Calibração de rugosidades absolutas e coeficientes de decaimento de cloro nas paredes das tubulações em redes de distribuição de água: abordagens com redes neurais artificiais e métodos Híbridos MIGHA-RNA

Abstract

A computational model of a water distribution network capable of performing hydraulic and chlorine transport simulations is a valuable resource for water utility companies. However, the complexity due to numerous parameters and associated uncertainties presents a significant challenge in constructing these models. In this research, calibration methods were developed and tested for pipe absolute roughness and chlorine wall decay coefficients (kw), using algorithms based on Artificial Neural Networks (ANN) and hybrid methods that combine the Alternative Hydraulic Gradient Iterative Method (MIGHA) with a Multilayer Perceptron (MLP). Four methods were developed and evaluated in three fictitious networks with distinct characteristics. Different scenarios were explored, varying the amount of observed data used in the tests. The results obtained by the developed methods were compared to each other and to those produced by calibrations using only MIGHA and Genetic Algorithms, considering the adjusted parameters and state variables of the model, such as pressure heads, flows, and chlorine concentrations returned by the calibrated networks. Additionally, an adaptation for MIGHA was designed to enable dynamic calibrations, and a graphical interface was developed for a software application using the developed algorithms. The results suggest that the G-ANN and G-Hybrid methods, which use a group-based calibration routine for similar pipes, tend to outperform, in most cases, the S-ANN and S-Hybrid models, which update parameters segment by segment. Further tests indicated that segment-based calibration methods have the potential to surpass those based on group methodology, provided that the parameter variation range used in the training and testing data generation for the MLPs is precise. Furthermore, this study also involved the modeling of the Federal University of Piauí (UFPI) network, where the G-Hybrid method was used to calibrate the model’s absolute roughness. After calibration, the mean absolute error for pressure heads was 0.40 mH2O, a result deemed satisfactory for the study conditions. The developed methods and adaptations represent a promising advancement in improving the reliability and efficiency of water distribution network calibration, contributing to more accurate and safer water system management.