Assessing machine learning and statistical models for pipe failure prioritisation in water distribution networks using shared data

Abstract

The progressive ageing of water distribution networks has significantly increased the frequency of failures in recent years, underscoring the critical need for predictive models capable of anticipating these events to optimise maintenance operations. In this work, we evaluate the performance of various Machine Learning (ML) algorithms for early failure detection, incorporating a collaborative data-sharing approach among water utility operators. The use of data from heterogeneous sources requires rigorous preprocessing for its homogenisation. This study systematises this methodology to allow for generalisation. Three ML algorithms—Logistic Regression (LR), Random Forest (RF) and Artificial Neural Networks (ANN)—are applied to datasets provided by different operators. The performance of these models is compared against a state-of-the-art statistical model (LEYP) and validated using data from the same utility (for training) and other utilities to assess portability.