Performance of two global models in forecasting extreme rainfall volumes over southern Brazil

Extreme rainfall in late April–early May 2024 led to the most severe flooding ever recorded in the Guaíba River Basin, southern Brazil.. Waters from this basin drain into Patos Lagoon before reaching the Atlantic Ocean. As a result, the exceptional precipitation volumes caused widespread flooding both along the river network and in cities surrounding the lagoon. The event affected 2,398,255 people in 478 cities (96% of the cities in the state of Rio Grande do Sul), causing 184 fatalities and leaving 25 people missing. The Guaíba Basin lies in a topographically complex region, with mountainous areas that amplify orographic precipitation and increase the difficulty of forecasting by global models. The event was associated with an atmospheric configuration conducive to persistent rainfall, characterized by an intensified subtropical jet, strong warm and moist air transport by a low-level jet, and the passage of cold fronts. Together, these factors promoted the development of mesoscale convective systems and produced exceptionally high rainfall accumulations. Nearly all National Institute of Meteorology (INMET) stations within the basin recorded more than 200 mm over five days, with peak totals reaching 540 mm, resulting in exceptionally large runoff volumes. This study evaluates how well the global GFS and ECMWF models forecast accumulated precipitation over the Guaíba Basin at lead times of up to 72 hours. Model precipitation forecasts were converted to basin-integrated rainfall volumes (m³) and evaluated against observations from INMET automatic stations interpolated onto the same grid. This volumetric approach captures the basin’s hydrological response more directly than traditional metrics based on point measurements or spatial averages. The results show that both models strongly underestimated the precipitation volume over the basin, with biases on the order of 10 to 18 billion m³ at lead times of 48 to 72 hours. Although the ECMWF showed better performance during the first 12–24 hours, both models quickly converged toward similarly underestimated solutions. This behavior indicates a failure to represent the persistence of the atmospheric circulation and the sustained moisture transport associated with the event. Such behavior suggests that the models were able to initiate precipitation but failed to maintain the synoptic and mesoscale forcing required to reproduce the observed hydrological magnitude of the event. This pattern is consistent with events characterized by atmospheric blocking and persistent low-level jets. These findings highlight important limitations of global models in forecasting persistent extreme events over complex river basins. They emphasize the need for hydrometeorological forecasting strategies that combine global and mesoscale models with ensemble prediction systems, regional adjustments, and volumetric metrics to better anticipate hydrological impacts and support early warning and disaster risk reduction.