short term wind power short-term wind power forecasting
SHORT-TERM SHORT TERM WIND POWER FORECASTING: A REVIEW OF STATUS AND CHALLENGES
Eric Grimit and Kristin Larson 3TIER®, Inc. I Pacific Northwest Weather Workshop March 5 - 6, 2010
© 2010 3TIER, Inc.
Integrating Renewable Energy » Variable generation driven by the weather adds many challenges for meeting energy demand. › Managing the reserve electric capacity and transmission systems becomes more difficult. › Example: Wind power in the Bonneville Power Administration (BPA) energy balancing authority area (mostly in the Columbia Basin)
Wind power ramp down opposite of load ramp up © 2010 3TIER, Inc.
Short-Term Wind Power Forecasting » State of the practice: › Hours to Days Forecast Horizons » Numerical weather prediction with locally trained statistical post-processing (e.g., MOS)
› Minutes to Hours Forecast Horizons NW P
OnSite Data
OffSite Data
» Autoregressive statistical models and supervised machine learning techniques » Blending with short-term NWP model output » Adaptive predictor selection for large input data sets, including off-site meteorological observations » Regime Regime-switching switching models
› Trained to minimize bulk errors for average power over all forecast intervals (e.g. RMSE over 1-hour 1 hour periods for 1 yr).
© 2010 3TIER, Inc.
Large Set Predictor Selection » Input Data Sources: On-site private wind project, met tower, and turbine data Nearby public met tower data Nearby private wind project and met tower data ASOS and meso-network (MADIS) stations within 300km o NWP model column output
» Variables: Meteorological variables: Wind speed Wind direction Temperature Pressure o
Large # of NWP model output fields
Derived variables: Pressure differences Lags L up tto 24 h hours o
Time derivatives
© 2010 3TIER, Inc.
» Total Predictor Set: O(100,000) predictors, most useless l and/or d/ correlated l d Requires massive dimension reduction or memory-efficient sparse matrix methods: Dimension reduction: pick top few lagged predictors first for each measurement instrument separately o Memoryy efficient jjoint selection: use a LARS, Elastic Net approach
Regime Switching Models » Models applied to wind power forecasting › Wind direction regime switching space-time model » Gneiting, Larson, et al. (2006)
› “Trigonometric Direction Diurnal Model” Model » Hering and Genton (2009)
› Markov switching autoregressi e model regressive » Pinson and Madsen (2009)
© 2010 3TIER, Inc.
Evolution of regime marginal probabilities using a Markov Switching AutoRegressive (MSAR) statistical model for 10-minute, 1-step ahead prediction of wind power at Horns Rev (Denmark). (Pinson and Madsen, International Journal of Forecasting, 2009)
Standard Wind Power Forecast Skill » Location Wind Project “A” (Columbia Gorge, WA)
» Time Period and Forecast Details Training Period: 03-May-2007 – 17-Aug-2008 Test Period: 18-Aug-2008 – 18-Aug-2009 Forecast Interval: 60 minutes (at HH:00) Forecast Leads: 120, 110, …, 70 minutes
Wind Project “A” Exp. # (input data)
Train % Imp. RMSE
Test % Imp. RMSE
1 (on-site) (on site)
10.0
10.2
2 (on-site, off-site)
15.5
14.6
3 (on-site, off-site, ASOS)
23.2
16.4
RMSE skill score (% Imp.) measured relative to 1-hr persistence forecast
© 2010 3TIER, Inc.
Wind Power Ramp Forecasting
» Usual goal is to minimize large deviations – forecasts are optimized to be conservative » Yields smooth forecasts – under-prediction of ramps © 2010 3TIER, Inc.
An Alternative Metric – Event-Based Scoring » With categorical (yes/no) forecasts of a binary event, there are four possible outcomes outcomes. » Costs can be associated with each outcome.
© 2010 3TIER, Inc.
Yes
No
Yes
Hit
Miss
No
Ramp O Observed
Ramp Forecast
False Alarm
Correct Negative g
Example Wind Power Ramp Forecast Skill » Location
» Location
Wind Project “B” (Columbia Gorge, WA)
Wind Project “C” (Columbia Gorge, WA)
» Time Period and Forecast Details
» Time Period and Forecast Details
Test Period: 2006-2007 (6 months)
Test Period: 2007
Forecast Interval: 60 minutes (at HH:00)
Forecast Interval: 60 minutes (at HH:00)
Forecast Lead: 1 day
Forecast Lead: 1 hour Event Ramp Size: > 20% of capacity
FCST OBS
YES
NO
YES
281
1534
NO
637
5766
CSI (threat score) = Hits / ( Hits + Misses + Falses) = 0.11
© 2010 3TIER, Inc.
