Proactively managing drinking water distribution networks: A data- driven, statistical modelling approach to predict the risk of pipe failure.

Water distribution networks are critical infrastructures, providing clean water to millions of people. 3 billion litres of water are lost through pipe failure every day in the UK, impacting serviceability. Statistical pipe failure models can reduce pipe failures by providing valuable insights to enhance decision-making and promote proactive management. This research aims to understand the complexity of pipe failures in water distribution networks and develop a methodology for a reliable pipe failure model that identifies the risk of failure. Through an embedded case study and data-driven approach, several Objectives have been undertaken that comprise the body of research delivered through several research papers. This study offers several contributions to the immediate field of pipe failure research. Firstly, the findings investigate new factors that form the various modes and mechanisms of pipe failure, using alternative methods not commonly used in pipe failure research are used, including Generalized Additive Model and Dijkstra’s algorithm, and using data from a large UK water distribution network. Secondly, the research develops a suitable methodology for predicting annual pipe failures using an advanced machine learning method; a methodology that is easily transferrable. Thirdly, the research provides a useful means of predicting the risk of failure and visualising the results. Fourthly, the research investigates the challenges of pipe failure models using a semi-structured interview approach to review current practice. Finally, the research contributes by exploring several different data-driven methods and an embedded case study design to contribute to the broader context of pipe failure modelling. The approach presented in this research provides a methodological framework to enhance decision-making for asset management of pipes in clean water networks. Furthermore, it highlights the main limitations, particularly data quality and quantity, data-pre-processing, and model development, highlighting areas for future progress.