Calibração automática do coeficiente de Manning em redes de drenagem urbana por inteligência computacional com análise de equifinalidade

Abstract

Urban stormwater management is a challenge intensified by urbanization, requiring the use of hydrological models like SWMM. The precision of these models, however, depends on the calibration of sensitive parameters, notably the Manning coefficient (n). This process is hindered by the scarcity of observed data, the time-consuming nature of manual calibrations, and the equifinality phenomenon, where multiple parameter sets can yield similar results. Literature establishes theoretical foundations but points to the need for hybrid approaches that integrate computational intelligence to overcome these limitations, especially in estimating parameters for unmonitored locations. This thesis aimed to develop a methodology focused on two computational tools to perform the automatic calibration of the Manning coefficient in sub basins and conduits of urban drainage networks. The methodology of Program 1 (sub-basins) applied Genetic Algorithms (GA) to optimize the N-Imperv and N-Perv coefficients, implementing a statistical strategy to filter the best solutions, seeking to mitigate equifinality. The methodology of Program 2 (conduits) used a hybrid approach: it applied the False Position Method for the calibration of sections with observed data and then used these results to train Machine Learning (ML) models, which estimated the coefficients in sections without observed data. The results from Program 1 revealed that calibration based on surface runoff flow data is superior to that based on depth data, allowing for the nearly perfect identification of N-Imperv (errors < 0.2%). However, N-Perv (pervious) proved to be a parameter of difficult identification and high sensitivity, whose calibration is strongly dependent on other model configurations, such as the selected infiltration method. The results from Program 2 confirmed the effectiveness of the hybrid ML approach, whose precision was directly proportional to the monitoring coverage. The research also empirically confirmed equifinality by demonstrating that large errors in Manning parameters can still produce hydraulically acceptable errors in water depth. The work concludes that the combination of physical models, optimization algorithms (GA), and machine learning techniques are viable strategies for automatic calibration, but obtaining high-precision results in drainage networks remains dependent on a significant amount of monitoring data.