0.0

500.0

1000.0

1500.0

2000.0

2500.0

3000.0

3500.0

4000.0

‐120.0

‐70.0

‐20.0

30.0

80.0

130.0

180.01 7 13 19 25 31 37 43 49 55 61 67 73 79 85 91 97 103

109

115

121

127

133

139

145

151

157

163

Wind Po

wer [M

W]

Price [$/M

Wh]

Time [hour]

DA price RT price Wind power

Does Wind Power Influence Electricity Markets Today?

Negative prices (LMPs)

Wind power ramping events

Midwest ISO Wind Power and Iowa* LMPs, May 11-17, 2009:

*MEC Interface13

U.S. DOE’s 20% Wind Energy by 2030 Report

Explores “a modeled energy scenario in which wind provides 20% of U.S. electricity by 2030”

Describes opportunities and challenges in several areas– Turbine Technology– Manufacturing, materials, and jobs– Transmission and integration– Siting and environmental effects– Markets

Wind power forecasting is identified as a key tool to better handle uncertainty and variability from wind power in system operations

14

Brief Overview of Argonne’s Wind Power Research

Environmental Impacts of Wind Power– Impact on critical wildlife habitats– Visual impact analysis

Wind Turbine Reliability– Improved coatings and lubricants– Better gear box reliability

Wind Power Forecasting and Electricity Markets– Improved statistical forecasting models– Use of forecasting in operational decisions

Funded by DOE EERE’s Wind and Water Power Program (since 2008)

15

16

Project Overview: Development and Testing of Advanced Wind Power Forecasting Techniques

Collaborators: Institute for Systems and Computer Engineering of Porto (INESC Porto), Portugal

Industry Partners: Horizon Wind Energy and Midwest ISO (MISO)

Sponsor: U.S. Dept. of Energy (Wind and Water Power Program)

The project consists of two main parts:Wind power forecasting– Review and assess existing methodologies– Develop and test new and improved algorithms

Integration of forecasts into operations (power system and wind power plants)– Review and assess current practices– Propose and test new and improved approaches, methods and criteria

Goal: To contribute to efficient large-scale integration of wind power by developing improved wind forecasting methods and better integration of advanced wind power forecasts into system and plant operations.

http://www.dis.anl.gov/projects/windpowerforecasting.html

Outline

Background– Brief Intro to Argonne National Laboratory– Wind energy in the United States

New Statistical Approaches to Wind Power Forecasting– Point forecasting– Uncertainty forecasting

Forecasting in Operational Decisions– System Operation: Unit Commitment and Dispatch– Wind Power Trading under Uncertainty

Concluding Remarks

17

Wind Forecasts are the Result of Combination of a Diverse set of Models and Input Data

18

NWP Output Data Weather Data Off-siteMet Data

Site Power Gen& Met Data

Forecast Results

Physical Models Statistical Models

Statistical Wind to Power Models

19

We can train neural networks or other mappers with any optimization algorithm and define a training criterion to generate the adequate mapper (wind to power)

A classical performance criterion is Minimum Square Error (MSE)We are exploring new criteria based on Information Theoretic LearningEntropy, correntropy, etc

g(x,w)

Training algorithm

Training criterion

x OT (target)

Neural Network

W2P Models with Information Theoretic Learning (ITL)

Non-Gaussian nature of wind power forecast errors–Mean Square Error (MSE) is only optimal under Gaussian distribution

The ITL idea…–ideal case is when the error pdf is a Dirac function - all errors equal (of

the same value)–all errors equal, means perfect matching between output Y and target T,

by adding a bias to the output neuronRenyi’s Quadratic Entropy combined with Parzen pdf estimation

dz)z(flogH 2Y2R

N

1i

2iY ),(G

N1)(f̂ Iyzz

2ikk2 )yz(

2

12

ikk e2

1),yz(G

Gaussian Kernel

20

Results Comparison: Forecast vs. Realized Values

Comparison of day-ahead forecasts and realized output for wind farm in the Midwest– Training based on mean square error (MSE)– Training based on Information Theoretic Learning (MCC, MEE, MEEF, cMCC)

