A Machine-Learning Approach for Monitoring Water Distribution Networks (WDNs)

Abstract

The knowledge of the simultaneous nodal pressure values in a water distribution network (WDN) can favor its correct management, with advantages for both water utilities and end users, and guarantee higher sustainability in the use of the water resource. However, monitoring pressure in all the nodes is not feasible, so it can be useful to develop methods that allow us to estimate the whole pressure field based on data from a limited number of nodes. For this purpose, the work employed an artificial neural network (ANN) as a machine-learning regression algorithm. Uncertainty of water demand is modeled through scaling laws, linking demand statistics to the number of users served by each node. Three groups of demand scenarios are generated by using a Latin Hypercube Random Sampling with three different cross-correlations matrices of the nodal demands. Each of the corresponding groups of pressure scenarios is employed for the training of an ANN, whose performance parameter is preliminarily used to solve the sampling design for the WDN. Most of the so-derived monitoring nodes coincide in the three cases. The performance of each ANN appears to be strongly influenced by cross-correlation values, with the best results provided by the ANN relating to the most correlated demands.

Two Stochastic Methods to Model Initial Geometrical Imperfections of Steel Frame Structures

“…The representation of the specified input distributions and the input correlations is also very accurate when the standard Latin Hypercube Sampling (LHS) method [43] is used, whereby a method to minimize the undesired correlations is implemented (Iman and Conover) [44]. Furthermore, it is easier to work with sampling methods when there is a correlation between parameters, e.g., [45][46][47]. ALHS has been preferred, as it is recommended for not so large numbers of input parameters [48].…”

Physics-Informed Graph Neural Networks for Water Distribution Systems

“…Curently, ML technologies have been proposed for specific tasks in WDS such as leakage detection (Fan, Zhang, and Yu 2021), modeling virtual sensors (Magini et al 2023), or demand prediction (Wu et al 2023). To the best of our knowledge, the complex task of state estimation has only been dealt with by (Xing and Sela 2022), addressing model hydraulics using GNNs.…”