A graph neural network using physical attributes to improve the system-wide nodal water-level prediction in sparsely monitored urban drainage systems

Abstract

Nodal water levels are a critical hydraulic parameter indicative of the operational status of urban drainage networks (UDNs), and their system-wide sensing is essential for evaluating system capacity and promptly identifying urban flooding and overflow pollution risks. However, due to financial constraints and installation challenges, the widespread deployment of sensors in UDNs is impractical. While developing approaches based on graph neural networks for system-wide sensing and prediction in sparsely monitored drainage systems is an effective solution, methods that rely solely on simple topological connectivity exhibit instability and significant prediction errors due to the complex and variable flow conditions within UDNs, influenced by multiple uncertainties. To address this challenge, we propose an edge-attribute-enhanced spatiotemporal graph convolutional network (Edge-STGCN) to improve prediction accuracy in sparsely monitored UDNs, offer a novel perspective to evaluate sensor placement strategies in different branches, and analyze the predictive utility of individual and combined edge attributes for model performance. Results revealed that with only 10% of nodes monitored, the Edge-STGCN model achieved reliable predictions at 88.6% of system-wide nodes, significantly outperforming the multilayer perceptron (14.5%) and STGCN (47.0%). Pipes near the outfall, particularly small-diameter branches whose invert elevations were higher than those of the main trunk pipes, were especially prone to uncertainty in prediction accuracy. Pipe invert elevation was an important contributor to the model’s prediction accuracy. The proposed method enables reliable predictions, informs sensor placement strategies, and provides data support for decision-making aimed at mitigating flooding and pollution.