Verification of Statistical Models for Lightning Prediction

William R. Burrows

Meteorological Research Branch, Meteorological Service of Canada,

Downsview, Ontario

The Canadian Lightning Detection Network provides continuous lightning detection over all Canada to about 65°N in the west and 55°N in the east. Canadian coverage is melded with the northern portion of the U.S. network. Statistical lightning prediction models are in development by the author for 5° latitude by 5° longitude boxes for May-September. The predictand is three-hour lightning report density (LRD) gridded to 22 km resolution. A flash report is weighted 1 when within 0-10 km of a grid point, and a decreasing weight of 1-0 when within 10-20 km. The predictand is translated into 11 categories, most based on a log-normal shape. The predictand is matched with 112 potential predictors derived from 0-12 hour forecasts by the Canadian Meteorological Center’s GEM numerical weather prediction model at 00Z and 12Z. The same models are applied for each 24-hr period. Models are built with the (non-linear) regression engine of Classification and Regression Trees. The probability of occurrence of LRD categories, calculated from the learning data in each CART terminal node, is forecast. Verification should account for the closeness of the forecast to observation in both space and time. A simple verification method has been developed thus far. Forecasts of the (very common) no-lightning category are subject to stringent space-time verification while the lightning categories enjoy a more relaxed verification.