Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.775816
Title: Pipe failure prediction and impacts assessment in a water distribution network
Author: Kakoudakis, K.
Awarding Body: University of Exeter
Current Institution: University of Exeter
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
Water distribution networks (WDNs) aim to provide water with desirable quantity, quality and pressure to the consumers. However, in case of pipe failure, which is the cumulative effect of physical, operational and weather-related factors, the WDN might fail to meet these objectives. Rehabilitation and replacement of some components of WDNs, such as pipes, is a common practice to improve the condition of the network to provide an acceptable level of service. The overall aim of this thesis is to predict-long-term, annually and short-term-the pipe failure propensity and assess the impacts of a single pipe failure on the level of service. The long-term and annual predictions facilitate the need for effective capital investment, whereas the short-term predictions have an operational use, enabling the water utilities to adjust the daily allocation and planning of resources to accommodate possible increase in pipe failure. The proposed methodology was implemented to the cast iron (CI) pipes in a UK WDN. The long-term and annual predictions are made using a novel combination of Evolutionary Polynomial Regression (EPR) and K-means clustering. The inclusion of K-means improves the predictions' accuracy by using a set of models instead of a single model. The long-term predictive models consider physical factors, while the annual predictions also include weather-related factors. The analysis is conducted on a group level assuming that pipes with similar properties have similar breakage patterns. Soil type is another aggregation criterion since soil properties are associated with the corrosion of metallic pipes. The short-term predictions are based on a novel Artificial Neural Network (ANN) model that predicts the variations above a predefined threshold in the number of failures in the following days. The ANN model uses only existing weather data to make predictions reducing their uncertainty. The cross-validation technique is used to derive an accurate estimate of accuracy of EPR and ANN models by guaranteeing that all observations are used for both training and testing, and each observation is used for testing only once. The impact of pipe failure is assessed considering its duration, the topology of the network, the geographic location of the failed pipe and the time. The performance indicators used are the ratio of unsupplied demand and the number of customers with partial or no supply. Two scenarios are examined assuming that the failure occurs when there is a peak in either pressure or demand. The pressure-deficient conditions are simulated by introducing a sequence of artificial elements to all the demand nodes with pressure less than the required. This thesis proposes a new combination of a group-based method for deriving the failure rate and an individual-pipe method for evaluating the impacts on the level of service. Their conjunction indicates the most critical pipes. The long-term approach improves the accuracy of predictions, particularly for the groups with very low or very high failure frequency, considering diameter, age and length. The annual predictions accurately predict the fluctuation of failure frequency and its peak during the examined period. The EPR models indicate a strong direct relationship between low temperatures and failure frequency. The short-term predictions interpret the intra-year variation of failure frequency, with most failures occurring during the coldest months. The exhaustive trials led to the conclusion that the use of four consecutive days as input and the following two days as output results in the highest accuracy. The analysis of the relative significance of each input variable indicates that the variables that capture the intensity of low temperatures are the most influential. The outputs of the impact assessment indicate that the failure of most of the pipes in both scenarios (i.e. peak in pressure and demand) would have low impacts (i.e. low ratio of unsupplied demand and small number of affected nodes). This can be explained by the fact that the examined network is a large real-life network, and a single failure of a distribution pipe is likely to cause pressure-deficient conditions in a small part of it, whereas performance elsewhere is mostly satisfactory. Furthermore, the complex structure of the WDN allows them to recover from local pipe failures, exploiting the topological redundancy provided by closed loops, so that the flow could reach a given demand node through alternative paths.
Supervisor: Farmani, R. ; Butler, D. Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.775816  DOI: Not available
Keywords: Water pipe failure ; Water supply ; Evolutionary Polynomial Regression ; Artificial Neural Network ; Pipe failure prediction ; Pressure drop ; Impacts assessment
Share: