Forecasting wind for the renewable energy market Matt Pocernich Research Applications Laboratory National Center for Atmospheric Research [email protected]
Jan 11, 2016
Forecasting wind for the renewable energy market
Matt PocernichResearch Applications Laboratory
National Center for Atmospheric [email protected]
Why is forecasting wind hard?
Turbulence
Inherently stochastic process
Issue of scales
Consequence – attempts at improving a deterministic forecasts have physical limitations.
04/21/23ENVR Workshop - October 2010
Results of Spectral Decomposition
(Rife, Davis, Liu 2004 MWR)
04/21/23ENVR Workshop - October 2010
Needs of the customer – a typical power curve
Pow
er
04/21/23ENVR Workshop - October 2010
OutlineComponents of a typical numerical weather
prediction system.
Ensembles Forecasting system
Methods of post processing
Verification of wind forecasts
Excitement to come
04/21/23ENVR Workshop - October 2010
Dynamic Integrated Forecast System - DICastTM
Performance
Ensemble Input + Dynamic Weighting + Bias Correction + Dynamic MOS = Optimized Forecast
04/21/23ENVR Workshop - October 2010
RTFDDA
Regional-scale NWP models WRF / MM5
MESONETs
GOES
Wind Prof
4-D Continuous Data Assimilation and Forecasts
Radars
Etc.
ACARS
Forecast
Cold start
tFDDA
Weather observations
WRF/MM5
Modified WRF/MM5:
Dx/Dt = ... + GW (xobs – xmodel )
where x = T, U, V, Q, P1, P2 …
W is weight function
All WMO/GTS
Farm Met
Yuabo Liu et al. Wind Energy Prediction - R & D Workshop. 11 -12 May 2010 © 2010 University Corporation for Atmospheric Research
04/21/23ENVR Workshop - October 2010
Assimilation of Wind Farm Data
Met Tower wind spd/dir
Turbine hubwind spd
Data QCand processing
Data combining and reformat
WRFRTFDD
A
All other weather Observations
Other met-towerweather Observations
Yuabo Liu et al. Wind Energy Prediction - R & D Workshop. 11 -12 May 2010 © 2010 University Corporation for Atmospheric Research
04/21/23ENVR Workshop - October 2010
22 March 2007Load Forecasting Workshop
NWP ForecastSolve the equations that describe
the evolution of the atmosphere
We cannot solve the equations analytically:
•Discretize them. Horizontally and vertically
•Close the equations by parameterizing small/fast physical processes.
22 March 2007Load Forecasting Workshop
Ensemble Forecasting – a very vague term
Random initial conditions
Multi-physics model
Multi-model (Poor man’s ensemble)
Time lag ensemble
04/21/23ENVR Workshop - October 2010
Win
d S
peed (
mp
h)
Win
d S
peed (
mp
h)
Forecast Hour
Multiple forecasts of the same event, designed to characterize uncertaintyObservational errorModel selection errorParameterization error
Run at many weather centers and forecasting companies
Forecast Hour04/21/23ENVR Workshop - October 2010
Challenges with ensembles
Tend to be under dispersive (not enough spread.)
Calibration for both reliability and sharpness.
Some methods includeensembleBMA (Chris Frayley + UW)quantile regression (Hopson)ensemble Kalman Filter (more later from Luca)
04/21/23ENVR Workshop - October 2010
Ensemble BMABias removal of each member using linear
regression.
Estimates weights and variance for each ensemble member which minimizes continuous rank probability score.
Essentially, dresses each ensemble member with a distribution.
Traditionally uses Gaussian distribution. For winds, use gamma.
Key work by Adrian Raftery, Tilman Gneiting, MacLean Sloughter and Chris Frayley (UW).
04/21/23ENVR Workshop - October 2010
Example of ensemble BMA forecasts
04/21/23ENVR Workshop - October 2010
Regime switching Algorithms(From M. Hering)
04/21/23ENVR Workshop - October 2010
04/21/23ENVR Workshop - October 2010
04/21/23ENVR Workshop - October 2010
04/21/23ENVR Workshop - October 2010
04/21/23ENVR Workshop - October 2010
04/21/23ENVR Workshop - October 2010
04/21/23ENVR Workshop - October 2010
Key Verification IssuesThe most common verification metrics are mean
absolute error, RMSE and Bias.
These do not address concerns like ramping events.
New statistical forecasts are created every 15 minutes with new physical model runs every 3 hours. We don’t have a developed concept or metrics for consistency.
Forecast value – cost/benefits is complicated. Value of weather forecast is used with load forecast. There are humans in the loop.
04/21/23ENVR Workshop - October 2010
Contingency table statistics
The most fundamental verification methods involve statistics derived from a contingency table. This requires forecasts and observations be categorized into discrete bins.
Basic contingency table statistics include hit rate, false positive rate, bias, false negative rate and percent correct.
Changes in power can be classified in such a way in the following manner. An increase (or decrease) in a forecast accompanied by a “similar”
increase (or decrease) in observed power is a good forecast. A change forecast in power, but not observed is a false positive. A change observed, but not forecast is a false negative A forecast of no-change, associated with no change is considered a
good, negative forecast.
The definition of a good forecast can be modified. Regions do not have to be defined by angular regions.
04/21/23ENVR Workshop - October 2010
Forecast vs. Observed Changes
04/21/23ENVR Workshop - October 2010
Agree in magnitude and direction
Disagree in magnitude and direction
Small values forecast and observed
False Positive and False Negative
Regions translated into a contingency table
04/21/23ENVR Workshop - October 2010
Changes in power forecast by short term forecast from in the first 3 hours
Percent Correct 42%
Gerrity Skill Score 0.2404/21/23ENVR Workshop - October 2010
0 hour lead time, 1- 3 hour duration
04/21/23ENVR Workshop - October 2010
GSS = 0.21PC = 48%
GSS = 0.21PC = 39%
GSS = 0.24PC = 42%
Criticisms of this approachFrom C. Ferro - U.Exeter
The classification of the observation as either neutral, down ramp or up ramp depends on the value of the forecast. That seems weird! It must lead to some difficulties in interpreting any analysis of the table.
Much easier to define categories using boundaries that are parallel to the observation and forecast axes.
My initial reaction to deal with this is not to use contingency tables at all but to model the continuous joint distribution.
04/21/23ENVR Workshop - October 2010
More data + better data = more fun
New instruments – LIDAR and SODAR
Better quality observations from existing stations.
Improvements in sharing data.
High quality networks of tall towers. (BPA)
04/21/23ENVR Workshop - October 2010
Space – Time processes?
04/21/23ENVR Workshop - October 2010
Concluding Remarks
04/21/23ENVR Workshop - October 2010