222 Probabilistic Ensemble Forecasting using Bayesian Model Averaging

Monday, 11 January 2016
Steven Beale, Amec Foster Wheeler Environment & Infrastructure, St. John's, NF, Canada; and T. Bullock

Probabilistic forecasts offer a number of advantages over traditional deterministic forecasts. The ability to provide confidence intervals, limits and risk levels allows forecasts to be generated that are of greater value to the end user. Growing interest in these forecasting methods is evident in the increasing number of publicly available ensemble forecast products (for example, NOAA's GEFS, EC's GEPS, FNMOC's Multimodel Ensemble and recently NCAR's WRF Ensemble). At Amec Foster Wheeler we are developing a probabilistic ensemble forecast framework using Bayesian Model Averaging. This framework allows us to combine any number of individual forecast models into a combined probability density function for an observable. This system includes forecast calibration (or debiasing) and probability tuning that responds to changes in model performance or weather predictability. The framework is versatile enough to be deployed for both simple observables (one dimension unbounded, eg. temperature) as well as complex observables (multidimensional bounded, eg. wind). This framework has been tested in multiple sites in Newfoundland & Labrador and Ontario and is being deployed as a part of the Amec Foster Wheeler operational forecasting system.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner