An Ensemble Learning Approach for Leakage Localization in Water Distribution Networks

Reliable and explainable leakage localization in water distribution networks (WDNs) is critical for minimizing non-revenue water, reducing inspection time, and improving the operational resilience of WDNs. Explainability is particularly important because leakage localization results directly inform high-cost and high-risk operational decisions, such as field inspections and pipe excavations, and must therefore be transparent and trustworthy to system operators. Machine Learning (ML) approaches have recently shown strong potential for leakage localization, which commonly formulated as a multi-class classification problem. In this setting, a WDN is first partitioned into different zones, and the ML model is then trained to predict the most likely zone containing the leakage. In previous work, four high-performing tree-based classifiers, including Random Forest, Gradient Boosting, XGBoost, and LightGBM, and three neural network models of increasing architectural depth (shallow NN, deep NN, and extra-deep NN) were trained and comparatively evaluated. While all implemented ML models performed well individually when the number of classes was small, their performance degraded to varying degrees as the number of classes increased. However, the differential performance across algorithms suggests potential for ensemble methods to highlight their complementary strengths. This work proposes an explainable ensemble ML framework for multi-class leak zone classification using pressure measurements that systematically combines the outputs from different ML models, thus strengthening the robustness beyond individual approaches. Building on prior evaluations of individual models, different classifiers are evaluated by integrating the outputs of multiple models using three complementary ensemble strategies: majority voting, which combines discrete leak localisation decisions; weighted averaging, which assigns reliability-based weights to individual model predictions; and stacking, where a meta-model is trained to learn how to optimally combine the outputs of several base classifiers.

Model performance is evaluated through multiple metrics including classification accuracy, precision, recall, and F1-scores, complemented by confusion matrix analysis and computational efficiency measurements. Additionally, a new metric, called Maximum Pipe Length Search (MPLS) is applied, which provides a physically interpretable measure of inspection effort on-side. MPLS quantifies the cumulative pipe length that would be inspected if operators followed the model-generated ranking of likely leakage zones until reaching the correct one. This metric bridges model predictions with actionable field strategies, offering a practical lens for utilities to compare model outputs in operational terms. This research investigates whether ensemble approaches can provide key advantages in the context of leakage localisation, including increased robustness to noisy sensor measurements, mitigation of the limitations of individual models, and improved generalisability across varying hydraulic and operational conditions.

Funding

This publication was produced as part of the "FOUND" project. This project is funded by the Federal Ministry of Agriculture and Forestry, Climate and Environmental Protection, Regions and Water Management (BMLUK) (Austria) (Project C300198).