Dual-stream graph convolutional networks enable accurate real-time chlorine prediction in sparsely monitored water...

Accurate real-time water quality prediction at unmonitored locations within water distribution networks (WDNs) is crucial yet challenging due to the inherent sparsity of monitoring sensors. This study introduces a novel Dual-stream Graph Convolutional Network (DGCN), explicitly integrating both upstream and downstream flow dynamics within WDNs. By training exclusively on data from sensor-equipped nodes (33 sensors) in a case study, the proposed model accurately predict chlorine residuals at unmonitored nodes, achieving mean absolute percentage errors (MAPE) below 5 % for 78 % of nodes (703/955) and MAPE <20 % for 98.7 % of network nodes. This model consistently outperforms benchmark Graph Convolutional Networks, particularly at locations distant from sensor nodes. The inclusion of a masking strategy during training further improves model robustness, making it effective in simulating real-world scenarios of limited sensor deployment. This study illustrates the DGCN model has a capability of accurately predicting water quality across a WDN with sparse observational data, holding promise for significantly improving real-time water quality optimisation and management with fast simulations.