Extended Abstract (776K)J5.6
A Real-Time Learning Technique to Predict Cloud-to-Ground Lightning
V. Lakshmanan, CIMMS/Univ. of Oklahoma and NOAA/NSSL, Norman, OK; and G. J. Stumpf
A short-term warning for intense cloud-to-ground (CG) lightning has the potential to be a very valuable new National Weather Service (NWS) operational product. A variety of forecast techniques and “rules-of-thumb” have been developed by several local NWS Weather Forecast Offices which manually combine WSR-88D radar data with thermodynamic information from local rawinsondes to predict CG lightning occurrence. Although a 0-3 hour forecast of CG lightning is available in AWIPS, this product provides no guidance on the likelihood of very high CG flash rates. The NWS Meteorological Development Laboratory (MDL) partnered with the NSSL to develop a prototype application to predict intense CG lightning.
Using the National Severe Storms Laboratory (NSSL) Warning Decision Support System – Integrated Information (WDSSII) as a prototyping environment, a multiple-sensor application which predicts the initiation and the advection of cloud-to-ground lightning has been developed. This application first combines radar data from multiple WSR-88D locations into a 3D Cartesian grid. The radar grid is then integrated with near-storm environment information extracted from the 20 km Rapid Update Cycle model (RUC20) initial analysis grids. This is done to diagnose radar reflectivity information within the thermodynamic levels in the mixed-phase region of storms, typically between 0° and -20°C, where charge separation and subsequent lightning is known to occur. A statistical clustering scheme is used to advect the 3D radar data, as well as any CG lightning density grids. A linear combination of Gaussian functions, called a Radial Basis Function (RBF), is used to combine these multiple-sensor fields into a prediction field of future CG flash density. The weights and standard deviations of the RBF are optimally determined by training it on historical data. This training is continually repeated every 15 minutes, based on the latest cloud-to-ground flash density information, and the radar and thermodynamic data from 1 hr before (for a 1 hour prediction interval). Thus, the RBF is trained to predict both advection of current lightning, and the initiation of new lightning. The weights and standard deviations correspond to the current storm-scale environment.
Joint Session 5, AI Applications with a Nowcasting Flavor (Joint between the Fourth Conference on Artificial Intelligence and the 21st International Conference on Interactive Information and Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology)
Tuesday, 11 January 2005, 9:00 AM-11:30 AM
Previous paper
Recorded presentation