Source-linked AI summary
Multiple Leak Localization in Water Distribution Networks through Graph Neural Networks
Arvin Ajoodani, Sara Nazif, Pouria Ramazi
TL;DR
Leak localization remains challenging when multiple leaks occur, while prior approaches often assume a single or predefined leak count or enumerate restricted combinations. This paper uses a graph convolutional network to localize a variable number of simultaneous leaks without specifying the count in advance. It reports stronger localization performance than conventional classifiers across the evaluated networks and conditions.
Problem
Many leak-localization studies assume one leak, while multiple-leak methods often fix leak counts or candidate locations and may require costly combination enumeration.
Source: §1 · pp. 2–3Method
A graph convolutional network performs node-level multilabel classification to localize variable numbers of leaks without exhaustive combination training, using sparse pressure information.
Source: §1 · p. 4, Conclusion · p. 23Results
The GCN achieved consistently stronger leak-localization performance than MLP, SVM, RF, and KNN classifiers across the evaluated networks.
Source: Conclusion · p. 23Takeaways & Limitations
Results on Hanoi and NET3 support the framework’s practical applicability under variable demand, leak locations and magnitudes, leak counts, and sparse sensor coverage.
Source: Conclusion · p. 23Further work should investigate additional hydraulic information, such as flow measurements, and more sample-efficient training strategies.
Source: Conclusion · p. 24
Abstract
from arXiv · showhide
Leak localization in water distribution networks is challenging when multiple leaks occur simultaneously, particularly under sparse monitoring. Existing data-driven methods often assume a single or predefined number of leaks or require extensive enumeration of leak combinations. This study proposes a graph neural network (GNN) model that localizes a variable number of simultaneous leaks without specifying the leak count in advance or exhaustively generating every possible combination of leak locations for training. The model is evaluated on two benchmark networks of different sizes, the 31-node Hanoi network and the larger 92-node NET3 network, under variable demand, sparse pressure-sensor coverage, variable leak magnitudes, and varying numbers of simultaneous leaks. The GNN outperforms multilayer perceptron, support vector machine, K-nearest neighbors, and random forest classifiers by up to 12% and 20% in accuracy respectively in Hanoi and NET3 networks. Consistent performance across both networks supports the scalability of the proposed approach.
1. Introduction
Prior leak-localization methods often assumed one leak, a fixed leak count, or a limited set of leak-location combinations. This study proposes a graph-based approach for localizing a variable number of simultaneous leaks without requiring exhaustive combination generation.
1. Introduction
- Many data-driven localization studies assumed a single leak, despite the possibility of concurrent leaks in large networks [Mazaev et al. (2022); Sun et al. (2020); Lučin et al. (2021); Ahmad et al. (2023); Chen et al. (2026)].The cited discussion links concurrent-leak likelihood to infrastructure aging, urbanization, and climatic factors.Source: §1 · p. 2
- Prior multiple-leak studies addressed two zones or experimental networks, but often fixed leak counts or limited candidate locations [Rodríguez-Argote et al. (2023); Fang et al. (2019)].Enumerating every combination becomes computationally prohibitive as candidate nodes increase, and fixed-count models may not generalize to variable leakage conditions [Romero-Ben et al. (2023)].Source: §1 · pp. 2–3
- Graph neural networks incorporate network topology through message passing to learn topology-dependent hydraulic relationships relevant to leak localization [Garðarsson et al. (2022)].WDNs are modeled with consumption nodes and connecting pipes, and prior work has used GNNs for localization with limited sensor data [Rolle et al. (2024)].Source: §1 · p. 3
- The proposed model performs node-level multilabel classification, allowing leak counts to vary without exhaustive training enumeration of leak-location combinations.The passage also states that the leak count need not be specified in advance.Source: §1 · p. 4
- The framework is evaluated under variable demand, sparse sensor coverage, and variable leak magnitudes on Hanoi and NET3, alongside KNN, SVM, MLP, and RF classifiers.The introduction names these benchmark networks and comparison classifiers but does not report comparative results here.Source: §1 · p. 5
2. Methodology
The methodology generates multi-leak hydraulic scenarios, reconstructs network-wide inputs from sparse pressure sensing, and classifies node-level leakage with a topology-aware GCN. It also compares the GCN with conventional classifiers across two benchmark networks.
