This study examines the benefits and limitations of a convolutional neural network (CNN) burst detection model that accounts for spatially distributed information of pressure responses in a water distribution system (WDS), i.e., the differences between measured and predicted pressure data. To that end, a 2D CNN is applied to a smart WDS where all pressures and advanced metering infrastructure (AMI) end-user demands are measured. Here, a well-calibrated hydraulic model for a WDS in Austin, TX is analyzed with measured AMI demands to predict pressure surfaces that are provided to a CNN. Alternative image data structures are examined to evaluate their importance and two different data types, raw pressure data and pressure responses, are evaluated to investigate the benefits of linking CNN with hydraulic information. In addition, the effect of field measurement errors on detection results is examined for a range of error magnitudes. Finally, burst detection results of partial and full pressure meters are assessed to study the benefits of pressure supplemented AMI systems. Based on the numerical results, several conclusions are posed. First, network layout information should be incorporated into the image data structure. In addition, CNN should incorporate hydraulic information within AMI demands rather than using raw pressure data. Lastly, large measurement errors can mask the impact of small bursts and SCADA systems are insufficient to detect these failures. Thus, pressure supplemented AMI systems are recommended.
- Advanced metering infrastructure
- Convolutional neural network
- Hydraulic model
- Water distribution system
ASJC Scopus subject areas
- Civil and Structural Engineering
- Water Science and Technology