Optimizing hydrological modeling on real urban catchment: impact of calibration data selection

Hydrological models are crucial in various engineering applications, including streamflow forecasting and flood risk estimation. Tools like the Stormwater Management Model (SWMM) are indispensable for efficacious water resource management. Calibrating these models is a necessary step to minimize parameter uncertainties and ensure accurate representation of a catchment area's hydrological response. However, the calibration process often faces challenges due to the need for extensive parameter adjustments. Sensitivity analysis (SA) is employed to mitigate these challenges by identifying and focusing on the most influential parameters, thereby streamlining the calibration process. In this work, the Morris method was applied to identify the sensitive parameters in the SWMM model, which were subsequently considered in the optimization process using Genetic Algorithms (GA). The results of the sensitivity analysis highly depend on the model output targets, such as total runoff volume and peak flow rate.

The traditional approach of dividing data into calibration and evaluation subsets is a fundamental practice in model development. Nevertheless, the impact of data allocation on model evaluation performance has not received sufficient attention in the literature. This study investigates the influence of calibration data selection on model performance, utilizing high-resolution experimental rainfall-runoff data from the urban catchment of Cascina Scala in Pavia, Italy. Four criteria—rainfall depth, mean intensity, hydrograph's center of mass, and maximum rainfall depth over five minutes—were employed to select the calibration set. From a total of 24 events, the four criteria were employed to select 8 events from 16 for calibration, while the remaining 8 events were designated for validation. The findings underscore that the selection of the calibration dataset substantially influences the optimally calibrated parameters, subsequently altering model performance.