...

AWST’s Approach to Wind Power Production Forecasting in Alberta John W. Zack

by user

on
Category: Documents
32

views

Report

Comments

Transcript

AWST’s Approach to Wind Power Production Forecasting in Alberta John W. Zack
AESO Forecasting Pilot Project Conference
Calgary, AB
April 24, 2007
AWST’s Approach to
Wind Power Production
Forecasting in Alberta
John W. Zack
AWS Truewind LLC
Albany, New York
[email protected]
Overview
•
•
•
•
•
AWS Truewind (AWST) Introduction
AWST’s Approach to Forecasting
Forecasting Time Scales
AWST’s Alberta Forecast System
Forecasting the Future of Forecasting
AWS Truewind
Integrated Services to the Wind Energy Industry
• Mapping
•
•
•
•
Energy Assessment
Project Engineering
Performance Evaluation
Forecasting
• Mapping and Project Development
– Utilizes AWST’s resources assessment tools: MesoMap and SiteWind
– Constructed regional wind maps for over 25 countries and 50 states and regions
– Been involved in over 15,000 MW of project development
• Forecasting
–
–
–
–
Based on AWST’s multi-model forecast system: eWind
Currently contracted to forecast for over 3,500 MW in NA and Europe
Evaluated in 3rd party forecast performance studies (e.g. EPRI report #1007339)
Selected as forecast provider to several major grid operators: CAISO, ERCOT etc.
Overview of AWST’s
General Approach to
Wind Power Production Forecasting
How do we produce forecasts?
AWST’s Forecast System
Physics-based Models
• Differential
equations for basic
physical principles are solved
on a 3-D grid
• Must specify initial values of all
variables at each grid point and
properties of earth’s surface
• Simulates the evolution of the
atmosphere in a 3-D volume
• Many different models
• Eta, GFS, MM5, WRF, MASS etc.
• Same basic equations with subtle but
critical differences
• Can be customized for an application
Physics-based Models
Key Performance Factors
• Initial values for all prognostic
variables must be specified for
every grid cell
• Boundary values must be
specified for all boundary cells
(usually from another model
with a larger domain)
•
Grid has finite resolution - some processes are at the
“sub-grid” scale and feedback to affect grid scale
•
Surface properties (roughness, heat capacity etc.) of
the earth must be specified or modeled
Statistical Models
• Empirical equations are
derived from historical
predictor and predictand data
(“training sample”)
• Current predictor data and
empirical equations is then
used to make forecasts
• Many types of models
• Time series, MLR, ANN, SVR
• More sophisticated does not
always mean better performance
Predict ors
P1 ,P2 ,...
SMLR
ANN
SVM
Predict and
F
Training
Algorithm
F = f ( P1 ,P2 ,...)
Statistical Models:
Performance Factors
• Type
& configuration of the
statistical model and training
algorithm
• Size, quality and
representativeness of the
training sample
• Input variables made
available for training
• The type of relationships
that actually exist
Issue: difficult to understand the reasons for observed performance
Plant Output Models
• Relationship of met
variables to power
production
• Could be physical or
statistical
• Often based on wind
speed but can consider
other variables
Plant-scale Power Curve: 1 Year of Data
Hourly Data
100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
0
2
4
6
8
10 12 14 16 18 20 22 24 26 28 30
Met TowerWind Speed (m/s)
Desired Data from Wind Plant
Measured Power Output vs. Wind Speed
Desired Wind Plant Data for Forecasting
•
•
Power production and turbine availability
Met tower within 5km and 100m elevation of
each turbine
•
•
One met tower with two levels of data
•
•
•
Wind speed and direction at hub height and T
and P at 2m
T and P at 2m and hub height
Wind speed and direction at hub height and
hub height - 30m
Consistent monitoring and calibration of data
100%
80%
60%
Power Curve
Reported
40%
Well-Behaved Data
20%
0%
2
4
6
8
10
12
14
16
18
20
100%
80%
60%
Power Curve
Reported
40%
One of the keys to
accurate forecasts…
High quality data
from the plant
Data with Issues
20%
0%
2
4
6
8
10
12
14
16
Avg. Met Tower Wind Speed (m/s)
18
20
Forecast Ensembles
• Uncertainty present in any
forecast method due to
– Input data
– Model type and configuration
