Version No. 1.0 Version Date 2/25/2008 Externally-authored document cover sheet Effective Date: 4/03/2008 The purpose of this cover sheet is to provide attribution and background information for documents posted to the California ISO website that were not authored by CAISO. Project Name Author Company Author Name Solar Symposium AWST Glenn VanKnowe Ph.D / John Zack Ph.D Author Title Title of document AWS Truewind s Approach to Solar Power Production Forecasting Date submitted 1/28/09 Other Comments Notes This document was submitted to the California Independent System Operator (CAISO) for posting to the CAISO website in conjunction with a Stakeholder-involved initiative or similar activity. The document was not produced by the CAISO and therefore does not necessarily reflect its views or opinion. Created by: IP&S Updated by: EA/ComPR/IPS/rls CAISO Page 1 of 1
Presented at the CAISO Solar Symposium Folsom, CA January 29, 2009 AWS Truewind s Approach to Solar Power Production Forecasting Glenn Van Knowe, Ph. D. and John W. Zack, Ph. D. AWS Truewind, LLC 185 Jordan Rd Troy, NY 12180 USA 2098 AWS Truewind, LLC
Overview The Nature of the Challenge AWST s Forecast Methods Forecast Performance Examples The Path to Improved Solar Forecasts
The Meteorological Challenge Objective: Predict the irradiance at the surface of the earth for a set of time intervals and look-ahead periods Obstacles: Variations in clouds and other atmospheric variables are driven by atmospheric features that originate, evolve and dissipate over a wide range of space and time scales under the control of a broad spectrum of physical processes. Current observational systems are able to measure only a small fraction of the variability associated with these features. Biggest challenge: Forecasting solar power ramps caused by rapidly forming and moving small-scale boundary layer clouds.
Solar vs Wind Forecasts Solar energy is generally more predictable than wind energy (especially in CA) Good solar sites have low cloudiness Forecast errors are small in absence of clouds Good wind sites have high winds => high variability Individual solar sites have more large (%) minutesscale ramps due to boundary layer clouds Satellites provide frequent, high resolution cloud data - the most significant factor affecting solar radiation variability - a big asset for short-term forecasts No equivalent data source available to provide high resolution spatial wind data which makes short-term wind forecasting very difficult
Methods Meeting the Challenge: AWST s esolar Forecast Systems Combination of physics-based (NWP) and statistical models Diverse set of input data with widely varying characteristics Importance of specific models and data types vary with look-ahead period Forecast products are generated from an ensemble of individual forecasts Input Data, Forecast Model Components and Data Flow for a State-of-the-Art Forecast System
Methods Physics-based Models also known as Numerical Weather Prediction (NWP) Models Differential equations for basic physical principles are solved on a 3-D grid Must specify initial values of all variables for each grid cell Simulates the evolution of the atmosphere over a 3-D volume Some forecast providers rely on government-run models; others run their own models Roles of Provider-run NWP Models Optimize model configuration and formulation for the forecasting of surface irradiance Use higher vertical or horizontal resolution over area of interest Execute simulations more frequently Incorporate data not used by government-run models Execute ensembles customized for surface irradiance forecasting
Methods 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 different model-generating methods available (linear regression, neural networks etc.) Roles of Statistical Models Correct systematic-errors in the NWP forecasts Account for processes on a scale smaller than the NWP grid cells (cloud advection?) Incorporate additional observational data received after the initialization of most recent NWP model runs not effectively included in NWP simulations Combine met forecasts and power data into power predictions (implicit plant output model)
Methods Solar Plant Output Models Relationship of met variables to power production for a specific Solar Power Generation Type Many possible formulations implicit or explicit statistical or physics-based single or multi-parameter Roles of Plant Output Models Facility-scale variations of production (within a solar plant) Handle plant type (e.g. PV versus concentration type) Operational factors (availability, plant performance variations etc)
Methods Forecast Ensembles Uncertainty present in any forecast method due to Input data Model type Model configuration Approach: perturb input data and model parameters within their range of uncertainty and produce a set of forecasts (i.e. an ensemble) Benefits An ensemble composite is often the best forecast Statistical weighting of individual forecasts often yields best performance Spread of ensemble provides a case-specific measure of forecast uncertainty
