Accurate and rapid prediction of water levels in urban drainage systems remains a key challenge due to sparse sensor coverage, complex hydraulic interactions, and uncertain rainfall inputs. Traditional hydrodynamic models often rely on detailed rainfall fields and high-resolution calibration, limiting their scalability. Data-driven approaches, in contrast, struggle to generalize under observational sparsity or unseen conditions. This study proposes a knowledge-based deep learning model that integrates prior knowledge with sparse real-time monitoring data to achieve accurate and robust water level prediction. The model propagates observational signals through network pathways, enabling correction across space and time without requiring explicit rainfall inputs. Comprehensive evaluations using synthetic and real rainfall events demonstrate that the model consistently achieves sub-decimeter accuracy across most nodes. Under extreme sensor sparsity, the model retains mean NSE values above 0.90. In real rainfall applications, peak prediction errors remain within 0.05 m, and overflow detection achieves high accuracy. Furthermore, spatial residuals align with known hydraulic sensitivities, enabling physically meaningful interpretation and guiding sensor deployment strategy. Compared with the data-driven model, the proposed model delivers more stable and accuracy results. Its generalization ability, low input requirement, and structural alignment support its deployment for real-time urban flood prediction and early warning. This study offers a scalable solution for smart drainage system management.
Keywords: Graph neural network; Knowledge-based model; Real-time monitoring data; Urban drainage system; Urban flood.
Copyright © 2026 Elsevier Ltd. All rights reserved.