The MISTRAL project (Meteo Italian Supercomputing Portal), funded under the EU’s Connecting Europe Facility (CEF), has come to an end after 30 months. The project’s main goals were to facilitate and foster the re-use of meteorological datasets by various communities, to provide added-value services using high-performance computing (HPC) resources. The project lead and provider of HPC facilities was CINECA, based in Bologna. Along with ECMWF, the other collaborating partners were Arpae Emilia-Romagna, Arpa Piemonte, Protezione Civile and Dedagroup (all in Italy).
ECMWF participated in MISTRAL to improve probabilistic 6-hour rainfall forecasts to help anticipate flash floods in Italy: the so-called ‘flash flood use case’. Localised very heavy rainfall is commonly associated with flash floods, but it is difficult to forecast it accurately in both magnitude and location. Ideally weather forecasts should be provided for points, and not for the large regions represented by the grid boxes of global models, such as those of ECMWF. The mismatch can be addressed by post-processing model output or by using convection-resolving limited-area models (LAMs). In MISTRAL, ECMWF used both approaches in ensemble mode, with blending used to create the final products. We also applied post-processing to the LAM output before blending, to address the limitations of finite ensemble size that play a central role when convection is resolved.
Blending global and LAM output
For the global model post-processing, we employed an ECMWF approach based on decision trees (ecPoint-Rainfall), which relies on calibration against rain gauges. Two underlying premises of ecPoint-Rainfall are that the relationship of forecasts versus point-observations depends not on location but on weather types, and that each such weather type has a different sub-grid point rainfall distribution associated. ECMWF already runs ecPoint-Rainfall for 12-hour accumulation periods at ECMWF. In MISTRAL we redesigned the weather types for use with 6-hour totals. Use of 6-hour rainfall is more appropriate here, given the propensity for flash floods to be driven by extreme rainfall over short periods.
For the LAM ensemble, ECMWF used COSMO-2I-EPS, created by the COSMO-LAMI consortium (Consortium for Small-scale Modelling-Limited Area Model Italia). This is a 2.2 km LAM ensemble prediction system run for Italy and surrounding regions. The post-processing applied here was a new state-of-the-art scale-selective neighbourhood technique. This technique aims to selectively and dynamically preserve, at fine scales, the (heavy) rainfall signals that are reliable – assumed to be where ensemble member grid box totals agree – and to smooth out probabilities elsewhere, across neighbourhoods proportionate in size to the level of disagreement.
By blending, we aim to combine the two systems’ most skilful aspects. Blended output is available up to 48 hours (when the LAM runs end), with the ecPoint relative weight rising from zero at T+0 to one by T+48. From T+48 to T+240, only ecPoint is used. This strategy minimises the impact of fast error growth with lead time in the LAM and provides product continuity. ecPoint-Rainfall is specifically designed to improve the reliability and discrimination of the forecast. In verification, it shows particularly good results for large totals. By combining it with LAM-based output, we can be more specific at short lead times about where extreme totals are most likely. The final products are delivered on the COSMO grid, as percentiles and probabilities of exceeding specific thresholds for 6-hour rainfall. This serves Italian forecasting and feeds into a gateway for Europe. The use of CINECA HPC resources has been crucial to produce the real-time products operationally since the end of 2019.
Open data portal
The new products can be found in the new National Meteorological Open Data Portal (https://meteohub.mistralportal.it/app/maps/flashflood/) created in the framework of the project. In the portal you can also find other forecasts, satellite data, radar and SYNOP observations. With this project, ECMWF benefited greatly from further development of ecPoint and from a close look at convection-permitting ensemble output and the challenges of post-processing it. Gaining timely access to high-density observations for Italy, for verification and research, was another plus.