Model Post-Processing for the Extremes: Improving Forecasts of Locally Extreme Rainfall
November 16, 2015
Hosted by Russ Schumacher (advisor), Elizabeth Barnes, Daniel Cooley (Statistics)
This study investigates the science of forecasting locally extreme precipitation events over the contiguous United States from a fixed-frequency perspective, as opposed to the traditionally applied fixed-quantity forecasting perspective. Frequencies are expressed in return periods, or recurrence intervals; return periods between 1-year and 100-years are analyzed for this study. Many different precipitation accumulation intervals may be considered in this perspective; this research chooses to focus on 6- and 24-hour precipitation accumulations. The research presented herein discusses the beginnings of a comprehensive forecast system to probabilistically predict extreme precipitation events using a vast suite of dynamical numerical weather prediction model guidance.
First, a recent climatology of extreme precipitation events is generated using the aforementioned fixed-frequency framework. The climatology created generally conforms with previous extreme precipitation climatologies over the US, with predominantly warm season events east of the continental divide, especially to the north away from major bodies of water, and primarily cool-season events along the Pacific coast. The performance of several operational and quasi-operational models of varying dynamical cores and model resolutions are assessed with respect to their extreme precipitation characteristics; different biases are observed in different modeling systems, with one model dramatically overestimating extreme precipitation occurrences across the entire US, while another coarser model fails to produce the vast majority of the rarest (50-100+ year) events, especially to the east of the Rockies where most extreme precipitation events are found to be convective in nature. Some models with a longer available record of model data are employed to develop model-specific quantitative precipitation climatologies by parametrically fitting extreme value distributions to model precipitation data, and applying these fitted climatologies for extreme precipitation forecasting. Lastly, guidance from numerous models is examined and used to generate probabilistic forecasts for locally extreme rainfall events. Numerous methods, from the simple to the complex, are explored for generating forecast probabilities; it is found that more sophisticated methods of generating forecast probabilities from an ensemble of models can significantly improve forecast quality in every metric examined when compared with the most traditional probabilistic forecasting approach. The research concludes with the application of the forecast system to a recent extreme rainfall outbreak which impacted several regions of the United States.