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

Abstract

The use of statistical models to predict pipe failures has become an important tool for proactive management of drinking water networks. This targeted review provides an overview of the evolution of existing statistical models, grouped into three categories: deterministic, probabilistic and machine learning. The main advantage of deterministic models is simplicity and relatively minimal data requirements. Deterministic models predicting failure rates for the network or large groups of pipes perform well. These models are also useful for shorter prediction intervals that describe the influences of seasonality. Probabilistic models can accommodate randomness and are useful for predicting time-to-failure, interarrival times and the probability of failure. Probability models are useful for individual pipe models. Generally, machine learning approaches describe large complex data more accurately and can improve predictions for individual pipe failure models yet is complex and requires expert knowledge. Non-parametric models are better suited to the non-linear relationships between pipe failure variables. Census data and socio-economic data require further research. Choosing the most appropriate statistical model requires careful consideration of the type of variables, prediction interval, spatial level, response type and level of inference required.