Post-processing rainfall in a high-resolution simulation of the 1994 Piedmont flood

Abstract

In November 1994, a catastrophic flooding event occurred in the Piedmont region in Northwestern Italy over a period of about 3 days. The large time and spatial scales associated with this event prompted a number of reanalysis studies to assess the forecast skill of the models. This paper investigates another forecasting technique using the Weather Research and Forecasting (WRF) model coupled with post-processing techniques: the analog ensemble (AnEn) and the convolutional neural network (CNN). The complex topography found in this region presents a challenge for numerical weather prediction (NWP) models especially for events such as these, where the orography is crucial in determining the distribution and amount of precipitation. By applying these post-processing techniques to WRF model output, significant improvements were observed in the accumulated precipitation fields during the flooding event in both techniques, although improvements using the CNN were at the expense of underestimating the highest precipitation.