Spatiotemporal graph convolutional network using sparse monitoring data for accurate water-level reconstruction in urban drainage systems

Abstract

A comprehensive monitoring of urban drainage network (UDN) is essential for maintenance, management, and sustainable urban development. However, limited sensor deployment hinders the acquisition of sufficient information. Conventional deep learning methodologies can predict and correct monitored data but struggle with unobserved data. Hydraulic models can simulate behaviors but face data collection challenges and low real-time performance. To address these issues, a novel spatiotemporal graph convolutional network (STGCN) model, based on graph neural networks, is proposed to reconstruct a real-time information system for UDNs. By extracting fundamental elements from limited monitoring data and UDN topology, the STGCN model effectively reconstructed unmonitored node data. The experimental results showed that the training efficiency and reconstruction accuracy of the model could be optimized by reducing the spatial data dimensionality to 0.6, adopting a passive-masked training strategy with a ratio of 4:3 for model-training sensors to loss-calculation sensors, and using a historical data input length of 3 h. This approach allowed for the reconstruction of water levels for 527 unmonitored nodes using only seven monitoring nodes, with a median mean absolute error of 0.038 m and an accuracy of 71.3 %. These results demonstrate that the STGCN model can accurately reconstruct unmonitored node data using low monitoring-node density and basic network topology, offering a practical solution to data-driven challenges in intelligent UDNs. The source code is available at https://github.com/holylove9412/UDNs_STGCN_model.