Haupt Solar Forecasting Final Presentation
Counting on Solar Production: Advances in Forecasting
Sue Ellen Haupt National Center for Atmospheric Research
SEPA Utility Solar Conference
San Diego
April 28, 2015
Why do we need to forecast Solar Power?
What happens to Solar Power when there are Clouds?
A Public-Private-Academic Partnership for Solar Power Forecasting
Value Chain: What is the value of solar power forecasting?
Clear Sky
Weather
Monitoring Observation
SURFRAD Clouds
Aerosols
Satellites Total Sky Imagers
Modelling
Forecasting
Dissemination & Communication
Power Conversion
WRF-Solar HRRR StatCast
DICast Point Forecast
TSICast
Pyranometers
NowCast CIRACast MADCast
Area Forecast Uncertainty Quant
Perception Interpretation
Projected Power Production Actual Power Production Load Balancing
Uses / Decision Making
Day Ahead Planning Real Time Operation
Outcomes
Unit Allocations
Reserve Estimates
Economic & social values
Production Cost Changes Reserve Analysis
Seamless Scaled Approach to Solar Power Forecasting
48
SunCast
Nowcasting - Overview 1. StatCast – regimes and data
3. CIRACast – Satellite-based Cloud
Advection
4. MADCast Multi-sensor 2. TSICast - Total Sky Imaging Advective Diffusive WRF Nowcasting
5. WRF-SolarNowcasting
Numerical Weather Prediction: WRF-Solar
WRF-Solar: Cloud-radiation-aerosol Interaction
WRF-Solar versus Standard WRF
Uncertainty Quantification Analog Ensemble Approach
300
500
5- 95% 25-75% An-En Mean
0 100
GHI (Wm−2 )
700
900
Station SMUD 67, forecast initialized at 12 UTC, 15 July 2014
0 4 8 13
19
25
31
37
43
49
Lead Time (Hour)
55
61
67
Engineering the System
Quantify Value - Metrics
Enhanced
Base
Model-Model Comparison
Economic Value
• • • •
Mean Absolute Error Root Mean Square Error Distribution (Statistical Moments and Quantiles) Categorical Statistics for Events
• •
Operating Reserves Analysis Production Cost
• • • • • • • • • • • • •
Maximum Absolute Error Pearson's Correlation Coefficient Kolmogorov-Smirnov Integral Statistical Tests for Mean and Variance OVER Metric Renyi Entropy Brier Score incl. decomposition for probability forecasts Receiver Operating Characteristic (ROC) Curve Calibration Diagram Probability Interval Evaluation Frequency of Superior Performance Performance Diagram for Events Taylor Diagram for Errors
•
Cost of Ramp Forecasting
Quantify Value - Metrics GEM
1
GFS Init Time
2
StatCast CIRACast MADCast WRF-SolarNow
PRELIMINARY RESULTS Each component has a “sweet spot” when it can contribute skill to nowcast. It is now a matter of building this information into the NowCast integrator
3 4
HRRR Legend Rank PRELIMINARY RESULTS • GEM strongest component • HRRR provides good skill at short lead times • GFS and NAM provide fair to good skill at longer lead times
Operationalization
SMUD – 100 + 50 MW
LIPA – 32 MW SCE – 350 Comm + 325Q + 1000 Dist MW
HECO– 43 MW
Xcel – 90 MW
DeSoto Plant – 25 MW
Software Dissemination • WRF-Solar – New radiation scheme – New cloud physics parameterization – Improved GODDARD parameterization for equation of time – High frequency output – Fast radiation scheme (NREL) – Shallow convection scheme (PSU) – Satellite data assimilation – I/O Parallelization documentation and scripts – Climatological Aerosol information Black – already released Brown – will release
• StatCast • Power Conversion scripts & software • MADCast • MET enhancements
Distributed Solar Forecasting Determine load cutout due to DG solar (higher load on cloudy days) Obtaine historical production data (gridded or point) and determine representativeness
UCAR Confidential and Proprietary. © 2014, University Corporation for DPV installations AtmosphericCapacity Research. Allofrights reserved.
