HR-PIGNN A High Resolution Prediction Method for Urban Drainage Network: Combining Graph Neural Networks and Discrete Form Physics Informed Neural Networks

Abstract

This article presents a novel hybrid model combining data and mechanisms for high-resolution water level prediction in pipeline networks. The model utilizes graph convolutional neural networks to integrate network topology information for precise predictions and incorporates de Saint-Venant system equations through physics-driven neural networks. Compared to traditional data-driven, mechanism-driven, and hybrid methods, this model achieves 5-minute, 1-centimeter resolution predictions while maintaining computational efficiency and high accuracy. In an experimental drainage system in Suzhou, China, the model's RMSE for predicting water levels and pipeline flow rates is 0.014 and 0.012, respectively, with NSE values of 0.802 and 0.883. The model's computation time for 24 hours of data at 5-minute intervals is 0.981 seconds.