Water Pressure Prediction in a Water Supply Network Using the Informer Framework

Abstract

Water pressure scheduling is important for the normal operation of urban water supply networks, and water pressure prediction provides the basis and guidance for water pressure scheduling. Informer, as a deep learning model proposed in recent years, brings a new approach to the study of water pressure prediction. In this study, Informer is used to model the water pressure data of the water supply network to achieve long-term water pressure prediction. In order to address the challenge of hyperparameter optimization in deep learning, the Bayesian and HyperBand (BOHB) hyperparameter optimization algorithm, combining BO and HB, is utilized to construct a water pressure prediction framework for automatic hyperparameter optimization with BOHB-Informer. Through experiments on real urban water supply network cases, Informer has 22.5%∼43% lower error than the recurrent neural network (RNN) model in long-term water pressure prediction, and the prediction error of the Informer model optimized by the BOHB-Informer framework is further reduced by 23.7%∼34.1%. The experimental results demonstrate that the framework exhibits excellent parameter tuning performance and prediction accuracy. It can be a valuable auxiliary guidance tool for water pressure scheduling in water supply networks.