Geoadditive Bayesian regression models for water mains failure rate prediction

Abstract

Application of pure linear deterioration models for Water Distribution Networks (WDNs) is not effective in the representation of the physical degradation of water pipes because of the theoretical approach of water pipes deterioration or simply the uncertainty related to the specific form of effects that a covariate has on the response variable. Polynomial approaches are convenient to represent the complexity of the physical phenomena. However, even high degree polynomials wiggly estimate the relationships and are unsatisfactory in some regions where they fail to fit the observed data. Flexible regression techniques that enable automatic data-driven estimation of nonlinear relations between covariates and response constitute an alternative approach that is able to represent the physical deterioration process. In this study, a Geoadditive Bayesian regression model with smooth nonlinear splines functions for the continuous covariates and spatially distributed effects for the geospatial information of the pipes is applied to predict the failure rate of metallic water mains. The results highlight nonlinear dependency between continuous covariates and the response variable. A map representing the effect of the covariates and the geospatial location of the pipes on the response variable is produced. This map can be used as an early indicator to localize areas where the effect of covariates on the failure rate is high and prioritize them for inspections and maintenance.