• Approach: perturb input data and model
parameters within their range of uncertainty
and produce a set of forecasts
• Benefits
– Ensemble composite is often the best forecast
– Spread of ensemble provides a case-specific measure of
uncertainty
Forecasting Time Scales
How does the wind power production
forecasting challenge vary with the
look-ahead period?
Hours-ahead Forecasts
• Must forecast small scale
weather features
– Large eddies, local-scale circulations
– Rapidly-changing, short life-times
– e.g. cloud features, mountain
circulations, sea/land breezes
• Typically poorly defined by
current observing systems
• Tools:
– Difficult to use physics-based models
– Autoregressive statistical models on wind farm time series data
– Supplement with offsite predictor data
• Errors grow rapidly with increasing look-ahead time
Days Ahead Forecasts
• Little skill in forecasting
small-scale features
• Forecast skill mostly
from medium and large
scale weather systems
• Well-defined by current
sensing networks
• Tools:
– Physics-based model simulations are the best tool
– Statistical models used to adjust physics-based output (MOS)
– Regional & continental scale weather data are most important
• Errors grow slowly with increasing look-ahead time
AWST’s
Alberta Forecast System
How is it configured for the Alberta application?
Why was it configured that way?
General Approach & Philosophy
• Apply AWST’s extensively used eWind system
– Forecast met variables with physics and statistical models
– Use plant output model to obtain power production forecast
• Employ sophisticated quality control of input data
• Configure and customize physics-based and
statistical models for optimum performance in Alberta
• Employ an ensemble forecasting scheme
– Ensemble members created by varying factors that are most
significant in producing uncertainty in forecasts in Alberta
• Input data
• Model parameters
– Construct an optimal composite forecast based on recent
performance of all ensemble members
Physics-based Simulations
Types
• Two physics-based models
– MASS: Mesoscale Atmospheric Simulation System
• Developed and maintained by MESO, Inc (AWST partner)
• Has a 25-year history of development and application
• Customized by AWST/MESO for wind energy forecasting in 1990’s
– WRF: Weather Research and Forecasting model
• New community model developed by US consortium including NCAR & NWS
• In widespread use for several years
• Four types of physics-based simulations
–
–
–
–
MASS with initial and boundary data from GEM (Canada)
MASS with initial and boundary data from NAM (US)
WRF with initial and boundary data from NAM (US)
WRF with initial and boundary data from GFS (US)
• Output used as input into statistical models
Physics-based Forecast Simulations
Configurations
• Nested grids over
Alberta and vicinity
• 10 km or less
horizontal grid cell size
• Initialized every 6 or 12
hrs depending on grid
point data availability
• Length of simulations:
66 or 72 hrs
• 3-D grid point output
data saved every hour
Short-term (0-6 hrs)
Met Variable Forecasting
• Ensemble of 12 different forecast methods
• Separate statistical forecast model for each look-ahead
hour for each forecast methods
• Ensemble based on varying several factors (emphasis on
statistical models and onsite and offsite data)
– Type of statistical algorithm (Linear regression, ANN etc.)
– Training sample size and time period
– Type and amount (# of variables) of input data
• Time series of met and power data
• Off-site met tower or remotely sensed data
• Physics-based model output data
• Statistical procedure used to construct an optimal
composite forecast from ensemble members based on
recent performance
Short-term forecasting (0-6 hrs)
Customization for Alberta
• Use offsite data types available in Alberta
• Analyze time series of data from Alberta wind
farms
– Determine autocorrelation structure
– Use most appropriate statistical procedures
• Identify significant offsite time-lagged spatial
relationships for each forecast site
– Analyze patterns and relationships in high-resolution
numerical simulations and observed data
Intermediate-term
Met Variable Forecasting (7-48 hrs)
