EPANET, the EPA's open-source hydraulic simulation engine first released in 1993, models water distribution networks as systems of pipes, junctions, tanks, pumps, and valves — computing pressure, flow, and water quality parameters at each timestep using the gradient method for hydraulic solution. For three decades, EPANET has been the industry standard for water distribution modelling, but its application has been largely limited to planning and design scenarios rather than real-time operational intelligence.
The convergence of SCADA telemetry streams, cloud computing, and graph neural network architectures has enabled a fundamentally new capability: the EPANET digital twin — a continuously calibrated hydraulic model that runs in parallel with the physical network, updated in real-time from SCADA pressure and flow measurements, and augmented with ML-driven anomaly detection that identifies deviations between modelled and observed hydraulic behaviour.
Graph Neural Networks (GNNs) are particularly well-suited to pipe network anomaly detection because water distribution systems are, by definition, graph-structured data — nodes (junctions, tanks, pumps) connected by edges (pipes). A GNN trained on historical pressure and flow telemetry learns the normal correlation patterns between measurement points across the network topology, enabling it to detect localised anomalies (pipe bursts, unauthorised withdrawals, sensor failures) as spatial pattern deviations rather than simple threshold exceedances.
Independent validation studies on datasets from UK water companies and Australian utilities have demonstrated that GNN-based pipe burst detection achieves true positive rates of 91–96% at false positive rates below 5% — compared to false positive rates of 35–60% for traditional pressure zone monitoring approaches. The practical operational impact is significant: at 5% false positive rate, a network with 1,000 pressure sensors generates approximately 2 false alarms per week rather than 350.
The Smart Water Platform's Hydraulic Digital Twin module implements EPANET-2.2 as the hydraulic simulation backbone, integrates SCADA telemetry for continuous model calibration using the Kalman Filter ensemble method, and applies a spatiotemporal GNN for real-time anomaly scoring. Detected anomalies are georeferenced against the GIS pipe network registry and trigger automated field investigation work orders with GPS routing to the predicted burst location.