The Application and Evaluation of Analog Ensemble Method to Day-Ahead Wind Power Production Forecasting in California

Accurate day-ahead forecasting of wind power production is an important factor in the integration of large amounts of wind power into the electric power supply. Day-ahead wind power forecasting is typically accomplished by applying statistical methods to correct systematic errors in Numerical Weather Prediction (NWP) model forecasts of wind speed at turbine level. Regardless of the statistical method applied, its skill depends on its ability to divide the sample of forecasts into relevant sub-samples with similar error patterns within the sample and different patterns between sub-samples.

The AE method selects a customized ensemble of historical outcomes based on the similarity between the current NWP forecast and historical NWP forecasts from the same model. The ensemble is unique for each forecast site and forecast look-ahead time and therefore provides a customized correction for each individual forecast value. The AE forecast ensemble can be assembled from either the observed outcomes from the historical cases (observed outcome method) or the model forecast errors from the historical cases applied to correct the current model forecast (model error method).

This study applies Analog Ensemble (AE) to forecasting the hourly day-ahead wind power production at 55 wind farms within the control area of the California Integrated System Operator (CAISO). The forecasts were generated over an historical sample using data that would have been available at 0300 Pacific Prevailing Time. In this case, all input data were extracted from the 0000 UTC run of 4 different NWP simulations run by AWS Truepower. Two simulations use the Mesoscale Atmospheric Simulation System (MASS) while the other two use the Weather Research and Forecasting (WRF) model. The model runs are based on first guess and boundary condition data from the 0000 UTC National Center for Environmental Prediction (NCEP) of either the North American Mesoscale (NAM) model or the Global Forecast System (GFS). The similarity between the current and historical NWP forecasts was computed using an equal weighting of 80-m U and V wind components and 50-m turbulent kinetic energy (TKE) extracted at each forecast site. Deterministic forecasts were created by averaging the ensemble members, giving more weight to the members that better matched the current forecast.

The conference presentation will provide a brief overview of the AE method and show the dependence of AE forecast performance on (1) choice of observed outcome or model error method, and (2) seasonal windows for ensemble member selection. It will also explore the impact of AE methods in reducing the error of an optimized ensemble of raw model and linear regression based model error correction methods.