Journal Article

Graph Neural Networks for Pressure Estimation in Water Distribution Systems

Lazovik, Alexander, Degeler, Victoria, Truong, Huy, Tello, Andrés

2024

Overview

Pressure and flow estimation in water distribution networks (WDNs) allows water management companies to optimize their control operations. For many years, mathematical simulation tools have been the most common approach to reconstructing an estimate of the WDNs hydraulics. However, pure physics‐based simulations involve several challenges, for example, partially observable data, high uncertainty, and extensive manual calibration. Thus, data‐driven approaches have gained traction to overcome such limitations. In this work, we combine physics‐based modeling and graph neural networks (GNN), a data‐driven approach, to address the pressure estimation problem. Our work has two main contributions. First, a training strategy that relies on random sensor placement making our GNN‐based estimation model robust to unexpected sensor location changes. Second, a realistic evaluation protocol that considers real temporal patterns and noise injection to mimic the uncertainties intrinsic to real‐world scenarios. As a result, a new state‐of‐the‐art model, GAT with Residual Connections, for pressure estimation is available. Our model surpasses the performance of previous studies on several WDNs benchmarks, showing a reduction of absolute error of ≈40% on average. Plain Language Summary Water management practitioners have resorted to mathematical simulation tools to reconstruct pressure, flow, and demand in order to improve their control operations. However, pure physics‐based methods need to deal with partially observable data, high uncertainty, and extensive manual calibration. We combine physics‐based modeling and graph neural networks, a data‐driven approach, to address the pressure estimation problem and overcome those limitations. Our work has two main contributions. First, a random sensor placement strategy makes our estimation model resilient to unexpected sensor location changes. Second, a realistic evaluation protocol that considers real temporal patterns and noise injection to mimic the uncertainties of real‐world scenarios. As a result, a new state‐of‐the‐art model, GAT with Residual Connections, for pressure estimation is available. Our model surpasses the performance of previous studies on several water distribution networks benchmarks, showing a reduction of absolute error of ≈40% on average. Key Points Mathematical Simulation Tools and graph neural networks were combined for pressure estimation in water distribution networks Random sensor placement during model training is a good strategy for robustness against unexpected sensors' location changes Time‐dependent patterns and Gaussian noise injection enable a realistic evaluation protocol for pressure estimation models

• Publisher: John Wiley & Sons, Inc,Wiley
• Subject: Benchmarks / Calibration / Error reduction / Flow estimation / Fluid flow / fluid mechanics / Graph neural networks / Hydraulics / Injection / Mathematical models / Modelling / Neural networks / Physics / Placement / Pressure / Pressure estimation / protocols / Sensors / state estimation / Traction / Uncertainty / water / Water distribution / water distribution networks / Water distribution systems / Water engineering / Water management