Sounding-derived indices for neural network based short-term thunderstorm and rainfall forecasts

摘要

A neural network-based scheme to do a multivariate analysis for forecasting the occurrence and intensity of a meteo event is presented. Many sounding-derived indices are combined together to build a short-term forecast of thunderstorm and rainfall events, in the plain of the Friuli Venezia Giulia region (hereafter FVG, NE Italy). For thunderstorm forecasting, sounding, lightning strikes and mesonet station data (rain and wind) from April to November of the years 1995-2002 have been used to train and validate the artificial neural network (hereafter ANN), while the 2003 and 2004 data have been used as an independent test sample. Two kind of ANNs have been developed: the first is a "classification model" ANN and is built for forecasting the thunderstorm occurrence. If this first ANN predicts convective activity, then a second ANN, built as a "regression model", is used for forecasting the thunderstorm intensity, as defined in a previous article. The classification performances are evaluated with the ROC diagram and some indices derived from the Table of Contingency (like KSS, FAR, Odds Ratio). The regression performances are evaluated using the Mean Square Error and the linear cross correlation coefficient R. A similar approach is applied to the problem of 6 h rainfall forecast in the Friuli Venezia Giulia plain, but in this second case the data cover the period from 1992 to 2004. Also the forecasts of binary events (defined as the occurrence of 5, 20 or 40 mm of maximum rain), made by classification and regression ANN, were compared. Particular emphasis is given to the sounding-derived indices which are chosen in the first places by the predictor forward selection algorithm.