21

Statistical Methods for Uncertainty Estimation

Kernel Density Forecast (KDF)– Forecasts the full probability density function– Based on Kernel Density Estimation (KDE)

• Quantile-Copula• Nadaraya-Watson

– Choice of kernel function important• depends on the type of variable

– Time adaptive formulations

Quantile Regression (QR)– Estimates a set of quantiles (or intervals)– Commonly used for wind power forecasting– Linear and splines quantile regression– Potential problem: Quantiles may cross

22

Kernel Density Estimation

Illustration of Kernel Density Forecast

23

Forecast the wind power pdf at time step t for each look-ahead time step

t+k of a given time-horizon knowing a set of explanatory variables (NWP

forecasts, wind power measured values, hour of the day)

0

0.20.4

0.60.8

1

0

5

10

15

20

Wind S

peed (m/s)

Wind Power (p.u.)

Time-adaptive Nadaraya-Watson Estimator

1∙

1∙ ∙ ∙

1∙

forgetting factor

Recursive KDE Estimator Exponential Smoothing

for stationary data streams

for nonstationary data streams

|∙ , 1 ∙ ∙

∙ 1 ∙

knowledge of the model at time instant t, which is updated using recent values of measured wind power and NWP data

Time-adaptive Nadaraya-Watson Estimator

1

Bessa, Sumaili, Miranda, Botterud, Wang, Constantinescu, “Time-Adaptive Kernel Density Forecast: A New Method for Wind Power Uncertainty Modeling,” 17th Power System Comp. Conf., Stockholm, Sweden, 2011.

24

Probabilistic Forecast Evaluation : U.S. Midwest Wind Farm

calibration plot sharpness plot

Trade-off between calibration and sharpness

NW: KDF Forecast with NW estimator

SplinesQR: Splines Quantile Regression

25

Quantile/interval forecast:

Wind Power Forecasting – Main Findings

Point Forecasting– Mapping from wind to power by artificial neural networks– Development of training algorithms based on Information Theoretic Learning (ITL)– Testing model on wind farms in the U.S. Midwest– ITL criteria give substantial reductions in forecasting error

Uncertainty Forecasting– Development of time-adaptive Kernel Density Forecasting (KDF) algorithms– Testing model on wind farms in the Midwest and EWITS data– KDF tends to give slightly better calibration than quantile regression, whereas sharpness tends

to increase– Other advantages of KDF is that it provides a full probability density function

Adequate scenario generation and reduction– Very important for multi-stage decision problems

26

Outline

Background– Brief Intro to Argonne National Laboratory– Wind energy in the United States

New Statistical Approaches to Wind Power Forecasting– Point forecasting– Uncertainty forecasting

Forecasting in Operational Decisions– System Operation: Unit Commitment and Dispatch– Wind Power Trading under Uncertainty

Concluding Remarks

27

Handling Uncertainties in System/Market Operation

What are the impacts on the system?– Reliability (curtailment,..)– Efficiency (system cost, price..)

Source of uncertainty

Operating Reserve

∆ Load ∆ Generating capacity

Operating Reserves(spin and non-spin)

∆ WindPower

????

Increase operating reserves?

Change commitment strategy?- Stochastic UC

[MW]

28

Wind power forecasting

A Stochastic Unit Commitment (UC) Model w/Wind Power Uncertainty

Formulation using wind power forecast scenarios (s) w/probabilities (probs):

A two-stage stochastic mixed integer linear programming (MILP) problem– First-stage: commitment– Second-stage: dispatch

Objective function (min daily expected cost)

Energy balance (hourly)

Spinning Reserve balance (hourly)

Unit commitment constraints(ramp, min. up/down)

Z. Zhou, A. Botterud, J. Wang, R.J. Bessa, H. Keko, J. Sumaili, V. Miranda, “Application of Probabilistic Wind Power Forecasting in Electricity Markets”, Wind Energy, accepted, Dec. 2011. 29

∙ ,,

,,

∑ , , , , ∀ ,

, , α , , , ∀ ,

, , 1 , , , ∀ , Non-spinning Reserve balance (hourly)

Commitment Constraints (i, t)

Wang J, Botterud A, Bessa R, Keko H, Carvalho L, Issicaba D, Sumaili J, and Miranda V, Wind power forecasting uncertainty and unit commitment, Applied Energy, Vol. 88, No. 11, pp. 4014-4023, 2011.

Operating Reserves vs. Stochastic UC

Dynamic reserve requirement (spinning + non-spinning) +

Deterministic UC

30

Forecast quantiles Reduced scenario set

Stochastic UC + scenario set

Commitment schedule

Commitment schedule

Real-time dispatch

Real-time dispatch

Realized generation

Illinois Case Study: Assumptions

210 thermal units: 41,380 MW– Base, intermediate, peak units

Wind power: 14,000 MW– 2006 wind series from 15 sites in Illinois

(EWITS dataset)– 20% of load

Peak load: 37,419 MW– 2006 load series from Illinois

No transmission network

120 days simulation period (July 1st to October 31st, 2006)– Day-ahead unit commitment w/wind power

point forecast– Real-time reliability assessment commitment

(RAC) w/ probabilistic forecast 31

Case study focus is to analyze:-Use of probabilistic forecasting methods-Operating reserves vs. stochastic UC

0

5000

10000

15000

20000

25000

30000

35000

40000

111

122

133

144

155

166

177

188

199

111

0112

1113

2114

3115

4116

5117

6118

7119

8120

9122

0123

1124

2125

3126

4127

5128

61

Loa

d/W

ind

Pow

er (M

W)

Hour

Wind and Load in July-October 2006

Load

Wind

4.78%

20.60%

19.71%

30.95%

Generation Capacity

Combine Cycle Turbine

Gas Turbine

Nuclear

Steam Turbine

0

200

400

600

800

1000

1200

1400

1600

P1 PF-F0 PF-F1 PF-F2 PF-F3 PF-D1 PF-D2 PF-D3 SF-S0 SF-S1 SF-S2

Cos

t (M

$)Unserved loadUnserved nonspinning reserveUnserved spinning reserveStart-upFuel

Overview of total cost (Illinois, 4-months period )

32

Point forecast with no additional reserve too risky Stochastic unit commitment has the lowest total costs Dynamic reserves perform slightly better than fixed reserves Overall, more operating reserves lead to lower costs within the same categories

Fixed reserves Dynamic reserves Stochastic UCPerfect forecast

Z. Zhou, A. Botterud, J. Wang, R.J. Bessa, H. Keko, J. Sumaili, V. Miranda, “Application of Probabilistic Wind Power Forecasting in Electricity Markets”, Wind Energy, accepted, Dec. 2011.

Summary: Wind Power Uncertainty in System Operation

Probabilistic wind power forecasts can contribute to efficiently schedule energy and operating reserves under uncertainty in wind power generation

Dynamic operating reserves (derived from forecast quantiles)+ Well aligned with current operating procedures+ Lower computational burden- Does not capture inter-temporal events- Uncertainty not represented in objective function

Stochastic unit commitment (with forecast scenarios)+ Captures inter-temporal events through scenarios+ Explicit representation of uncertainty in objective function - More radical departure from current operating procedures- High computational burden

Important factors to consider in evaluating probabilistic approaches– Quality of probabilistic forecast– Risk preferences of system/market operator

Broader market design issues– Market timeline, deviation penalties, system flexibility, demand response

33

Profit, πh, from bidding into day-ahead market in hour, h:

p – priceq – quantitypen – penaltydev – deviation from schedule

Wind Power Trading under Uncertainty in LMP Markets

What is the optimal strategy?How much to bid into DA market?

Three stochastic variables:

What is the impact of risk preferencesand market design?

