On the generalizability and transferability of water pipe break models between utilities

Water distribution networks are aging, pipe break rates are increasing, and budgets are limited. These factors have put the utilities under pressure to maintain their pipe networks in an acceptable condition. Proactive management techniques are increasingly applied to maintain the pipe networks using risk-based approaches. One key part of these approaches is predicting the pipes’ future probability of failure. Machine learning (ML) techniques are providing promising results in predicting pipes’ future failures when being validated on existing pipe failure statistics. However, small utilities lack the required amount of data to train such models, and hence, are left behind in utilizing the predictive power of ML. To address this, we aim at testing the generalizability and transferability of random survival forest (RSF) models between utilities with training data from 9 Norwegian case studies. Local utilities’ models were trained on individual utilities' datasets and their abilities to predict other utilities' pipe breaks were tested. Global models are then trained with multiple utilities’ datasets and their abilities to predict other utilities, which were not included in the training dataset, were also tested. The use of RSF is advantageous in two ways: it utilizes the power of ensemble ML and accounts for right-censorship. The results are visualized in a comparative way and indicate that the global models can predict other utilities with sufficient accuracy. The global models are also preferred over the local models in many cases. However, if a representative and large enough utility is available, it can predict the target utility’s pipe breaks as accurately as the global models.