Ensemble Convective Rainfall Nowcasting by integrating Numerical Weather Prediction models and Neural Networks: the ICREN project

Convective events pose a significant threat to society due to the associated heavy rainfall, large hail, strong winds, and lightning. Location and timing determination of convective precipitation is still a challenge for modern meteorology. Despite the good skills of current weather forecasting tools in the prediction of the large-scale environment facilitating the onset of convective phenomena, the multitude of spatial scales involved in such events makes their characterization, observation, and forecast a difficult task. The problem is further complicated by their rapid temporal development, which lasts from minutes to a few hours depending on the specific case.

Recent research indicates that the predictability of these events can be strongly improved accounting for local meteorological observations. 

The goal of the ICREN (Intense Convective Rainfall Events Nowcasting) project is to enhance the nowcasting of convective events by:

• exploiting the information made available by local standard and non-conventional observations of meteorological variables

• integrating physically based Numerical Weather Prediction (NWP) models with data-driven black box Neural Networks (NNs). 

The NWP model is used to support the NN by means of pseudo-observations (forecasted variables); while the fast computational speed of the NN enables advancing predictions in time and generating ensemble forecasts of convective phenomena.

The project is carried out in the Seveso River basin (almost 300 km²) in Northern Italy. In this region, convective events trigger floods and flash floods heavily impacting the large urban area of Milan.

Within the project, the Weather Research and Forecasting (WRF) NWP model is employed. By using three nested grids, the model achieves a 2 kkm x 2 km spatial resolution over the test area. To optimize the prediction of meteorological variables required by the NN, the model assimilates lightning observations and GNSS-derived Zenith Tropospheric Delays (ZTDs), both of which enhance the representation of local atmospheric humidity.

Several NN models have been trained on standard meteorological data, GNSS ZTDs, and radar-derived parameters—including the position, velocity, and attenuation of convective cells—to identify the architecture best suited for predicting 10-minute accumulated rainfall from 10 minutes up to 1 hour following the detection of a convective event in the test area.

The best-performing models are used to generate ensemble predictions of rainfall events by suitably perturbing the input variables.

Results from the WRF model, the NN predictions and the ensemble forecasts will be presented along with initial integration outcomes for selected convective events occurring in the test area in 2019.

 

This work is supported by the ICREN-PRIN project (MUR- CUP: D53D23004770006).