Leak Detection under Sparse Instrumentation: Implications for Data-Driven Methods in Water Distribution Systems

Despite accelerating digitalization and the wide availability of artificial intelligence tools, many water distribution systems (WDS) still lack system-scale customer monitoring, advanced metering infrastructures, or up-to-date calibrated hydraulic models. For the majority of such systems, leak detection remains predominantly data-driven, relying on time series from a sparse set of pressure sensors and flow meters, the placement of which is typically determined by expert knowledge. Sensor data may be augmented by simplified network graphs or uncalibrated hydraulic models, yet insights remain limited and the problem highly imbalanced, as the small number of (engineered) leak events restricts both model training and evaluation.

Under these challenging conditions, combinations of time series analysis and machine learning models have shown strong potential for automated, data-driven leak detection. However, the amount, quality, and structure of data required for robust model development and evaluation can hinder practical implementation. A limited number of devices, operational constraints, and environmental risks often lead to temporary installations and repeated sensor relocation, further reducing the availability of consistent training data.

This work investigates the potential of hybrid modeling approaches, transfer learning strategies, and the encoding of distribution system structure for data-driven leak detection under these constraints. Data quality and temporal granularity are examined at the sensor level. By mapping edge computing concepts to structural units inherent to WDS (e.g., district metered areas), the performance of anomaly detection algorithms is evaluated across different sensor combinations and spatial scales.

Model development and evaluation are based on high-resolution hydraulic (pressure and flow) time series and operational data collected over three years in an operational WDS. Beyond assessing the proposed methodologies, this study enables an in-depth discussion of the limitations of data-driven leak detection under conditions of incomplete instrumentation and expert sensor placement.

Acknowledgements: The data for this work was generated during research project “ADAM - Advanced data-driven modeling of water distribution systems” funded by Vienna Water.