An explainable machine learning approach to the prediction of pipe failure using minimum night flow

Abstract

Both minimum night flow (MNF) and pipe failures are common ways of understanding leakage within water distribution networks (WDNs). This article takes a data-driven approach and applies linear models, random forests, and neural networks to MNF and pipe failure prediction. First, models are trained to estimate the historic average MNF for over 800 real-world DMAs from the UK. Features for this problem are constructed from pipe records which detail the length, diameter, volume, age, material, and number of customer connections of each pipe. The results show that 65% of the variation in historic average MNF can be explained using these factors alone. Second, a novel method is proposed to deconstruct the models' predictions into a leakage contribution score (LCS), estimating how each individual pipe in a DMA has contributed to the MNF. In order to validate this novel approach, the LCS values are used to classify pipes based on historic pipe failure and are compared against models directly trained for this. The results show that the LCS performs well at this task, achieving an AUC of 0.71. In addition, it is shown that both LCS and directly trained models agree in many cases on an example real-world DMA.