Physics-informed neural network for hydraulic prediction in open-channel water transfer projects with sparse...

Open-channel water transfer projects play a crucial role in addressing regional water supply-demand imbalances, and real-time, comprehensive, and accurate acquisition of their hydrodynamic spatiotemporal evolution is essential for ensuring safety and efficiency of water conveyance and optimizing scheduling strategies. While hydraulic monitoring systems and numerical simulations are potential solutions, the former struggles to balance the number of monitoring points with cost constraints to achieve comprehensive and economically feasible measurements, and the latter requires clear boundary conditions and key parameters that often pose challenges in practical scenarios. This paper presents a Physics-Informed Neural Network (PINN)-based method applied to predicting hydraulic transients in open channels, incorporating sparse monitoring data and physical laws. Results from numerical simulations and field tests demonstrate that the PINN model can accurately predict open-channel hydrodynamics and Manning's coefficient in various operational scenarios, and is robust to sensor noise. A series of sensitivity analyses were conducted to investigate the effects of neural network structural parameters, as well as the number and location of monitoring points, thereby determining the optimal neural network structure and monitoring point configuration. The findings provide a foundation for real-time, comprehensive and accurate hydrodynamic predictions in open-channel transfer projects from sparse monitoring data.