Leak detection and localization in water distribution networks using the mixture of experts model

Leakage issues in water distribution networks pose significant challenges, and mitigating water loss is critical for ensuring a sustainable and resilient water supply system. Intelligent detection methods make effective use of existing sensor data for the reliable leak identification across the network. Advanced leakage detection and localization schemes employ a hybrid approach combining hydraulic modeling with data-driven technologies. However, these efforts predominantly focus on detection and localization using single models. In this study, an integration of multiple models for enhanced leak detection and localization, which initially trains each individual model and the gating model and subsequently outputting detection results by weighting each model through the gating mechanism within the Mixture of Experts (MoEs) framework, was proposed. To assess the performance of the proposed methodology, experiments with datasets derived from the LeakDB benchmark, including Hanoi water network data and Net2 water network data, were conducted. The hydraulic model results indicate that leak detection accuracy is consistently above 90%, underscoring the stability of the MoE. This helps not only to minimize water losses and related energy costs of water treatment but also to guide water departments toward a more sustainable and resilient future.