A Short-Term Hail Prediction System Based on Numerical Weather Modeling and Machine Learning

In the United States, hail storm causes billions of dollars damages to the properties and claims around $850 million insurance loss each year. The population growth in bigger cities such as Denver, Dallas Fort Worth, and Chicago have made substantial amounts of property damage from hail events. It is essential to predict and give warning to people about the hail storms and its path at least 1 hour prior. The focus of this study is to develop a short term hail prediction system for the Dallas Fort Worth (DFW) urban area using a numerical weather model (NWP) and machine learning techniques. The Collaborative Adaptive Sensing of the Atmosphere (CASA) deployed a dense X-band radar network and developed the end-to-end warning system for severe weather events like floods, and tornadoes for DFW urban area. CASA also implemented the real-time hail detection algorithm over DFW radar network. The Weather Research and Forecasting (WRF) model is used to simulate the hail storms over the DFW area.

The High-Resolution Rapid Refresh (HRRR) model analysis is used as initial and boundary conditions. WRF model domain covers 301 x 301 km2 area over DFW urban region with 1 km grid resolution. The WRF Three-Dimensional Variational (3DVAR) data assimilation system is used to assimilate high spatial and temporal Next Generation Weather Radar (NEXRAD, S-band) and CASA’s X-band radar observations in the WRF model. The one-dimensional HAILCAST model is used to estimate the maximum surface hail size. The Understory hail sensor network measurements are used to validate the WRF-HAILCAST model estimated hail size. The first phase of the research work focused on comparing the understory hail sensor measurement and CASA’s real-time hail products over DFW hail storms. The comparison results showed both measurements are matched in terms of space and time.

The second phase of this research work is focused to simulating and validating the WRF-HAILCAST model on two hail storm events 1) June 06, 2018, 2) March 13, 2019, which produced large hail and strong winds over DFW urban area. The result shows that hail is not predicted in the control simulation. Whereas the simulation shows the hail when the high spatial and temporal NEXRAD and CASA radar observations are assimilated into the WRF model. The maximum hail size estimated from CASA-HAILCAST model after radar data assimilation is closer to the hail size measured by hail sensors. Even though the maximum hail size estimated from WRF-HAILCAST model closer to hail sensor measurements, the WRF-HAILCAST underestimates the hail size.

The main objective of this research work is to overcome the existing inbuilt one-dimensional HAILCAST model with the machine learning based model. The final phase of the work to developing a machine learning based algorithm to estimate the surface hail size from WRF model simulated vertical profiles of the storm. The understory hail sensor measurements will be used in machine learning based model development. The enhancement from machine learning model and the detailed results will be presented in the conference.