Neural network-based leak localization in water distribution networks using the gravity center of pressure measurements
A novel methodology for leak diagnosis in urban water distribution systems (WDS) is proposed. Small leaks are simulated using a well-calibrated EPANET model of the WDS. Considering only the known topology of the WDS, and pressure head values recorded at some nodes, the center of gravity of pressure...
| Autores: | , , , , |
|---|---|
| Tipo de recurso: | artículo |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/456424 |
| Acceso en línea: | https://hdl.handle.net/2117/456424 https://dx.doi.org/10.1016/j.jwpe.2025.108348 |
| Access Level: | acceso embargado |
| Palabra clave: | Leak localization Neural network LSTM Deep learning Urban water management Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic |
| Sumario: | A novel methodology for leak diagnosis in urban water distribution systems (WDS) is proposed. Small leaks are simulated using a well-calibrated EPANET model of the WDS. Considering only the known topology of the WDS, and pressure head values recorded at some nodes, the center of gravity of pressure is computed. Under nominal (leak-free) operation the position of the center of gravity varies predictably, but leaks cause variations on its position. Sensor-measurements with a duration of 24 h are used to compute residual coordinates from leak-free operation and used to train a LSTM neural network implemented in MATLAB for leak classification. Results are presented for the leak localization task considering two levels of resolution: identifying the general sector and pinpointing the specific node where the leak occurs. Tests are performed on a benchmark and real-world WDS obtaining a good performance with simulated data under steady-state and variable demand conditions. The impact of measurement noise is addressed by including the measured outflow from the reservoir as a third dimension to the training data. |
|---|