A comparative study of statistical and machine learning methods to infer causes of pipe breaks in water supply networks

Abstract

Water supply pipes age, deteriorate and break, which puts at risk the continuous provision of safe potable water endangering the public health in cities. Risk management methods are increasingly applied to optimise the capital investment for pipe replacement and rehabilitation, taking into account the probability and hydraulic impact of pipe breaks. As part of this process, however, historic pipe break data and statistical methods should be utilised to gather causal insights for past breaks to inform operational changes and/or capital investment decisions in order to reduce future breaks. This paper presents a comparative study of statistical and machine learning methods to carry out an exploratory causal analysis for historic pipe breaks in an operational water supply network. Regression models for count data and probabilistic models have been developed. The performance of these models was assessed and enhanced with the introduction of interactions and the inclusion of different network representations.

An evolution of statistical pipe failure models for drinking water networks: a targeted review

“…The Zero Inflation Poisson (ZIP) model handles zero-inflation by allowing different probabilities through two components; the first generates zero counts, and the second generates counts with probability, some of which may be zero. In a study by Konstantinou & Stoianov (2020), several models were compared and evaluated on a network of 374 km between 2003-2016, using 550 pipe failures. A ZIP GLM and negative binomial GLM were amongst those used but were the only models that could manage the zero-inflated dataset for individual pipes.…”

“…One assumption of the Poisson model is that the response variance is equal to the mean, yet this is seldom the case in many pipe failure data, especially for zero-inflated data, resulting in over dispersion (Asnaashari et al 2009). When this assumption is violated, the methodology does not meet the Poisson distribution assumptions (Asnaashari et al 2009;) and overestimate pipe failures with residuals showing significant error and bias (Konstantinou & Stoianov 2020), and other studies suggesting no useful results (Yamijala et al 2009). By definition, it can be suggested that predicting failure rates is problematic for individual pipes and would be best suited to models using large pipe groups or network-wide models.…”

“…Advancing into machine learning for pipe failure modelling is an obvious choice since these approaches are capable of modelling complex relationships between variables and compliment the increased data collection over the past decade. Machine learning has been shown to improve predictive accuracy compared to deterministic and probabilistic approaches (Konstantinou & Stoianov 2020), although some studies show similar accuracy (Chen et al 2019;Robles-Velasco et al 2020), and when methodologies include large groups of pipes, deterministic models show comparable accuracy, as shown by Giraldo-González & Rodríguez (2020). Machine learning models can be used for regression, classification, or as probabilistic classifiers.…”

The challenges of predicting pipe failures in clean water networks: a view from current practice

“…Where data is available, it is often framed for a particular use elsewhere in the business, meaning additional time spent on data cleaning and transformation. The involuntary omission of important variables results in variable bias, preventing the estimator from converging correctly, causing an inaccurate representation of pipe failures (Tang et al 2019;Konstantinou & Stoianov 2020). If pipe failure models are to aid decision makers effectively, water companies must prioritise data collection, ensuring enough data related to pipe failures is readily available and collected in useful formats.…”

“…development. As such, there remains a degree of scepticism of pipe failure models Konstantinou & Stoianov 2020) and the discord between failure predictions and those observed in the field, potentially hindering effective decision making.…”