Sensitivity of Short-term Wind Ramp Forecast Error in Complex Terrain to NWP Model Configuration

One of the most challenging aspects of forecasting short-term wind power generation is the accurate prediction of the timing and amplitude of ramp events (large, rapid changes in wind power production). Numerical Weather Prediction (NWP) models have been shown to routinely miss the timing of ramps by 1-to-3 hours as well as significantly overpredict the wind speed at hub height, especially in the presence of complex terrain. In order to address these issues, this study aims to diagnose the sensitivity of ramp event timing and amplitude to the physics-based configuration options, resolution and initialization data by running a series of Weather Research and Forecasting (WRF) model simulations for the Tehachapi Pass Wind Resource Area (TWRA) in California, which is the area of largest installed wind capacity and greatest potential for further growth in the state. The primary objectives of this study are to (1) identify the model attribute (physics, resolution, etc.) to which the wind ramp forecast is most sensitive in different wind regimes, (2) estimate the relative forecast uncertainty associated with each key model attribute, and (3) determine an optimal model configuration for a 1-year forecasting experiment that will use input from a targeted sensor network in Tehachapi Pass.

Three primary TWRA flow regimes were studied – diurnal cycles (May - July), monsoonal flows (Aug - Sep) and mid-latitude events (Dec - Feb). Five up ramps and five down ramps were chosen for each of the three flow regimes for a total of 30 cases from Aug 2014 – June 2015. Cases were chosen based on three categories - the 60 minute ramp rate (the change in power generation per minute), the structure of the ramp, and the availability of observational data. Cases with small ramp rates, a noisy temporal evolution or a lack of data were excluded from selection. Historical forecasts were generated using a baseline configuration of WRF initialized with data from the Rapid Refresh (RAP) model. These forecasts were initialized 6 hours prior to the observed ramp and run out to 15 hours. Once the baseline simulations were completed, the sensitivity of forecast error was investigated for model physics, specifically the boundary layer, land surface, radiation and water phase-change physics submodels using different schemes in the WRF model. Forecast error sensitivity was assessed by varying only one submodel or submodel component at a time. Additional sensitivity experiments were also conducted for the model grid spacing and topographic resolution as well.

Error statistics for the forecasts were calculated for the look-ahead time in hours after forecast issue time in 15-minute intervals. Verification values for forecast bias, mean absolute error (MAE) and root mean square error (RMSE) were reported in percent of installed capacity for 17 operational wind farms located in the TWRA. In addition to 15-minute power production error statistics, we also analyzed ramp timing, amplitude, and duration errors. An analysis of the spatio-temporal evolution of the modeled vs. observed events was also performed for each case to determine which aspects of the evolution are replicated by the model and which are not.

The conference presentation will summarize the results for a group of representative case studies, including error statistics for individual model attributes and a suggested optimal WRF configuration. This configuration will then be used as a baseline for a 1-year real-time forecasting experiment in the TWRA.