Assimilation of Local Ground Stations and Radar Data to Improve the Prediction of the 9–10 September 2017 Thunderstorm in Livorno, Italy

More and more frequently the countries bordering the Mediterranean Sea experience very intense rainfall on limited areas, causing flash flooding and casualties. Many regional and international European meteorological centers are engaged in a great effort to improve the prediction capability of such extreme events. Because of their exceptional intensity and their small spatial and temporal scale, one of the most promising approach points at high resolution short term forecast, possibly with the assimilation of local near real-time data. This presentation aims at summarizing the results of the study of the meteorological predictability of the heavy precipitation event that hits the town of Livorno, located in central Italy (Tuscany), during the night between 9 and 10 September 2017. The event was characterized by accumulated precipitation exceeding 200 mm in 3 hours, associated with a return period higher than 500 years. As a result, all the largest streams of the Livorno municipality flooded several areas of the town, causing the loss of nine human lives. We used the WRF model at 3 km of horizontal grid spacing (as the operational forecast mode which currently runs at the regional Weather Service of Tuscany Region, LaMMA), to reconstruct the dynamics of the event triggered by a strong convective activity. Ground stations (P, T, RH, wind), radiosonde (RAOB) data, and Radar reflectivity observations (X- and S- bands) collected a few hours before the rainfall peak were also assimilated into WRF hindcast runs, to test their effect on the numerical prediction of the event, both in terms of quantification of its intensity, spatial localization and timing. Specifically, a control run without any data assimilation that provided the baseline scenario was compared with runs that assimilated ground stations observations only, radiosonde observations only, radar only, ground stations and radiosonde, and all observations. For all these runs, data assimilation ended between 3 and 6 hours before the beginning of the event, and the run continued during the event without any data assimilation. This allowed us to highlight the effectiveness of the assimilation of weather observations for this severe thunderstorm in the immediate hours before the event. Simulations were verified using radar, satellite and rain-gauges data.