Added value of a multi-model ensemble of convection-permitting rainfall re-forecasts over Italy

The interest towards the development of regional high-resolution retrospective datasets, allowing an enhanced representation of past meteorological states, has been progressively increasing. Convection-permitting (CP) datasets (i.e., hindcasts, based on historical model integrations, or reanalyses, including the additional assimilation of historical observational data) have demonstrated to improve the representation of precipitation compared to convection-parameterized counterparts. The benefits involve the spatial structures of rainfall fields, the timing and peak of the diurnal cycle of summer precipitation, and the frequency of wet days/hours. This is of particular relevance for enhancing the characterization of severe precipitation events with the aim to prevent and minimize their impacts on terrestrial ecosystems, and on human and animal life. However, the simulation of convective-related phenomena is highly model-dependent, implying the inability to sample the full range of natural variability with single-model experiments. This is exacerbated for km-scale simulations owing to the intrinsic chaotic behavior underlying convective processes.

Recently, the development of Multi-Model Ensembles (MMEs) of CP regional climate models over Europe has demonstrated to efficiently tackle this issue and reduce the simulation error associated with single model outputs. This approach could benefit also retrospective estimates in order to retrieve a complete, homogeneous, and optimized assessment of past atmospheric states. In case of precipitation, this could be valuable also for potential downstream modeling applications such as forcing hydrological forecasting systems to obtain improved historical series of high-impact flood events.

This work presents the first MME of retrospective CP datasets over Italy based on four reanalyses/hindcasts recently produced, with the aim to assess the added value of their joint employment. The datasets are obtained by dynamically downscaling the global reanalysis ERA5 using different numerical models: MERIDA-HRES (based on WRF-ARW), the hindcast based on the model MOLOCH, and SPHERA and VHR_REA-IT (both based on COSMO). The reference dataset for comparison is GRIPHO, the first Italian pluviometer-based hourly analysis. The investigation over a decade (2007-2016) includes various aspects such as the annual and seasonal dependence of daily and hourly mean rainfall intensity and frequency, heavy precipitation occurrences, and their summer diurnal cycles. The results indicate that a superior dataset performing always best is not detected, while large inter-model variability characterizes summer precipitation. The ensemble aggregation systematically improves rainfall estimates over single datasets, resulting in more adherent spatial fields and lower root-mean-squared errors and relative biases when compared to the observations, at the expense of reduced spatial variability of the distributions.