• Ensemble of 16 different forecast methods
• Separate statistical forecast model for each lookahead hour for each forecast method
• Ensemble based on varying several factors
(emphasis on physics-based models)
–
–
–
–
Type of statistical algorithm
Training sample size and time period
Type of physics-based model
Source of physics-based model initialization and boundary data
• Statistical procedure used to construct an optimal
composite forecast from ensemble members based
on recent performance
Intermediate-term Forecasting
Customization for Alberta
• Use of Environment Canada’s models for
initialization and lateral boundary conditions
• Customize surface property databases for Alberta
– Standard databases often have misrepresentations
• Customize physics-based model to optimally
simulate phenomena important in Alberta
– Low level jets (reverse turbulence profile)
– Strong downslope winds (chinooks)
– Shallow cold air surges
• Configure model grids to have high grid resolution in
areas critical to wind farm wind variation
Plant Output Model
• Two models
– Plant-scale power curve
– Power curve deviation model
• Wind direction
• Atmospheric stability
• Why two models?
– Data quality and quantity
• Forecast obtained by using
ensemble composite of met
variable forecasts as input
Plant-scale Power Curve: 1 Year of Data
Hourly Data
100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
0
2
4
6
8
10 12 14 16 18 20 22 24 26 28 30
Met TowerWind Speed (m/s)
Plant Output Model Issue
Turbine Availability Reporting
• Reporting of actual
turbine availability has
been very inconsistent in
other applications
• Projected (scheduled)
availability often left at
100%
• Example depicts both
missing and probably
inaccurate actual
availability data. 100%
availability was specified
for all hours on the chart
Availability Data Problem Example
Reported PIRP Data - June 2006
Measured
100
90
80
70
60
50
40
30
20
10
0
0
5
10
15
20
Wind speed (m/s)
25
30
Forecasting the Future of Forecasting
What is being (can be) done to
improve forecast performance?
How will forecasts be improved?
(Top Three List)
• (3) Improved physics-based/statistical models
– Improved physics-based modeling of sub-grid and surface processes
– Better data assimilation techniques for physics-based models
– Learning theory advances: how to extract more relevant info from data
• (2) More effective use of models
–
–
–
–
Enabled by more computational power
Higher resolution, more frequent physics-based model runs
More sophisticated use of ensemble forecasting
Use of more advanced statistical models and training methods
• (1) More/better data
– Expanded availability and use of “off-site” data in the vicinity of wind
plants, especially from remote sensors
– A leap in quality/quantity of satellite-based sensor data
New Remote Sensing Technology
• Low-power, low-cost, dual-polarization phased array
Doppler radars on cellular towers being developed by
CASA (Center for Adaptive Sensing of the Atmosphere)
– Target price: $50 K per unit
– Commercial availability: 2009-2010
– Small enough to mount on cell towers;
– Will measure atmosphere below 1 km that is not visible to current National Weather
Service NEXRAD Doppler radar (72% of atmosphere below 1 km is not visible to
current NWS system)
– Provide winds with resolution in 100’s of meters out to 30-50 km
– New data every few minutes
• Attempt currently being made to organize a field project in
Tehachapi Pass in California to evaluate the value of this
technology to short-term wind energy forecasting
Summary
• AWST’s Approach to Forecasting: eWind
– Widely used and verified for wind power productions
forecasts in North America
– Based on an ensemble of forecasts from several physicsbased and statistical models using different datasets
– Extensive customization for forecasting in Alberta
• General Points About Forecasting
– Forecast quality has strong dependence on quality and
quantity of data from the wind generation facilities
– Forecast systems can and should be customized to meet the
requirements of a particular application
– Forecast technology is changing rapidly - need system/team
that can keep pace with the evolution of forecasting
technology
Fly UP