2002 Annual

Wednesday, 16 January 2002
A neural network approach to warm season rainfall forecasting in the Upper Midwest
Bradley R. Temeyer, Iowa State University, Ames, IA; and W. A. Gallus Jr. and C. G. Carmichael
A neural network has been developed to predict quantitative precipitation amount (QPF) and probability of precipitation (PoP) in the Upper Midwest during the warm season, to determine if the approach of Hall et al. (Wea. Forecasting, 1999), which worked well for prediction of these parameters in the Dallas-Forth Worth, TX region, could be applied to other areas. Some of the primary mechanisms resulting in warm season rainfall in the Upper Midwest, such as convergence at the nose of the low-level jet, differ from those in Texas. Therefore, we not only examined the parameters that Hall et al. found to be most significant for rainfall prediction in Texas, but expanded our search to include other measures of forcing and thermodynamics.

The neural network was designed using rawinsonde data and Eta model forecasts for Omaha, Nebraska from 355 days during the warm seasons (May 1 - September 30) of 1998-2000. Observed precipitation amounts were taken as an average of several sites in the Omaha area. A network trained on 50% of the cases evidences a high degree of forecast skill for heavier rainfall amounts. Light amounts (generally .25 inches or less in 24 hours) tend to be overforecasted, with high probabilities of detection but also high false alarm rates. Heavier rainfall amounts (.5 or 1.0 inches in 24 hours) are somewhat underforecasted, with nearly perfect false alarm rates (near zero). A preliminary comparison with rainfall forecasts from a 10 km version of the NCEP Eta model suggests that the neural network system is more skillful for the heavier rainfall amounts (.5 inches or more). For instance, for rainfall exceeding 1.0 inches, the equitable threat score (ETS) of the neural network system is .39, whereas the 10 km Eta model has an ETS of .08. The correlation of predicted amounts with those observed is much better in the neural network than in the standard grid point Eta model.

Supplementary URL: