High Fidelity Digital Twin-Based Anomaly Detection and Localization for Smart Water Grid Operation Management.

• Developed a generic digital twin (DT) framework for smart water grids (SWG). • Integrated data analytics and hydraulic model calibration for constructing high-fidelity SWG DT. • Automated model training including data cleansing, outlier detection and event classification. • Applied DT-based approach to three supply zone with pipeline length of more than 1000 km. • Achieved event detection accuracy of greater than 80% with lead time of 1 to 3 days. Smart Water Grid (SWG) plays a critical role in sustaining cities economic and social development, but challenges remain in fully realizing the benefits of SWG. While Digital twin (DT) has been discussed in some literature for possible SWG applications, there has been limited, or no technical framework developed to facilitate SWG operation and management. In this paper, a generic framework is developed for constructing SWG high fidelity Digital Twin (DT) by integrating digital thread with various digital models for visualization, data-driven analysis, physics-based simulations, and decision-making support. Both the physics-based models and data-driven models are trained/retrained and calibrated/re-calibrated respectively by using the data collected with the sensors installed throughout a SWG. The information derived from the SWG DT can be diagnostic, predictive, and prescriptive to significantly augment users' intelligence for improving SWG operation and management. One important application of digital twin augmented intelligence is illustrated to timely detect and localize anomaly events, which may include, but not be limited to, pipe bursts and unauthorized water usages. The solution is tested on the selected areas in Singapore to construct the ever-green DT by calibrating and recalibrating the models for near real-time SWG operation management. The case study was conducted for three supply zones with the total pipeline length of more than 1000 km and 40 weeks of monitoring data, collected by 89 pressure monitoring stations and 8 flow meters at inlets and boundary. More than 3300 data-driven models are trained for optimizing the model performance to achieve accuracy of greater than 80% F1 score for detecting anomaly events, which are subsequently localized within 400 m with 2–3 days lead time. [ABSTRACT FROM AUTHOR]