Associate Professor, Centre for Water Systems, University of Exeter, Exeter, UK
PhD student, Centre for Water Systems, University of Exeter, Exeter, UK
Senior Lecturer, University of West London, London, UK
Professor, Centre for Water Systems, University of Exeter, Exeter, UK
Due to high economic, environmental and social costs resulting from pipe bursts in water distribution systems, development of a reliable and accurate prediction model to assess susceptibility of a pipe to failure is of paramount importance. This paper aims to consider the impact of both static and dynamic factors on pipe failure for long and mid-term predications. Length, diameter and age of pipes are the static and weather is the dynamic factors for the prediction model. To improve the performance of the pipe failure prediction models, the K-means clustering approach is considered. Evolutionary Polynomial Regression (EPR) is used as the pipe failure prediction model. To prepare the database for the prediction model, homogenous groups of pipes are created by aggregating individual pipes using their attributes of age, diameter and soil type. The created groups were divided into training and test datasets using the cross-validation technique. The K-means clustering approach is employed to partition the training data into a number of clusters with similar features based on diameter and age of the pipe groups. An EPR model is developed and calibrated for each data cluster. To predict pipe failures for new (unseen) data, the most suitable cluster is identified and the relevant EPR model is used to obtain the most accurate prediction. The proposed approach is demonstrated by application to a water distribution system in the UK. Comparison of the results shows that the cluster-based prediction model is able to significantly reduce the prediction error of pipe failures. Temperature-related factor is identified as the main dynamic factor influencing the t mid-term prediction of pipe failures. An EPR model is employed to predict the annual variation in the number of failures. Mid-term and long-term prediction models are developed to present the relationship between number of pipe failures and temperature-related factors for better operation and long term for capital investment respectively.