The Storm Prediction Center’s Weather Generator

Despite advancements in storm-scale numerical weather prediction, direct deterministic prediction of tornadoes (including their number and exact locations) remains a difficult scientific challenge. As a part of Project IMPACTS (Integrated Machine-based Predictive Analytics for Convective Threats to Society), the NWS Storm Prediction Center (SPC) has developed Monte Carlo-like methods to statistically simulate tornadoes (the “weather generator”) and their characteristics, including number, location, path length, path width, direction, and intensity.

These simulations require SPC Day 1 Convective Outlooks (COs) and historical databases of tornado properties maintained by the SPC. Our analysis shows that the probability distribution of the number of tornadoes on a given convective day is well-represented by exponential Weibull distributions (for each Day 1 CO tornado probability) where the size of the area encompassed by that tornado probability is the independent variable. We have additionally identified two distinct distributions of observed tornado intensities associated with SPC Day 1 COs: Distribution 1, skewed toward weaker tornadoes, exists in the warm season (May - October) events when no SPC significant tornado (“sig tor”) contour is present in the forecast; Distribution 2, skewed toward stronger tornadoes, exists for all other events (all tornado days in November – April and for warm season days when a “sig tor” contour is issued). Given a tornado, one rated EF-2 or higher is 2.5 times as likely on days governed by Distribution 2 than on those days governed by Distribution 1.

We present results of SPC’s weather generator applied to all Day 1 COs from 2007 – 2016. When the weather generator is connected to a societal impact prediction model developed by Simmons et al. (2017), the combined method successfully identifies the most impactful tornado outbreaks of the last decade, including 16 April 2011, 27 April 2011, and 24 May 2011. We also generated “practically perfect” analyses of all convective days, with at least one tornado, dating back to 1 January 1950. These practically perfect analyses are intended as a rough approximation of what a forecaster possessing “perfect” prior knowledge might have drawn as a Day 1 CO. From 2007 – 2016, the results of the weather generator these analyses generally agree with the results when the simulator is fed the actual Day 1 CO. Prior to 2007, the weather generator, using these practically perfect analyses, is able to identify historically impactful tornado events like the 1974 Super Outbreak, the 1965 Palm Sunday Outbreak, the 3 May 1999 Outbreak, and the 31 May 1985 Outbreak. We conclude that this method is a robust way of statistically estimating the potential societal impacts of tornadoes using SPC’s Day 1 Convective Outlook product.