Forecasting of water consumption by integrating spatial and temporal characteristics of short-term water use in cities

Abstract

Urban water consumption prediction is an important basis for urban water resource management and optimal scheduling of water supply. Most existing prediction methods disregard the spatial–temporal characteristics of water consumption data and utilize incomplete data mining, resulting in prediction models with low accuracy. To accurately predict short-term water demand, a data-cleaning algorithm based on statistical distributions and velocity constraints is proposed in this study. The long- and short-term time-series network (LSTNet), automated spatial–temporal graph prediction (AutoSTG), and attention-based spatial–temporal graphical convolutional neural network (ASTGCN) deep learning models were investigated for automatically capturing the spatial–temporal dynamic features. Additionally, the short-term predictions for 1, 3, 6, 12, and 24 h forecasting periods were investigated based on the water consumption data of water plants in Tianjin. The results revealed that the proposed data-cleaning algorithm can effectively manage single-point anomalies and continuous outliers, thus improving prediction accuracy. The LSTNet, AutoSTG, and ASTGCN deep learning models, particularly the ASTGCN model based on the spatial–temporal attention mechanism, significantly improved the prediction accuracy and adaptive spatial–temporal feature extraction. The final prediction model that combines data-cleaning algorithms with the ASTGCN deep learning method achieved root mean square error, mean absolute error, accuracy, and relative squared error values of 566.4 m3, 377.7 m3, 0.964, and 0.112, respectively. The proposed forecasting framework can improve short-term water use forecasting at multiple sites in the city and provide support for fine-grained water allocation.