Challenges for Ramp Event Prediction and Validation
» Timing › Example: Day-ahead forecasts of wintertime frontal passage on the west coast can be off by > 6 hr
» Magnitude › Example: Small errors in offshoreonshore pressure gradients can result in poor land/sea l d/ b breeze iintensity t it
» Location › Example: Thunderstorm locations are difficult to forecast due to poor simulation of their initiation, development and decay processes
» Frequency q y › A ramp event is a rare event › At the project level, events of 20% or more occur less than 10% of the time › State-of-the-art forecasts have more false alarms and missed events than hits © 2010 3TIER, Inc.
Wind Power Ramp Event “Capture” Alberta Electric System Operator Wind Power Forecast Pilot Study (2007-08) Forecaster CSI (threat scores)
Alberta Region Forecaster
(Source: ORTECH Power, “Wind Power Forecasting Pilot Project Part B: The Quantitative Analysis Final Report”, 2008)
© 2010 3TIER, Inc.
A Non-Deterministic Approach Goal: Quantify all the Uncertainties
DAY AHEAD RAMP EVENT PREDICTION
Example Ramp Event Probability Forecast and Observation Time Series
2-month hourly time series (Jan-Feb 2007)
» System developed for BPA research project in 2007 » Utilizes NWP ensembles for day-ahead ramp event prediction » Includes statistical calibration (e.g., Bayesian model averaging) © 2010 3TIER, Inc.
Integrating the Ramp Forecast » Goals: › User-defined thresholds trigger alerts › Full probability distribution input into stochastic unit commitment and economic dispatch system › System operation can automatically optimize use of reserve capacity and avoid id ttransmission i i b bottlenecks ttl k › A proactive SmartGrid!
Risk of down ramp event above users’ th h ld threshold
© 2010 3TIER, Inc.
Challenges (Future Work) » Improved Ensemble NWP in Short-Range (0-6 hrs) › Assimilation of more on-line data
» Weather Regime Detection and Transition Prediction › Concepts p from low-frequency q y atmospheric p regime g transition studies
» Large-Set Predictor Selection › Dimension reduction › Mutual information criterion › Alternative metrics (asymmetric, event-based scores, etc.)
» Forecast Integration and Utilization › Packaging for energy management systems › Human factors © 2010 3TIER, Inc.
© 2010 3TIER, Inc.
Eric Grimit and Kristin Larson 3TIER®, Inc. I Pacific Northwest Weather Workshop March 5 - 6, 2010
© 2010 3TIER, Inc.
Integrating Renewable Energy » Variable generation driven by the weather adds many challenges for meeting energy demand. › Managing the reserve electric capacity and transmission systems becomes more difficult. › Example: Wind power in the Bonneville Power Administration (BPA) energy balancing authority area (mostly in the Columbia Basin)
Wind power ramp down opposite of load ramp up © 2010 3TIER, Inc.
Short-Term Wind Power Forecasting » State of the practice: › Hours to Days Forecast Horizons » Numerical weather prediction with locally trained statistical post-processing (e.g., MOS)
› Minutes to Hours Forecast Horizons NW P
OnSite Data
OffSite Data
» Autoregressive statistical models and supervised machine learning techniques » Blending with short-term NWP model output » Adaptive predictor selection for large input data sets, including off-site meteorological observations » Regime Regime-switching switching models
› Trained to minimize bulk errors for average power over all forecast intervals (e.g. RMSE over 1-hour 1 hour periods for 1 yr).
© 2010 3TIER, Inc.
Large Set Predictor Selection » Input Data Sources: On-site private wind project, met tower, and turbine data Nearby public met tower data Nearby private wind project and met tower data ASOS and meso-network (MADIS) stations within 300km o NWP model column output
» Variables: Meteorological variables: Wind speed Wind direction Temperature Pressure o
Large # of NWP model output fields
Derived variables: Pressure differences Lags L up tto 24 h hours o
Time derivatives
© 2010 3TIER, Inc.
» Total Predictor Set: O(100,000) predictors, most useless l and/or d/ correlated l d Requires massive dimension reduction or memory-efficient sparse matrix methods: Dimension reduction: pick top few lagged predictors first for each measurement instrument separately o Memoryy efficient jjoint selection: use a LARS, Elastic Net approach
Regime Switching Models » Models applied to wind power forecasting › Wind direction regime switching space-time model » Gneiting, Larson, et al. (2006)
› “Trigonometric Direction Diurnal Model” Model » Hering and Genton (2009)
› Markov switching autoregressi e model regressive » Pinson and Madsen (2009)
© 2010 3TIER, Inc.
