Desempenho operacional e custos em obras de redes coletoras de esgoto: uma abordagem com aprendizado de máquina e dados reais em Fortaleza-CE

Abstract

Sewer network construction projects are characterized by high operational variability, strong dependence on local implementation conditions, and recurrent challenges related to physical and financial planning. In this context, this dissertation investigated the operational performance and implementation costs of sewer collector networks and household connections based on real field data collected from a large-scale project carried out in Fortaleza, Ceará, Brazil, between October 2022 and August 2025. The analyzed database comprised approximately 160 km of sewer network, 21,000 household connections, and 35 months of continuous monitoring of productive, geometric, operational, climatic, and economic variables. The research was structured into two complementary pillars. The first evaluated the physical performance of construction crews through machine learning models applied to operational productivity analysis. The second investigated operational costs associated with sewer network implementation, considering labor, equipment, fuel consumption, materials, and execution constraints. The study employed Linear Regression, Ridge Regression, Random Forest, Gradient Boosting, and Support Vector Regression algorithms, combined with a temporal validation protocol based on Nested Cross-Validation and TimeSeriesSplit, aiming to preserve chronological causality and reduce optimistic evaluation bias. The results revealed a significant discrepancy between ex-post performance and forecast-safe prospective performance, highlighting the limitations of predictive generalization in real-world urban construction scenarios. Although high goodness-of-fit metrics were obtained during historical adjustment, the models exhibited substantial degradation under external temporal validation, indicating that operational and construction-related variables present highly unstable dynamics over time. Nevertheless, the modeling process enabled the identification of relevant drivers associated with productivity and costs, including excavation depth, pavement type, groundwater lowering, seasonality, and crew composition. Based on these findings, an operational simulation prototype was developed to support comparative scenario analysis and planning decision-making in sanitation projects, using empirically calibrated parameters derived from the observed historical database. The research contributes to advancing machine learning applications in sanitary engineering by simultaneously demonstrating both the explanatory potential of these techniques and the limitations of their predictive application in real and temporally dynamic operational environments.