A Spatial-Temporal Approach to Characterise WDSs Behaviour Against Water Leakage Events

Water distribution systems (WDSs) are essential water resources infrastructure increasingly challenged by recurring anomalies such as leaks and bursts, which can compromise reliability and exacerbate extreme events like flooding or water scarcity. Effective management of leakage requires timely characterisation of system behaviour to reduce water loss, enhance resilience, and support operational decision-making. Traditional approaches, relying on hydraulic models, flow measurements, or demand estimation, are computationally intensive and unsuitable for real-time application, while pressure-based alternatives face limitations in handling indicator assumptions, overlapping leaks, and spatial–temporal dynamics. This thesis proposes a pressure-only, model-free framework that reconstructs WDS behaviour before, during, and after leakage events. The methodology integrates temporal and spatial representations of system deviation: a residual pressure matrix compares observed data against a baseline from a no-leak period, a network-level indicator quantifies temporal deviations, and residual heatmaps visualise spatial impacts. These features address the thesis objectives: capturing deviations efficiently, improving interpretability of anomalies, detecting leaks without hydraulic calibration or flow measurements, and demonstrating robustness across multiple scenarios. The methodological choices in this thesis were guided by the need for a solution that can operate under the practical limitations of real-world networks. Pressure measurements are typically the most widely available and reliable sensor data in WDSs, making a pressure-only strategy more realistic and scalable than approaches dependent on flowmeters or hydraulic calibration. A residual-pressure formulation was selected because it allows deviations to be quantified without relying on model assumptions, and the chosen filtering steps (Fourier decomposition and Savitzky–Golay smoothing) provided the most effective balance between noise reduction and signal preservation. Together, these design choices support a framework that is computationally efficient, easily deployable, and adaptable across different network configurations. Application to the L-Town benchmark network demonstrates that the framework can capture the onset, escalation, and recovery of multiple leak events, including overlapping anomalies, without requiring hydraulic calibration or flow measurements. Overall, the proposed approach provides a scalable, interpretable, and computationally efficient tool that enhances operational practicality, supports resilience planning, and lays the groundwork for real-time decision-support in water network management.