Uncertainty Quantification and Spatial Biases Assessment in Precipitation Forecasts: A Methodology for Real-Time Flood Forecasting Applications

Meteorological forecasts are crucial for mitigating flood impacts, but they are always affected by uncertainty, particularly regarding the accurate prediction of the location for intense convective precipitation. This issue is especially critical for flood forecasting in small watersheds, where even slight displacements in the predicted rainfall position can lead to significant flow forecast inaccuracies. This study (i) proposes a methodology to assess spatial biases of rainfall forecasts produced by a convection permitting meteorological model, (ii) identifies whether the model has "preferential" misplacement directions in forecasting convective events, and (iii) suggests a method to cope and deal with this uncertainty in hydrological predictions. In this study, 64 significant convective rainfall events have been analyzed by comparing the quantitative precipitation forecast (QPF) from the Modello Locale in Hybrid Coordinates (MOLOCH) meteorological model with observed rainfall fields, covering a large portion of the Lombardy Region (northern Italy). The model's average displacement error is quantified using the fractions skill score (FSS), yielding a value of 20 km. By deriving each rainfall event's displacement vector through pattern matching, a systematic misplacement tendency has been identified, with a consistent forecast shift by the model toward the northeast direction in the "hydraulic node of Milan" study area. The bidimensional rainfall displacement probability density function is then obtained through kernel density estimation (KDE). This distribution can be used as a generator of shifted rainfall forecasts, substantially creating an ensemble from a high-resolution deterministic model, aimed at taking into account the uncertainty associated with possible QPF misplacements. The methodology can be generalized and applied to any river basin and limited-area meteorological model. SIGNIFICANCE STATEMENT: This study aims to improve flood forecasting by addressing uncertainties in predicting the location of intense rainfall, particularly in small watersheds. It analyzes 64 rainfall events in northern Italy using the Modello Locale in Hybrid Coordinates (MOLOCH) meteorological model to assess precipitation forecast errors. A key finding is that the model tends to misplace rainfall by about 20 km, mainly shifting it toward northeast. The study introduces a method to take into account these spatial errors in hydrological predictions, using statistical techniques to generate a range of possible rainfall scenarios. This approach helps in quantifying uncertainty, improving flood predictions, and supporting forecasters' decisions and is potentially suitable for other hydrometeorological forecasting systems in other regions.