Season Spring Summer Fall Winter
Higher load days
p-value
Sample size (clear/cloudy)
Cloudy Clear Cloudy Cloudy
0.032 0.0003 0.638 0.016
41/17 28/8 30/24 36/21
Correlation of hourly gridded PV percent capacity to those from a grid box near the NREL-WTC site. Also shown are BVSD PV (blue dots) and Sun Edison PV installations (green dots), and METAR sites (black dots).
Distributed Solar Forecasting Built system to forecast solar power and upscale to CO, plus provide info for feeders Solar forecast to impact load forecast – will allow to grow with increased deployment
5 fold cross validation shows a Mean Absolute Error of just over 0.05 magnitude in % PV capacity UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Distributed Solar Forecasting NVISTA 20141105
Evaluation Difficult to evaluate given lack of actual production data (behind the meter) Used data from school district – had some QC problems
NVISTA 20141102
Normalized RMSE
Scatter plot shows majority of forecasts along Y=X line
Gridded Atmospheric Forecasts:
GRAFS-Solar Observations
NWP Models Current: NAM Future: GFS WRF-Solar GEM RAP/HRRR
Initial Grid Interpolated to 4 km CONUS Grid 1-Hour Averaging Archive data near observation sites
Current: SMUD Future: MADIS OK Mesonet BNL SURFRAD Xcel DeSota ARM
Statistical Correction Current: DICast Point Correction Future: Gradient Boosted Regression Trees Cubist Support Vector Machines Analog Ensemble
Output Products Maps of solar irradiance Single point forecasts % of clear sky irradiance Future: Other met. variables
GRAFS Results at SMUD Sites
GRAFS
• •
A new forecast is generated every hour Individual images are generated for each lead-time – Currently hourly out to 60 hours.
Comments: • Solar Power Forecasting is advancing rapidly • Start from Value to Stakeholders • Plan for evaluation • Much of the advances will be widely available • GRAFS is a new community gridded forecasting system being developed • Partners continue to be added • Final outcome to advance solar energy through better, more economical grid integration
Questions?
Sue Ellen Haupt National Center for Atmospheric Research
SEPA Utility Solar Conference
San Diego
April 28, 2015
Why do we need to forecast Solar Power?
What happens to Solar Power when there are Clouds?
A Public-Private-Academic Partnership for Solar Power Forecasting
Value Chain: What is the value of solar power forecasting?
Clear Sky
Weather
Monitoring Observation
SURFRAD Clouds
Aerosols
Satellites Total Sky Imagers
Modelling
Forecasting
Dissemination & Communication
Power Conversion
WRF-Solar HRRR StatCast
DICast Point Forecast
TSICast
Pyranometers
NowCast CIRACast MADCast
Area Forecast Uncertainty Quant
Perception Interpretation
Projected Power Production Actual Power Production Load Balancing
Uses / Decision Making
Day Ahead Planning Real Time Operation
Outcomes
Unit Allocations
Reserve Estimates
Economic & social values
Production Cost Changes Reserve Analysis
Seamless Scaled Approach to Solar Power Forecasting
48
SunCast
Nowcasting - Overview 1. StatCast – regimes and data
3. CIRACast – Satellite-based Cloud
Advection
4. MADCast Multi-sensor 2. TSICast - Total Sky Imaging Advective Diffusive WRF Nowcasting
5. WRF-SolarNowcasting
Numerical Weather Prediction: WRF-Solar
WRF-Solar: Cloud-radiation-aerosol Interaction
WRF-Solar versus Standard WRF
Uncertainty Quantification Analog Ensemble Approach
300
500
5- 95% 25-75% An-En Mean
0 100
GHI (Wm−2 )
700
900
Station SMUD 67, forecast initialized at 12 UTC, 15 July 2014
0 4 8 13
19
25
31
37
43
49
Lead Time (Hour)
55
61
67
Engineering the System
Quantify Value - Metrics
Enhanced
Base
Model-Model Comparison
Economic Value
• • • •
Mean Absolute Error Root Mean Square Error Distribution (Statistical Moments and Quantiles) Categorical Statistics for Events
• •
Operating Reserves Analysis Production Cost
• • • • • • • • • • • • •
