The Texas component of the WFIP was a collaboration among the DOE, NOAA and a 7-member private sector team managed by AWS Truepower, LLC. MESO, Inc was a member of the team and played a key role in the forecast system experiments. During the one-year field experiment from August 2011 to September 2012, an array of sensors was deployed at targeted locations by NOAA, DOE, and the private sector team members to collect additional atmospheric data within the experimental area. The data from those sensors were assimilated into a set of advanced prediction systems to generate real-time 0- to 6-hour ahead wind power forecasts for individual wind farms and regional aggregates of wind farms on the electrical grid system operated by the Electric Reliability Council of Texas (ERCOT).
The additional data from the project sensors along with the large array of standard meteorological data gathered routinely over the region were used as input into an ensemble of physics-based and statistical forecasting methods that generated a set of deterministic and probabilistic forecasts of (1) average wind power production over 15-minute and 1-hour intervals for the 0- to 6-hour look-ahead period and (2) the occurrence of large power production change events (known as ramps) during the same 6-hour forecast window.
The physics-based component of the experimental WFIP-South forecast system was composed of an ensemble of twelve Numerical Weather Prediction (NWP) systems, which were executed by members of the project team. One member was the High Resolution Rapid Refresh (HRRR) model executed hourly on a 3-km grid by NOAA's Earth System Research Laboratory. A second member was the Advanced Regional Prediction System executed every 6 hours on a 2-km grid by the University of Oklahoma. The other ten members were executed by MESO, Inc. and consisted of ARPS, the Weather Research and Forecast (WRF) model, and Mesoscale Atmospheric Simulation System on a 5-km grid run every two hours using different combinations of initialization datasets, data assimilation procedures, and model physics. The data from additional sensors deployed for this project as well as the data from a set of participating wind farms within Texas were assimilated into most of the ensemble members. However, the data from the project sensors were withheld from some ensemble members to gauge their impact on the forecasts.
A Model Output Statistics (MOS) procedure was applied to the forecasts from each NWP system. The MOS is designed to correct systematic errors of relevant NWP meteorological variables (e.g. wind speed and direction) at forecast sites (i.e. the wind farms). Several MOS strategies based on variations of sample selection strategies and statistical prediction tools were used to generate an ensemble of statistical predictions from each NWP system. The MOS output for the individual NWP systems was then used as input to an Optimized Ensemble Model (OEM), which created a composite deterministic or probabilistic forecast from the set of MOS-adjusted NWP forecasts. In addition to the NWP forecasts, statistical predictions based purely on recent observational data were also included in the ensemble. Two OEM training strategies were tested. One was based on a rolling sample of the last 30 days. A second approach was based on a customized analog training sample that was constructed by matching key weather parameters for the current forecast period with those for cases in a historical archive. The objective of the regime-based approach was to weight the individual members of the ensemble according their performance in weather patterns that were similar to the one expected during the forecast period.
An extensive analysis of forecast performance was performed. This analysis included an evaluation of each physics-based and statistical component of the forecast system as well as the performance of final ensemble composite forecasts. It also included an evaluation of the variations in forecast performance by time of year, time of day, weather regime, and other factors.
The presentation will include (1) an overview of the physics-based and statistical modeling systems, type of forecasts produced and methods employed to evaluate the forecasts, (2) 0- to 6-hour deterministic wind power production forecast performance from the components of the modeling system and the forecast system composite, as well as performance by the time of year, time of day, and weather regime, (3) 0- to 6-hour ahead probabilistic and deterministic wind power ramp rate forecast performance as a function of ramp event type, and (4) forecast performance sensitivity to amount and type of data assimilated by the NWP models and size of the forecast ensemble.