2.1 Data Generation
- Scenarios varied leak count and locations, sampled a time step, perturbed demand by 0.8–1.2, and assigned each leak a random emitter coefficient from 0 < C ≤ 10.Leak duration was not investigated.Source: §2.1 · pp. 5–7
2.2 Input and Target Construction
- Sparse sensor readings were graph-interpolated into network-wide pressure inputs, with one binary target label for each leak-candidate node.Reconstruction used hydraulic head to account for node elevation.Source: §2.2 · pp. 8–9
2.3 Graph Convolutional Network
- The GCN classified each candidate node using its pressure-related feature and pipe connectivity, aggregating information from neighboring nodes across successive layers.Its node-level outputs allow multiple nodes to be identified as leaking in one scenario.Source: §2.3 · p. 10
2.4 Conventional Classifiers
- MLP, SVM, KNN, and RF provided comparison classifiers, each producing a binary prediction per leak-candidate node from the same pressure-related inputs.Source: §2.4 · p. 11
2.5 Model Training and Evaluation
- Models used binary cross-entropy and Adam; an 80/20 development-test split and five-fold cross-validation selected hyperparameters, with mean validation F1-score guiding selection.Training stopped when loss improved by less than 1% over 10 consecutive iterations; dataset growth stopped after two increments each improved mean cross-validation F1-score by less than 0.01.Source: §2.5 · pp. 12–13
2.6 Benchmark Networks and Experimental Setup
- The framework was tested on Hanoi, with 31 nodes and four pressure sensors, and NET3, with 92 nodes and eleven sensors.Hanoi has 34 pipes and one reservoir; NET3 has 117 pipes, two reservoirs, and three tanks.Source: §2.6 · p. 15
3. Results and Discussion
The GCN outperformed conventional classifiers across evaluated metrics, and node-level accuracy increased as distance to pressure sensors decreased. This pattern held across the Hanoi and NET3 networks.
3.1 Model Performance and Effect of Dataset Size
- The GCN outperformed RF, MLP, SVM, and KNN across evaluated metrics.The comparison used the same reconstructed pressure-related features, while the GCN additionally incorporated topology through graph-based neighborhood aggregation.Source: Figure 5 · p. 18, §3.1 · p. 19, Figure 6 · p. 19
- A small sampled fraction of possible leak-location combinations supported effective localization and scalability across the tested network sizes.The scenario-space comparison reports approximately 8.5 × 10−22% coverage for about 42,000 NET3 scenarios.Source: §3.1 · p. 20
- All models performed worse on larger, more complex NET3 than on Hanoi, but the GCN advantage over conventional classifiers became more pronounced.The authors suggest topology-aware learning provides greater benefits as network size and complexity increase.Source: §3.1 · p. 20
3.2 Node-Level Accuracy and Sensor Proximity
- Hanoi’s overall accuracy was 0.92, and NET3’s was 0.82; node-level accuracy increased as average shortest-path distance to sensors decreased in both networks.Hanoi node-level accuracy ranged from 0.84 to 0.97; the NET3 analysis reported the same distance-accuracy pattern. ASPD intervals were chosen to balance node counts across distance groups.Source: §3.2 · pp. 20, 21, 23, Figure 7 · p. 21, Figure 8 · p. 22
Conclusion
The GCN framework localizes variable numbers of simultaneous leaks under sparse monitoring without exhaustive combination generation. Evaluations across varied conditions in Hanoi and NET3 support practical applicability within the study’s tested settings.
Conclusion
- The framework localizes a variable number of simultaneous leaks under sparse monitoring, using graph convolution to incorporate spatial relationships alongside hydraulic information.It does not require the leak count to be specified in advance and can train on sampled rather than exhaustive leakage scenarios.Source: Conclusion · p. 23
- Evaluations varied demand, leak locations and magnitudes, leak counts, and pressure-sensor coverage; results on Hanoi and NET3 support applicability under those conditions.The authors describe future work on additional hydraulic information, including flow measurements, and more sample-efficient training to improve performance and reduce data requirements.Source: Conclusion · pp. 23–24