220 - Graph Neural Networks-Based Water Quality State Estimation in Water Distribution Networks

Maintaining a sufficient chlorine residual throughout water distribution networks (WDN) is of utmost importance for protecting public health. Ubiquitous monitoring of chlorine concentrations (CCs) throughout WDNs is inhibited by the vast scale of these systems and the excessive cost of deploying and operating water quality sensors. In this study, we introduce a Graph Neural Networks (GNN)-based framework for water quality state estimation (WQSE) to improve WDN water quality observability. The developed algorithm aims to estimate chlorine concentrations at all network junctions relying solely on the data collected by a limited number of sensors, eliminating the necessity for an extensive sensor network. Two different GNN WQSE models are introduced, namely, Static Prediction (SP), and Dynamic Prediction (DP), in which water quality predictions are made based on fixed and mobile sensor measurements, respectively, without relying on additional network information. The two GNN models were applied to a benchmark system to demonstrate their performance in predicting CC under various scenarios. Preliminary results showed the ability of the proposed GNN models to learn the underlying relationships between sensor and non-sensor junctions, resulting in accurate CC predictions for intermediate junctions. The results also revealed a correlation between the prediction error and the distance from the sensor. More importantly, the GNN model was shown to be robust against measurement noise, while being significantly dependent on the input sensor design.