Feedforward Deep Neural Network-Based Model for the Forecast of Severe Convective Wind

Abstract

AbstractIn this study, we compared and analyzed the environmental characteristics in Shaanxi between 2016–2018 (before and after the occurrence of convective wind) based on both severe and non-severe wind samples extracted from the fifth generation of European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis (ERA5) data and automatic weather station data from the China Meteorological Administration. Through the establishment of deep feedforward neural networks, these environmental characteristics were then used to define a model for the forecast of severe convective wind events. For \(\sim\)75% (or > 75%) severe convective wind samples, the differences between 1 h after the occurrence of severe convective wind and when it occurred, 1 h after it occurred and 1 h before it occurred in MLCAPE, MUCAPE, SBCAPE, MLCIN, MUCIN, SBCIN, MLEL, MUEL, SBEL, SBLCL and temperature were negative, while those in MLLI, MULI, SBLI, sea level pressure (SLP), and precipitation were positive. This suggests that, for \(\sim\)75% (or > 75%) severe convective wind samples, atmospheric stability increased or SLP increased, or temperature decreased; moreover, precipitation was found to occur after severe convective wind. Finally, a certain degree of differentiation was noted between the parameters associated with the severe and non-severe convective wind samples. We designed two schemes (each containing three kinds of experiments) and trained a feedforward deep neural network to predict severe convective wind events. The network experiment including the higher number of elements was found to provide the highest threat score (Ts).KeywordsSevere convective windEnvironmental characteristicsFeedforward deep neural network