All Climate Events
Webinar: Improved Statistical Postprocessing of Precipitation for the NWS Forecasts and Hydrologic Applications
Monday, 26 March 2018, 1:30
Monday, March 26, 2018. 1:30PM. Webinar: Improved Statistical Postprocessing of Precipitation for the NWS Forecasts and Hydrologic Applications. Thomas M. Hamill, NOAA Earth System Research Lab. Sponsored by Center for Satellite Research and Applications. More information here.
Accurate numerical forecasts of precipitation have many purposes, including providing forcings to hydrologic models used to predict river flow and Great Lakes water levels. Commonly, the raw model guidance may exhibit both random and systematic errors, such as an overabundance of light precipitation and not enough heavy precipitation. If the raw guidance is used to provide precipitation forcings to hydrologic predictions, the accuracy of those predictions will suffer. A potential solution is the statistical postprocessing of the numerical guidance, using past forecasts and observations (or analyses) to statistically adjust the current model guidance. Statistical postprocessing of precipitation amount is quite challenging; to have an adequate training sample of unusual events such as heavy precipitation, either a long-time series of past forecasts and analyses are needed (i.e., reforecasts) or advanced statistical methods to leverage shorter training data sets.
In this seminar [she] will describe an advanced precipitation postprocessing algorithm in development that is intended for operational use in the NWS roughly a year hence. Each of the ensemble predictions are statistically postprocessed to adjust for biases in the mean amount and for excessive confidence. Short training data sets (the last 60 days of forecasts and analyses) are used. Biases in the mean amount are addressed through "quantile mapping" using cumulative distributions functions (CDFs) of analyzed and forecast precipitation. To address the small training sample size, the training sample used to populate CDFs at a particular location are supplemented by data from other locations with similar precipitation climatologies and terrain characteristics. The final probabilities are determined from a weighted linear combination of "dressed" ensemble members.
The resultant probabilistic forecasts are greatly improved, as will be demonstrated with verification statistics and some case studies. After presentation of results, [she] welcome discussion about the potential relevance of this algorithm for forcing hydrologic predictions of the Great Lakes.