Evaluating pipe failure risk in water distribution networks using GIS fuzzy overlay, ROC curves and machine learning logistic regression

This article presents a predictive model for assessing pipe failure risk in urban water distribution networks by integrating logistic regression – a supervised machine learning approach – with Geographic Information System (GIS) tools. The model incorporates key variables including pipe diameter, length, age, and internal pressure to estimate the likelihood of failure. Real-world failure and pipe data were collected from District 1 of Tehran's water distribution network over a seven-year period (2015–2021). Following statistical normalization and classification, logistic regression was applied to derive predictive equations for each variable. These equations were then used to generate raster-based GIS risk layers, assigning a failure probability between 0 and 1 to each spatial pixel. By applying fuzzy logic operators (AND and OR), the individual layers were integrated into a composite risk map that visualizes failure likelihood with fine spatial resolution. The results indicate that pipes with diameters of 60–150 mm, lengths under 200 meters, and ages between 15 and 25 years, along with pressures above 60 meters of water, exhibited significantly higher failure probabilities. This spatial modeling approach enables utility managers to identify high-risk areas for preventive maintenance.