Maximum Absolute Error Pearson's Correlation Coefficient Kolmogorov-Smirnov Integral Statistical Tests for Mean and Variance OVER Metric Renyi Entropy Brier Score incl. decomposition for probability forecasts Receiver Operating Characteristic (ROC) Curve Calibration Diagram Probability Interval Evaluation Frequency of Superior Performance Performance Diagram for Events Taylor Diagram for Errors
•
Cost of Ramp Forecasting
Quantify Value - Metrics GEM
1
GFS Init Time
2
StatCast CIRACast MADCast WRF-SolarNow
PRELIMINARY RESULTS Each component has a “sweet spot” when it can contribute skill to nowcast. It is now a matter of building this information into the NowCast integrator
3 4
HRRR Legend Rank PRELIMINARY RESULTS • GEM strongest component • HRRR provides good skill at short lead times • GFS and NAM provide fair to good skill at longer lead times
Operationalization
SMUD – 100 + 50 MW
LIPA – 32 MW SCE – 350 Comm + 325Q + 1000 Dist MW
HECO– 43 MW
Xcel – 90 MW
DeSoto Plant – 25 MW
Software Dissemination • WRF-Solar – New radiation scheme – New cloud physics parameterization – Improved GODDARD parameterization for equation of time – High frequency output – Fast radiation scheme (NREL) – Shallow convection scheme (PSU) – Satellite data assimilation – I/O Parallelization documentation and scripts – Climatological Aerosol information Black – already released Brown – will release
• StatCast • Power Conversion scripts & software • MADCast • MET enhancements
Distributed Solar Forecasting Determine load cutout due to DG solar (higher load on cloudy days) Obtaine historical production data (gridded or point) and determine representativeness
UCAR Confidential and Proprietary. © 2014, University Corporation for DPV installations AtmosphericCapacity Research. Allofrights reserved.
Season Spring Summer Fall Winter
Higher load days
p-value
Sample size (clear/cloudy)
Cloudy Clear Cloudy Cloudy
0.032 0.0003 0.638 0.016
41/17 28/8 30/24 36/21
Correlation of hourly gridded PV percent capacity to those from a grid box near the NREL-WTC site. Also shown are BVSD PV (blue dots) and Sun Edison PV installations (green dots), and METAR sites (black dots).
Distributed Solar Forecasting Built system to forecast solar power and upscale to CO, plus provide info for feeders Solar forecast to impact load forecast – will allow to grow with increased deployment
5 fold cross validation shows a Mean Absolute Error of just over 0.05 magnitude in % PV capacity UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Distributed Solar Forecasting NVISTA 20141105
Evaluation Difficult to evaluate given lack of actual production data (behind the meter) Used data from school district – had some QC problems
NVISTA 20141102
Normalized RMSE
Scatter plot shows majority of forecasts along Y=X line
Gridded Atmospheric Forecasts:
GRAFS-Solar Observations
NWP Models Current: NAM Future: GFS WRF-Solar GEM RAP/HRRR
Initial Grid Interpolated to 4 km CONUS Grid 1-Hour Averaging Archive data near observation sites
Current: SMUD Future: MADIS OK Mesonet BNL SURFRAD Xcel DeSota ARM
Statistical Correction Current: DICast Point Correction Future: Gradient Boosted Regression Trees Cubist Support Vector Machines Analog Ensemble
Output Products Maps of solar irradiance Single point forecasts % of clear sky irradiance Future: Other met. variables
GRAFS Results at SMUD Sites
GRAFS
• •
A new forecast is generated every hour Individual images are generated for each lead-time – Currently hourly out to 60 hours.
Comments: • Solar Power Forecasting is advancing rapidly • Start from Value to Stakeholders • Plan for evaluation • Much of the advances will be widely available • GRAFS is a new community gridded forecasting system being developed • Partners continue to be added • Final outcome to advance solar energy through better, more economical grid integration
Questions?
