A Hybrid Framework and Sensitivity Analysis to Enhance Usability and Robustness of Leak Detection and Localization in Water Distribution Networks

Abstract Submission: Leakages in water distribution networks pose substantial challenges to water utilities, resulting in water and revenue losses, and other potential cascading effects. Over the years, various model-based and data-driven methods have been proposed in scientific literature and practical applications for leak identification and localization. Model-based approaches identify and localize leaks by comparing the simulation output from calibrated hydraulic models with sensor data collected in the real water distribution network. Data-driven approaches identify leaks only based on analysis of flow-pressure data using time series analysis and other mathematical models. Both methodological approaches have trade-offs between information requirements and expected performance: data-driven methods do not require a well-calibrated hydraulic model, but this advantage comes with the cost that they are less suitable than model-based approaches for leakage localization, as they do not inherently consider the spatial characteristics of water distribution networks. Here we present a hybrid method that combines the strengths of both approaches with the aim of enhancing robustness, and usability of leak identification and localization methods. We enhance usability by combining both model-based and data-driven leak detection and localization methods in a common framework. We then assess robustness with a hierarchical global sensitivity analysis using Sobol’ indices to determine the influence of each source and combinations of aleatoric (e.g., leak scenarios and network topologies) and epistemic uncertainties (e.g., data quality, hydraulic model parameters, detection method hyperparameters). Results obtained by testing our framework on data from different networks and different scenarios allow formulating recommendations for practical use by water utilities.

Learning Objectives/Expected Outcome (Optional) :