Lifecycle operational reliability assessment of water distribution networks based on artificial neural network

Abstract

In urban operation, a reliable water distribution network (WDN) is the foundation and various factors may undermine the reliability of WDNs. Due to the complexity and nonlinearity of WDNs, however, assessing the reliability of WDNs is still a challenging task.

The present study proposes a framework for addressing this issue. First, an artificial neural network (ANN) based model is established to predict the probability of pipe failure. The accuracy of the ANN is verified by a normalized confusion matrix. In addition, impacts of input factors, including physical and socioeconomic factors, on failure probability are analyzed with a Shapley Additive exPlanations (SHAP) method. Service age and population density of the district are proved to be the two most important factors while pipe material and housing area constructed are the least critical factors.

Second, a failure scenario simulation model considering leak and burst is introduced. A general hydraulic equilibrium equation is established considering the decrease of transmission capacity caused by leakage and the leak area is considered as a random variable. In terms of bursts, an algorithm based on depth-first search (DFS) is used to isolate burst pipes with least pipes impacted.

Finally, above techniques are combined into a framework for assessing lifecycle operational reliability of a real-word WDN. In this paper, the reliability of WDNs is analyzed by Monte-Carlo simulation (MCS) method. To obtain a precise result, 5000 simulations are repeated. The final nodal reliability is equal to the ratio of the qualified simulation times (nodal head greater than 20 m) to the total simulation times. The results show that (1) service age, pipe length, population density and housing area constructed have a positive effect on failure while area of district and pipe diameter have a negative one; (2) nodal reliability decreases with the increase of failure probability and pipe roughness; (3) the distance between source and node and loop configuration appears to have a comprehensive effect on the reliability of nodes; (4) the more pipes with two valves in the WDN, the less nodes influenced by a burst occurring to a distant pipe and thus the reliability is enhanced.