Probabilistic forecast of failure occurrences in water networks with short maintenance records

Managers of drinking water utilities need technical decision tools for network maintenance. One of these tools, used at the Cemagref, allows for probabilistic forecasting of failures. This is a parametric proportional hazard model with covariates, assuming a Weibull distribution for the inter-arrival failure times. It has already been shown to be effective with complete maintenance historic records despite the fact that parameterisation cannot be transposed from one service to another. Up to the present, most of the services have only short maintenance records (less than 10 years). In order to assess the usefulness of the model for short records, we have processed data of the rural service of the Charente-Maritime area (Atlantic Coast of France). These networks involve pipes made of different materials: PVC, asbestos-cement, cast and ductile iron, with maintenance records available since 1988. A difficulty could arise a priori from the lack of data over the period running from the date of laying of the main to the beginning of observations. Counting the first inter-arrival times from the starting date of observations does not handicap the model; the method accounts for right censoring of the data. Firstly, using the failure records from 1988 to 1996, we parameterise a specific model for each material (the set of relevant covariates varies according to the material). Secondly, we forecast the number of failures, which may occur in the year 1997, using a Monte-Carlo method. We are thus able to compute the differences between actual and forecasted numbers of failures, and then assess their statistical significance by a Chi-square test. This has lead to encouraging results. The model is effective even for networks with short maintenance records, and therefore should be integrated in Geographical Information Systems.