Author(s): Sanghoon Jun
Linked Author(s):
Keywords: Advanced metering infrastructure; Artificial intelligence; Convolutional neural network; Leakage; Smart meter; Water distribution network
Abstract: This presentation evaluates convolutional neural network-based leak detection model over several water distribution systems (WDSs) that have distinct network characteristics such as pipe diameter distribution and supplied demands. The model identifies leaks by examining spatially distributed images of pressure responses (i.e., differences between measured and estimated pressures) generated by comparing advanced metering infrastructure observations with predictions from a well-calibrated hydraulic model. For each WDS, ten different leak magnitudes are tested to determine the detectable leak sizes for varying systems. Three performance indicators, recall, precision and F1 score are used to quantify the detectability. In the results, the important features of WDSs that impact detectable leak sizes are demonstrated.
Year: 2026