Convective Rainfall Nowcasting: comparison between Numerical Weather Prediction models and Neural Networks in view of an integrated approach

The prediction of convective storms, even a few hours in advance, could help reduce the impact of associated phenomena such as heavy rainfall, strong winds, lightning, and large hail. Although highly beneficial to society, accurately forecasting where and when these phenomena will occur remains a major challenge. This is due both to the wide range of spatial scales involved and to the rapid temporal evolution of these events, which typically last from minutes to a few hours. Recent research indicates that the predictability of such events can be significantly improved by incorporating local meteorological observations.

In this context, the ICREN project (Intense Convective Rainfall Events Nowcasting) investigated the possibility of enhancing the nowcasting of convective events in the Seveso River Basin, located in the Lombardy region of Northern Italy, where such events frequently trigger floods and flash floods, severely impacting the urban area of Milan.

The aim of the project was to exploit information provided by local standard and non-conventional meteorological observations through an ad hoc model that integrates physically based Numerical Weather Prediction (NWP) models with data-driven black-box Neural Networks (NNs). The NWP model supports the NN by providing pseudo-observations in the form of forecasted variables, while the fast numerical NN is used to advance the predictions in time and to generate ensemble forecasts of convective phenomena.

This presentation mainly focuses on the research activities devoted to the development of data-driven models and their intercomparison. Furthermore, it illustrates how these models perform with respect to NWP model predictions, both before and after the assimilation of local observations, in order to address the main research question of the project: namely, whether data-driven models are able to integrate NWP predictions at a very local scale and to rapidly advance these predictions in time. In other words, is there an advantage in coupling these two types of models, and to what extent?

Although NN model accuracy decreases with forecast lead time, the predictions outperform those of the NWP models in terms of localization of convective phenomena, confirming that their combination can enhance current NWP forecasting capabilities.