Calibration of Probabilistic Forecasts: The Bayesian Processor of Ensemble

One of the key objectives of the program aimed at the development of the Natinoal Oceanic and Atmosphereic Administration's (NOAA) Next Generation Global Prediction System (NGGPS) is the production of calibrated and skillful probabilistic forecasts. NGGPS supports a collaborative project with federal research laboratory (the Global Systems Division of NOAA Research), operational (Meteorological Development Laboratory of NOAA's National Weather Service), and academic (University of Virginia) participation to move the Bayesian Processor of Ensemble (Krzysztofowicz and Toth, 2008; Krzysztofowicz 2010) system towards NOAA operations.

The BPE will fuse predictive information from multiple sources (e.g., ensemble and higher resolution unperturbed forecasts from multiple centers) with a climatological prior distribution to provide well calibrated and informative probabilistic forecasts in the form of quantiles, probability density and cumulative distribution functions, and calibrated ensemble members. Due to its use of a rich set of distribution functions and a meta-Gaussian formulation, all continuous variables can be calibrated with BPE.

This presentation will describe the development of BPE algorithms and software and their testing in an operationally relevant environment. The performance of BPE will be compared with two systems used operationally at the National Weather Service (NWS): the Ensemble Kernel Density Model Output Statistics (EKDMOS) technique developed for observation sites by the Meteorological Development Laboratory (MDL), and the North American Ensemble Forecast System (NAEFS) statistical post-processing technique developed for applications on model grids by the Environmental Modeling Center (EMC) of the National Centers for Environmental Prediction (NCEP) and its Canadian Meteorological Center (CMC) collaborators. Using common training datasets and other constraints, the performance of the systems will be compared in terms of statistical reliability (calibration), statistical resolution (informativeness), and computational efficiency.