7.7 Improving Forecasts for Locally Extreme Rainfall: A Probabilistic Approach

Tuesday, 12 January 2016: 4:30 PM
Room 226/227 ( New Orleans Ernest N. Morial Convention Center)
Gregory R. Herman, Colorado State University, Fort Collins, CO; and R. S. Schumacher

Methods are explored for developing a quality Probabilistic Forecast System (PFS) tailored for prediction of 6-hour and 24-hour extreme rainfall events, with event return periods ranging between 1 and 100 years, over the contiguous United States (CONUS). Output from a plethora of numerical models, including the Global Ensemble Forecast System (GEFS), National Severe Storms Laboratory's Weather Research and Forecasting Model (NSSL-WRF), Colorado State University WRF (CSU-WRF), High Resolution Rapid Refresh (HRRR), operational Global Forecast System (GFS) and North American Mesoscale (NAM) models, Short Range Ensemble Forecast (SREF) system, operational ARW and NMM High-Resolution Windows (HIRESW-ARW, HIRESW-NMM), and 4km NAM nest (NAM-NEST) are examined as input to the PFS. Techniques for generating weights for different NWP models based on individual model skill verification over the statistical model training period are assessed. Additionally, ensemble methods for generating forecast exceedance probabilities from individual member forecasts are explored, including traditional democratic voting, uniform ranks, and neighborhood-based approaches. More sophisticated machine learning algorithms, such as logistic regression, support vector machines, and random forests, are also employed to calibrate PFS probabilities using ensemble member precipitation forecasts in addition to other pertinent atmospheric fields when available. All statistical model configurations are assessed using cross-validation over the approximately 5-year period from June 2009 through August 2014. PFS performance is quantified by a combination of forecast reliability, Fractions Skill Scores (FSS), and Value Score (VS). From the perspective of each of skill, value, and reliability, preliminary findings from this work indicate that more advanced techniques for probability generation yield greatly improved forecasts relative to more basic methods which are commonly employed.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner