Ensemble-derived estimation of wind uncertainty using a linear variance calibration for probabilistic weather applications

This study examines the calculation of uncertainty in the meteorological data as derived from an ensemble, and its effects when used as additional input to a probabilistic weather application such as an atmospheric transport and dispersion (AT&D) model. In the event of the release of a dangerous atmospheric contaminant, an AT&D model is often used to provide forecasts of the resulting contaminant dispersion affecting the population. These forecasts should also be accompanied by accurate estimates of the forecast uncertainty to allow more informed decisions about the potential hazardous area.

The Second-Order Closure Integrated Puff (SCIPUFF) AT&D model is used to demonstrate that the use of wind variances from a meteorological (MET) model ensemble can be used to represent the MET uncertainty effects on an AT&D prediction. Furthermore, these wind variances, which are readily available in existing MET ensemble products, should be calibrated to represent more accurately the actual MET uncertainty (variance). This study applies a Linear Variance Calibration (LVC) method to NCEP Short-Range Ensemble Forecast (SREF) predictions to evaluate the impact of a calibration of its derived wind variances on the resulting SCIPUFF predictions. The SCIPUFF forecasts are then compared to a baseline “truth” SCIPUFF simulation driven by the MET from a high-resolution (4-km) dynamic analysis.

Preliminary results have shown that calibration of MET ensemble wind variance has a positive impact on SCIPUFF mean concentration and hazard-area predictions. The LVC model parameters (slope and y-intercept) are further investigated here through the use of an ideal stochastic model. Examination of the ideal stochastic model results indicates that for ensemble sizes less than several hundred members, sampling error causes the ensemble to under-represent the actual ensemble variance. This introduces an anomalously larger y-intercept in the LVC calculations and a corresponding smaller slope. The size of these anomalies is dependent on the variance of actual errors, with larger error variances resulting in larger anomalies. It is shown that with large sample size, these anomalies in the LVC parameters approach zero.