Enhanced Prediction of Water Pipeline Failures Using a Novel MTGNN-Based Classification Model for Smart City Infrastructure

Accurate prediction of waterpipe failures is critical for urban infrastructure resilience and operational efficiency, especially as smart city initiatives increasingly rely on data-driven approaches to manage infrastructure assets. Traditional machine learning models for predicting pipeline failures have limitations, often failing to capture the spatial dependencies intrinsic to interconnected water networks and falling short on short-term prediction accuracy. This study introduces a novel Multi-Task Graph Neural Network (MTGNN)-based classification method specifically designed to address these limitations by leveraging spatial and temporal dependencies. Using the Kitchener waterpipe failure dataset, the MTGNN-driven approach achieves precision rates of 0.69, 0.98, and 0.95 for six-month, one-year, and two-year prediction horizons, respectively. This improved accuracy highlights the model's potential to enhance proactive maintenance strategies and infrastructure sustainability in digital city environments. This work is the first to employ MTGNN for waterpipe failure prediction, offering a methodological advancement in short-term prediction and introducing a binary classification model that improves prediction precision across multiple time horizons.