Storm-Scale Weather Analysis and Prediction at the NOAA National Severe Storms Laboratory Using a Localized Particle Filter

The nonlinear evolution of model forecast errors in convective weather regimes presents a major obstacle for predicting severe weather events. Ensemble forecasts provide an effective means of characterizing this uncertainty, which is a crucial step towards assimilating observations in weather models. In practice, these errors are assumed to follow a parametric form such as a Gaussian distribution. While this strategy works effectively for many applications in atmospheric science, it is inappropriate when the underlying dynamics produce highly skewed, bounded, or even bimodal error distributions---all of which are typical when clouds are present.

This presentation summarizes recent efforts at NOAA NSSL developing and testing a particle filter-based data assimilation technique that avoids many of the assumptions made by current methods. Using severe weather outbreaks selected from the 2016 season, the new method, called the local particle filter, is compared with a conventional ensemble Kalman filter data assimilation method applied in the NSSL Experimental Warn-on-Forecast System for ensembles (NEWS-e) framework. These experiments demonstrate the feasibility of applying localized particle filters for real problems in atmospheric science, and demonstrate potential benefits of the method for convective-scale forecasting. Results from this study also provide insight into how remotely sensed data can be used most effectively for real-time numerical weather prediction.