Botterud A., Zhou Z., Wang J., Bessa R.J., Keko H., Sumaili J., Miranda, V., “Wind Power Trading under Uncertainty in LMP markets,” IEEE Transactions on Power Systems, in press (available online).

)ˆ(ˆˆ DAh

RTh

RTh

DAh

DAhh qqpqp

devhDeviation penalty?

)( hdevpen

34

A Model for Wind Power Trading: Objective Functions and Risk Preferences

1) Risk Neutral: Expected Profit

2) Risk averse: Conditional Value at Risk (CVaR)

3) Risk averse or risk prone: Expected Utility

where

35

Conditional Value at Risk (CVaR)

Profit

Pro

babi

lity

dens

ity

th

CVaR is the expected value of the profit below threshold, th

Objective function: Max [E(Profit) + w*CVaR ]

36

Utility Function

Risk Prone

RiskAverse

RiskNeutral

Decision Maker’sPreference(Utility Function)

0.0

0.5

1.0

LowestProfit

HighestProfit

Objective function: Max E(Utility)37

Analysis of Horizon Wind Farm

Three months of data for forecasted and realized wind power and prices (synchronized)– Oct 22 2009 – Feb 19 2010– 2904 hours

Wind power forecasts– Kernel Density Forecast

• NW (default)

Price forecast is simple average of moving window– Normal distribution– No price-wind correlation

Summary of realized LMPs in this period at MISO hubs:

38

CINERGY FE ILLINOIS MICHIGAN MINNAvg DA 29.5 30.6 26.0 30.7 24.5Avg RT 29.0 30.2 24.5 30.1 23.7

StDev DA 10.5 10.7 11.3 11.5 13.2StDev RT 17.4 18.2 19.2 20.3 20.5#neg DA 4 4 21 9 29#neg RT 104 106 394 140 443

Corr DA-RT 0.55 0.51 0.55 0.48 0.58Corr DA-wind -0.02 -0.02 -0.05 0.00 -0.05Corr RT-wind -0.05 -0.05 -0.09 -0.01 -0.14

One Hour: Expected Profit vs. CVAR (no penalty)

39

avg

E*

C*

U*(β=‐3)

U*(β=3)‐10

‐8

‐6

‐4

‐2

0

2

6.15 6.2 6.25 6.3 6.35 6.4 6.45 6.5 6.55 6.6

CVAR [$/M

W]

Expected Profit [$/MW]

All bids

Optimal bids

E- expected value, C- CVaR, U – utility, avg – average forecast

One Hour: Optimal Bid Depends on Deviation Penalty

40

DA bid as function of deviation penalty

0.0

0.2

0.4

0.6

0.8

1.0

0 1 2 3 4 5 6 7 8 9 10

DA Bid Quantity

, qDA

Penalty [$/MWh]

E

C

U(β=‐3)

U(β=3)

avg

E- expected value, C- CVAR, U – utility, avg – average forecast

4 months simulation: Realized Profit vs. Deviation ($0/MWh penalty)

41

Wind Power Producer

Sys

tem

Ope

rato

r

Conflict of interest between wind power producer and system operator!

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

27000 28000 29000 30000 31000 32000

Avg Ab

s Deviatio

n [M

W]

Total Profit [$/MW]

avg

zero

E*

C*(w=0.1)

C*(w=0.3)

U*(β=‐3)

U*(β=3)

4 months simulation: Realized Profit vs. Deviation ($5/MWh penalty)

42

Now the interests are better aligned, but wind power profits significantly reduced- Just and reasonable treatment of wind power

Wind Power Producer

Sys

tem

Ope

rato

r

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

22000.0 24000.0 26000.0 28000.0 30000.0

Avg Ab

s Deviatio

n [M

W]

Total Profit [$/MW]

pf(median)

zero

E*

C*(w=0.1)

C*(w=0.3)

U*(β=‐3)

U*(β=3)

Wind Power Trading under Uncertainty: Main Findings

Trade-off between risk and return important for merchant wind power– Model assist in analyzing this trade-off under uncertainty in wind power and prices

Optimal day-ahead bid driven by price expectations (without penalty)– Day-ahead prices on average higher than real-time prices– Risk averse strategies give lower DA bids

Importance of market design– Potential of conflicting objectives between wind power producer and system operator– Deviation penalty brings optimal bids closer to expected forecast and reduces system

deviations, but reduces wind power revenue

43

Outline

Background– Brief Intro to Argonne National Laboratory– Wind energy in the United States

New Statistical Approaches to Wind Power Forecasting– Point forecasting– Uncertainty forecasting

Forecasting in Operational Decisions– System Operation: Unit Commitment and Dispatch– Wind Power Trading under Uncertainty

Concluding Remarks

44

Concluding Remarks

Rapid growth in renewable energy in the United States the last few years– Most investments in wind power so far– Increasing interest in solar energy– Vulnerable to policies/incentives and gas/electricity prices

A large-scale wind power expansion requires new operational approaches– How to efficiently handle increasing uncertainty and variability?

• System operator: Reserve requirements, unit commitment, dispatch• Wind power producer: Offering wind power into the electricity market

– Make efficient use of the information in the wind power forecast• Improved forecasting models (probabilistic forecasts, ramp forecasts)• Stochastic models to aid decisions under uncertainty

More general challenges– How to design markets that better accommodate wind power and other renewables?– How to make industry move up the technological ladder: adaptive, probabilistic methods– What will be the impact of a smarter grid and a more flexible demand side?

45

Selected Project References for More Details

46

More information: http://www.dis.anl.gov/projects/windpowerforecasting.html

Zhou Z., Botterud A., Wang J., Bessa R.J., Keko H., Sumaili J., Miranda V., “Application of Probabilistic Wind Power Forecasting in Electricity Markets,” Wind Energy, accepted, Dec. 2011.

Botterud A., Zhou Z., Wang J., Bessa R.J., Keko H., Sumaili J., Miranda, V., “Wind Power Trading under Uncertainty in LMP markets,” IEEE Transactions on Power Systems, in press (available online), Sept. 2011.

Bessa R.J., Miranda V., Botterud A., Zhou Z., Wang J., “Time-Adaptive Quantile-Copula for Wind Power Probabilistic Forecasting,” Renewable Energy, Vol. 40, No. 1, pp. 29-39, 2012.

Wang J., Botterud A., Bessa R., Keko H, Carvalho L., Issicaba D., Sumaili J., Miranda V., “Representing Wind Power Forecasting Uncertainty in Unit Commitment,” Applied Energy, Vol. 88, No. 11, pp. 4014-4023, 2011.

Bessa R.J., Miranda V., Botterud A., Wang J., “‘Good’ or ‘Bad’ Wind Power Forecasts: A Relative Concept,” Wind Energy, vol. 14, no. 5, pp. 625-636, July 2011.

Botterud A., Wang J., Miranda V., Bessa R.J., “Wind Power Forecasting in U.S. Electricity Markets,” Electricity Journal, Vol. 23, No. 3, pp. 71-82, 2010.

Mendes J., Bessa R.J., Keko H., Sumaili J., Miranda V., Ferreira C., Gama J., Botterud A., Zhou Z., Wang J., “Development and Testing of Improved Statistical Wind Power Forecasting Methods,” Report ANL/DIS-11-7, Argonne National Laboratory, Sep. 2011.

Monteiro C., Bessa R., Miranda V., Botterud A., Wang J., Conzelmann G., “Wind PowerForecasting: State-of-the-Art 2009,” Report ANL/DIS-10-1, Argonne National Laboratory, Nov. 2009.

University of New South Wales Sydney, Australia, February 3, 2012

Wind Power Forecasting in Electricity Markets

Audun Botterud*, Zhi Zhou, Jianhui WangArgonne National Laboratory, USA*abotterud@anl.gov

Ricardo Bessa, Hrvoje Keko, Jean Sumaili, Vladimiro MirandaINESC Porto, Portugal

Project website: http://www.dis.anl.gov/projects/windpowerforecasting.html