Estimation of the short-term probability of failure in water mains

In this paper, the nonhomogeneous Poisson process (NHPP), hierarchical beta process (HBP), and a new-developed Bayesian simple model (BSM) are used for short-term failure prediction with several water utility failure data sets. The prediction curve was used to investigate the relative ranking of the pipes from the three models. The curve was demonstrated to be more suitable for practical situations compared to the traditional receiver operating characteristic (ROC) curve. The expected number of failures and the effect of incorporating covariates by dividing the pipe assets into smaller pressure cohorts were also examined. Based on the data sets used, the observations from the prediction curves show that the performance of the three models are very similar in terms of pipe ranking. However, the BSM is more advantageous because of its relative simplicity. The covariate, the number of known past failures, plays a very important role when considering the relative ranking of the pipes in the network. The NHPP and HBP are recommended if the total number of failures in the network is required. Subdividing one of the data sets (Data Set A) into a smaller pressure cohort showed improvement only for the prediction by the BSM. Unlike the NHPP, time is not specifically included in the HBP and BSM. Therefore, the HBP and BSM are not suitable for long-term predictions and cannot include time-dependent covariates.