Spatial Analysis of Leakage Events in Water Distribution Systems

Water utilities need practical methods to prioritise inspection and leakage detection activities across the distribution networks. This thesis develops a modelling framework that combines information on individual pipe characteristics with the historical leakage intensity of the connected surrounding network. The analysis is based on GIS and leakage records from the water distribution network managed by Irisacqua S.r.l. in north-eastern Italy, covering the period from 2016 to September 2025. Pipe-level prioritisation is first examined using Logistic regression, Poisson regression, a Generalized Additive Model (GAM), and Histogram Gradient Boosting (HGB). These models are evaluated using forward validation over the 2016--2023 development period and a ranking metric based on the proportion of subsequent leakage events captured within a limited fraction of total network length. A separate topological score is constructed from historical leakage events in the surrounding network. This information is then combined with the pipe-level models either by merging the rankings or by adding the topological index as a model predictor. The four pipe-level models perform similarly overall, with a small advantage for GAM and HGB. The topological ranking is only weakly correlated with the pipe-level rankings, indicating that it contains different information. When the two sources are combined, prioritisation generally improves. This result is observed both in the complete 2024 evaluation on later data and in a supplementary evaluation using partial 2025 data. The results suggest that historical leakage patterns in the connected surrounding network can provide useful additional information for pipe prioritisation.