Evolution of regime marginal probabilities using a Markov Switching AutoRegressive (MSAR) statistical model for 10-minute, 1-step ahead prediction of wind power at Horns Rev (Denmark). (Pinson and Madsen, International Journal of Forecasting, 2009)
Standard Wind Power Forecast Skill » Location Wind Project “A” (Columbia Gorge, WA)
» Time Period and Forecast Details Training Period: 03-May-2007 – 17-Aug-2008 Test Period: 18-Aug-2008 – 18-Aug-2009 Forecast Interval: 60 minutes (at HH:00) Forecast Leads: 120, 110, …, 70 minutes
Wind Project “A” Exp. # (input data)
Train % Imp. RMSE
Test % Imp. RMSE
1 (on-site) (on site)
10.0
10.2
2 (on-site, off-site)
15.5
14.6
3 (on-site, off-site, ASOS)
23.2
16.4
RMSE skill score (% Imp.) measured relative to 1-hr persistence forecast
© 2010 3TIER, Inc.
Wind Power Ramp Forecasting
» Usual goal is to minimize large deviations – forecasts are optimized to be conservative » Yields smooth forecasts – under-prediction of ramps © 2010 3TIER, Inc.
An Alternative Metric – Event-Based Scoring » With categorical (yes/no) forecasts of a binary event, there are four possible outcomes outcomes. » Costs can be associated with each outcome.
© 2010 3TIER, Inc.
Yes
No
Yes
Hit
Miss
No
Ramp O Observed
Ramp Forecast
False Alarm
Correct Negative g
Example Wind Power Ramp Forecast Skill » Location
» Location
Wind Project “B” (Columbia Gorge, WA)
Wind Project “C” (Columbia Gorge, WA)
» Time Period and Forecast Details
» Time Period and Forecast Details
Test Period: 2006-2007 (6 months)
Test Period: 2007
Forecast Interval: 60 minutes (at HH:00)
Forecast Interval: 60 minutes (at HH:00)
Forecast Lead: 1 day
Forecast Lead: 1 hour Event Ramp Size: > 20% of capacity
FCST OBS
YES
NO
YES
281
1534
NO
637
5766
CSI (threat score) = Hits / ( Hits + Misses + Falses) = 0.11
© 2010 3TIER, Inc.
Challenges for Ramp Event Prediction and Validation
» Timing › Example: Day-ahead forecasts of wintertime frontal passage on the west coast can be off by > 6 hr
» Magnitude › Example: Small errors in offshoreonshore pressure gradients can result in poor land/sea l d/ b breeze iintensity t it
» Location › Example: Thunderstorm locations are difficult to forecast due to poor simulation of their initiation, development and decay processes
» Frequency q y › A ramp event is a rare event › At the project level, events of 20% or more occur less than 10% of the time › State-of-the-art forecasts have more false alarms and missed events than hits © 2010 3TIER, Inc.
Wind Power Ramp Event “Capture” Alberta Electric System Operator Wind Power Forecast Pilot Study (2007-08) Forecaster CSI (threat scores)
Alberta Region Forecaster
(Source: ORTECH Power, “Wind Power Forecasting Pilot Project Part B: The Quantitative Analysis Final Report”, 2008)
© 2010 3TIER, Inc.
A Non-Deterministic Approach Goal: Quantify all the Uncertainties
DAY AHEAD RAMP EVENT PREDICTION
Example Ramp Event Probability Forecast and Observation Time Series
2-month hourly time series (Jan-Feb 2007)
» System developed for BPA research project in 2007 » Utilizes NWP ensembles for day-ahead ramp event prediction » Includes statistical calibration (e.g., Bayesian model averaging) © 2010 3TIER, Inc.
Integrating the Ramp Forecast » Goals: › User-defined thresholds trigger alerts › Full probability distribution input into stochastic unit commitment and economic dispatch system › System operation can automatically optimize use of reserve capacity and avoid id ttransmission i i b bottlenecks ttl k › A proactive SmartGrid!
Risk of down ramp event above users’ th h ld threshold
© 2010 3TIER, Inc.
Challenges (Future Work) » Improved Ensemble NWP in Short-Range (0-6 hrs) › Assimilation of more on-line data
» Weather Regime Detection and Transition Prediction › Concepts p from low-frequency q y atmospheric p regime g transition studies
» Large-Set Predictor Selection › Dimension reduction › Mutual information criterion › Alternative metrics (asymmetric, event-based scores, etc.)
» Forecast Integration and Utilization › Packaging for energy management systems › Human factors © 2010 3TIER, Inc.
© 2010 3TIER, Inc.
