Graph Convolutional Neural Network for Pressure Prediction in Water Distribution Network Sites

The safe operation of water distribution networks (WDNs) is crucial for ensuring the city dwellers' living standards. Accurate and multi-step predictions of pressure at key sites in WDNs can prevent the occurrence of pipe bursts in the future. Therefore, this study proposes an EMD-Graph-Wavenet-HGSRS model to predict the pressure at several monitoring sites in the WDNs. The LSTC-Tubal method is proposed to repair the abnormal pressure values of the WDNs. Then, the pressure features are enriched by EMD. The predefined adjacent matrix of monitoring points is obtained through the topology of WNDs. And, the enriched pressure features and the predefined adjacent matrix of the monitoring sites are input into the Graph-Wavenet model to predict the pressure values for the next 12 h. In addition, the Graph-Wavenet model is optimized by HGSRS in this study. The results of this study show that the MAE of EMD-Graph-Wavenet decreased by 24.36%, KGE increased by 6.73% compared to Graph-Wavenet. EMD-Graph-Wavenet-HGSRS (optimized by HGSRS) prediction outperforms EMD-Graph-Wavenet model. The MAE of Graph-Wavenet decreased by 40.91% and KGE increased by 11.91% compared to Bi-LSTM. The Bi-LSTM exhibited the best performance among these temporal models, whereas the baseline LSTM had the worst performance. The method proposed in this study can better predict the pressure extremes at each stage of the monitoring sites and provide guidance for the pressure management of actual WDNs.