84th AMS Annual Meeting

Monday, 12 January 2004: 9:00 AM
An Experimental Severe Weather Nowcast Algorithm Based on a Back-propagation Neural Network and a Comparison with an Algorithm Based on Multiple Linear Regression
Room 602/603
Yerong Feng, Guangdong Provincial Meteorological Observatory, Guangzhou, China; and D. H. Kitzmiller
Poster PDF (133.8 kB)
Two kinds of severe weather nowcast algorithms were developed in this experimental work using neural network (NN) approach and multiple forward-selection linear screening regression (LR) approach. The algorithms trend to utilize WSR-88D Doppler Radar data and Numerical Weather Prediction model outputs to provide probabilities of severe weather centered on a convective storm cell, 44 km on a side, 30 min after radar observation.

Detailed comparison of the skills of NN and LR algorithms in predicting severe weathers on dependent and independent datasets showed that NN approach, with its nonlinear and collective expression of all the candidate predictors, provided higher forecast scores when comparing to the forward-selection method based on linear expression of a few selected predictors. Thus NN is more capable of detecting characteristic patterns of severe weather events that may exist in the statistical datasets.

The algorithms were developed to serve as the Severe Weather Threat Index (i.e. probability to produce severe weather) in the Systems for Convective Analysis and Nowcasting (SCAN), a suite of Advanced Weather Interactive Processing System (AWIPS), to automatically generate probabilities that individual thunderstorms will produce severe weather such as tornado, large hail and damaging wind.

Supplementary URL: