One model fits all – on the generalizability of pipe deterioration models between utilities

For proactive management of water distribution pipe networks, one pre-requisite is to predict either pipe failure probability for risk assessment on a tactical level or life expectancies for budget allocation on a strategic level or preferably both. Machine learning methods are documented to provide promising results for predicting water distribution pipe failures. Still, they rely on a considerable amount of data, which is seldom available with sufficient quality. Especially, small municipalities lack the required amount and quality of data to utilize the benefits of machine learning models. This study aims to train a deterioration model based on random survival forests (RSFs) by using datasets from several utilities in Norway and testing the model’s usability for the individual networks to assess its generalizability. The benefit of using RSFs is twofold: 1) it can model complex relationships between the variables and output, 2) it accounts for right-censored data. The method is tested on pipe failure data from nine different, small to medium-sized water utilities. Furthermore, this work highlights the possibilities and limitations of such a machine learning approach.