7.4 Gridded Bayesian Linear Regression to Improve Storm Forecasts Using NCAR’s Real-Time Prediction System for Northeast United States

Wednesday, 10 January 2018: 11:30 AM
Room 19AB (ACC) (Austin, Texas)
Jaemo Yang, Univ. of Connecticut, Storrs, CT; and M. Astitha and C. S. Schwartz

This study presents the development and application of gridded Bayesian linear regression (GBLR) as a new post-processing technique to improve numerical weather prediction (NWP) of rain and wind storm forecasts over the northeast United States. In this approach, raw NWP model forecasts are corrected using regression coefficients estimated from modeled-observed pairs of training storms that minimize the systematic and random error of predicted atmospheric variables (wind speed, precipitation, etc.) for meteorological stations of the National Weather Service, and then interpolated back to the model domain. The GBLR model with ten predictors is developed for ten ensemble members of the National Center for Atmospheric Research (NCAR) real-time prediction system for a database composed of 92 storms, using leave-one-storm-out cross-validation (i.e. 91 storms for training dataset and 1 storm for out-of-sample). Forecast improvements based on error reduction using the GBLR approach are significant when compared to the ensemble mean (average of the ten-member forecasts). This presentation will also illustrate how the variances are optimized for the training partition in GBLR and discuss the verification strategy for grid points where no observations are available. The new technique is successful in improving wind speed and precipitation storm forecasts using past events and has the potential to be implemented in real-time.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner