GCN-GRU: Multi step prediction model for urban water consumption by integrating spatiotemporal graph convolution and multi head attention mechanism

Abstract

Urban water resource planning as well as the supply-demand balance are directly affected by immediate demand forecasts, making them vital for water resource management.   The present approaches to water demand forecasting just consider temporal variables and do not take account of the possible impact of geographical characteristics on them.  The reason for this is because short-term urban water demand forecasts are affected by several variables, many of which display complex nonlinear dynamic features.   Predictions end up being inaccurate because of this.   This research aims to address this problem by presenting a model that considers geographical and temporal characteristics in order to predict urban water consumption in the near term.   The first step is to detect and fix anomalies using the Prophet model.   In order to generate an adjacency matrix among variables, we use a maximum information coefficient, and to extract spatial attributes between variables, we use a graph convolutional neural network.  Afterwards, a multi-head attention method is used to enhance crucial aspects of water consumption statistics while reducing the influence of unimportant components.   The next phase involves projecting urban areas' short-term water demands using a three-layer long short-term memory system.   This study's proposed hybrid model outperforms state-of-the-art prediction methods in terms of accuracy and efficiency, with an average percentage absolute error reduction of 1.868-2.718%.   Not only does this study set the stage for future research, it also has the potential to aid cities in making better use of their water resources.