Leakage Prediction in Real-World Water Distribution Networks Using Multi-Objective Multi-Gene Genetic Programming

Understanding leakage is an important challenge within the water sector to minimise waste, energy use and carbon emissions in water distribution networks. Leakage is usually approximated as minimum night flow for each District Metred Area (DMA). However, not all DMAs have instruments to monitor leakage directly or the main dynamic factors that contribute to it. Therefore, here the leakage is estimated by using the recorded features of its pipes, making use of readily available asset data collected routinely by water companies. The problem is interpreted as a feature construction task for multiple-instance regression and uses a multi-objective multi-gene strongly typed genetic programming approach to create a set of features that are used by a linear regression model to estimate the average leakage in DMAs. The methodology is applied to a dataset of a large supply area for a water company with 790 DMAs, and the results are compared to a previous work which used human-constructed features. A non-dominated analysis is conducted on the appearance of particular features across multiple runs, and Shapley values are used to understand the impact and importance of each feature. The results show better performance with significantly fewer and less complex features. In addition, novel features are found that were not part of the human-constructed features. This article contributes to the wide body of application-focused literature that continues to demonstrate the effectiveness and creativity of digital evolution for real-world problems.