Predicting water distribution pipe failures using machine learning and cross-infrastructure data

Kozelj, Daniel (author), Fernández, David Abert (author)

Water pipeline failures in urban networks are a significant source of non-revenue water, service disruptions, and high maintenance costs. This study develops a machine learning model to predict pipeline failure probabilities and inform risk-based maintenance strategies. Trained on real-world assets and geospatial data from 2010 to 2025, the model incorporates standard pipe attributes – such as material, age, diameter, network type, and maintenance history – alongside spatially derived indicators of the surrounding infrastructure. Notably, it quantifies the predictive impact of adjacent infrastructure systems, including electricity grids, gas pipelines, district heating, sewage systems, and roads, utilizing spatial buffering and overlay techniques. Several of these cross-utility features, particularly road category, electricity voltage, and sewer type, showed meaningful predictive importance, reflecting their indirect but consistent influence on the risk of pipe failure. The ML model, built with the XGBoost algorithm and validated through stratified K-fold cross-validation, achieved high performance (ROC AUC: 0.9102, recall: 0.7750, accuracy: 0.8750). Despite lower precision due to class imbalance, the F1 score (0.2261) and LogLoss (0.2500) confirm its reliability. This study introduces a novel, spatially enriched approach to failure prediction, advancing urban infrastructure management through context-aware, data-driven insights.

• Language: English
• Keywords: water distribution systems, pipe failure prediction, machine learning, XGBoost, spatial analysis, condition assessment
• Work type: Article
• Typology: 1.01 - Original Scientific Article
• Organization: FGG - Faculty of Civil and Geodetic Engineering
• Publication status: Published
• Publication version: Version of Record
• Year: 2025
• Number of pages: Str. 53-64
• Numbering: Vol. 38, no. 68
• PID: 20.500.12556/RUL-171326
• UDC: 004.85:628.1(1-21)
• ISSN on article: 0352-3551
• DOI: 10.15292/acta.hydro.2025.05
• COBISS.SI-ID: 246363651
• Publication date in RUL: 22.08.2025

Record is a part of a journal

• Title: Acta hydrotechnica
• Publisher: Fakulteta za gradbeništvo in geodezijo
• ISSN: 0352-3551
• COBISS.SI-ID: 3664386

Licences

• License: CC BY-NC-SA 4.0, Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
• Link: http://creativecommons.org/licenses/by-nc-sa/4.0/
• Description: A Creative Commons license that bans commercial use and requires the user to release any modified works under this license.

Secondary language

• Language: Slovenian
• Title: Napoved okvar vodovodnih cevi s strojnim učenjem in podatki o sosednji infrastrukturi

Abstract

Okvare vodovodnih cevi v urbanih omrežjih so pomemben vzrok komercialnih izgub zaradi neobračunane vode, motenj v oskrbi in visokih stroškov vzdrževanja. Ta študija razvija model strojnega učenja za napoved verjetnosti okvar cevovodov in podporo strategijam vzdrževanja, temelječim na tveganju. Model, izurjen na podatkih o infrastrukturi in geoprostorskih podatkih iz obdobja 2010–2025, vključuje standardne lastnosti cevi – kot so material, starost, premer, vrsta omrežja in zgodovina vzdrževanja – ter prostorsko izpeljane kazalnike infrastrukture v neposredni soseščini. Posebej pomembno je, da model kvantificira napovedno vrednost sosednjih infrastrukturnih sistemov, vključno z električnim omrežjem, plinovodi, daljinskim ogrevanjem, kanalizacijo in cestnim omrežjem, z uporabo prostorske analize soseščine in tehnik prekrivanja. Več teh medinfrastrukturnih značilnosti, zlasti kategorija ceste, napetost električnega omrežja in tip kanalizacije, je pokazalo pomemben napovedni vpliv, kar odraža njihovo posredno, a dosledno povezavo z verjetnostjo okvare cevi. Model strojnega učenja, zasnovan z algoritmom XGBoost in validiran s slojevitim navzkrižnim preverjanjem (K-fold), je dosegel visoko zmogljivost (ROC AUC: 0,9102; priklic: 0,7750; natančnost: 0,8750). Kljub nižji preciznosti zaradi neravnovesja razredov rezultat F1 (0,2261) in LogLoss (0,2500) potrjujeta njegovo zanesljivost. Raziskava predstavlja nov, s prostorskimi podatki obogaten pristop k napovedovanju okvar in prispeva k naprednemu, na podatkih temelječem upravljanju urbane infrastrukture

• Keywords: vodovodni sistemi, napovedovanje okvar cevovodov, strojno učenje, XGBoost, prostorska analiza, ocenjevanje stanja