Time Scales How the Forecasting Problem Changes by Time Scale Minutes Ahead Clouds caused by BL plumes, turbulence, waves etc. Rapid and erratic evolution; very short lifetimes Mostly not observed by current sensor network Tools: Site trends, cloud image advection Difficult to beat diurnally adjusted persistence forecast Need: Very hi-res 3-D data from remote sensing Hours Ahead Clouds associated with mesoscale features Rapidly changing, short lifetimes Current sensors detect existence but not structure Tools: Site trends, cloud advection & NWP Outperforms persistence by a modest amount Need: Hi-res 3-D data from remote sensing Days Ahead Lows and Highs, frontal systems Slowly evolving, long lifetimes Well observed with current sensor network Tools: NWP with statistical adjustments Much better than a persistence or climatology forecast Need: More data from data sparse areas (e.g. oceans)
Time Scales PIRP Forecast Time Scales Hours-ahead Forecasts are the most critical for PIRP Approaches for this time scale Autoregressive trends from site data Pros: Simple and captures (& maybe conditional) climatological patterns Cons: Can t anticipate what is approaching the site at a particular time Cloud advection schemes Pros: Uses cloud image data (satellite or local imager) and estimate of winds at cloud levels to map current cloud patterns and anticipate movement towards sites Cons: Difficult to specify cloud levels from image data; does not account for development or dissipation Spatial-statistical cloud image models Pros: Uses cloud images and statistically accounts for advection, development & decay Cons: Huge data set and very complex models needed to adequately forecast the very large number of atmospheric cloud evolution patterns for a site Rapid Update Cycle (RUC) NWP simulations Pros: Explicitly models the physics of cloud movement, development and decay Cons: Large computational cost; difficult to adequately specify initial state
Products Forecast Products Deterministic Predictions Most likely MW production for a specific time interval (e.g. hourly) Tuned to minimize a performance metric (e.g. RMSE etc.) Often results in hedging for extreme event forecasts Probabilistic Predictions Confidence Bands Probability of Exceedance (POE) Values Event Forecasts Probability of events in specific time windows Most likely values of event parameters (amplitude, duration etc.) Example: large up or down ramps Situational Awareness Forecasts of significant weather regimes Produce events (large errors, ramps etc.) that impact user s applications Geographic displays of cloud patterns
Performance Example of an AWST Forecast Parameter: Global Irradiance Hours-ahead forecast for a site in Colorado Black line is the measured global irradiance Other lines are esolar deterministic forecasts produced at successive hours
Performance Example of Solar Forecast Performance Based on 4-hr forecasts and measurements of hourly average irradiance at a CA site for June 2006 Used to estimate errors for a 15 MW facility Hour Avg Insolation Ins Fcst MAE Ins Fcst MAE (%) Gen Fcst MAE PIRP GMC LDT (W/m2) (W/m2) % MWh US $ Note 1 Note 2 Note 3 Note 4 Note 5 6 34.98 10.78 30.81% 0.140 $0.20 7 435.25 81.07 18.63% 1.055 $1.53 8 707.57 59.16 8.36% 0.770 $1.12 9 826.63 40.01 4.84% 0.521 $0.75 10 884.67 34.89 3.94% 0.454 $0.66 11 924.14 32.23 3.49% 0.419 $0.61 12 946.46 28.31 2.99% 0.368 $0.53 13 941.67 37.72 4.01% 0.491 $0.71 14 935.63 41.68 4.46% 0.542 $0.79 15 904.98 66.04 7.30% 0.859 $1.25 16 866.86 76.74 8.85% 0.999 $1.45 17 796.43 121.03 15.20% 1.575 $2.28 18 727.78 93.68 12.87% 1.219 $1.77 19 580.75 88.27 15.20% 1.149 $1.67 20 273.21 65.33 23.91% 0.850 $1.23 Daily $16.55 Monthly $496.37 Average 2006 monthly GMC for a ~15 MW wind plant in PIRP was about $1,700 (assuming a $1.45 per MWh deviation charge - same as in the solar example) Note: assumes ideal plant performance
Challenges Meteorological Challenges Forecasting Boundary Clouds Cause of most difficult solar power ramp forecast problems Example to the right: Red: Clear sky Blue: With boundary layer clouds
Future 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 Use of regime-based Model Output Statistics (MOS) approach. (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 solar plants, especially from remote sensors Substantial potential to improve 0-6 hr forecasts A leap in quality/quantity of global satellite-based sensor data
Summary AWS Truewind employs an ensemble of physics-based (NWP) and statistical models that provide forecasts of meteorological parameters to a site-specific solar plant output model Many types of forecasts can be generated from the ensemble of individual forecasts (deterministic, probabilistic, event-oriented, situational awareness etc.) The importance of specific forecasting techniques and data types varies with the look-ahead time scale For PIRP, forecasts on the hours-ahead time scale will be critical AWST s fundamental approach to the PIRP time scale: Initially: Sites trends with background trends from standard NWP Ultimately: Cloud image and other remotely sensed data assimilated into rapid update cycle (hourly) 3-D NWP forecast simulations