M. Kutyłowska
Wroclaw University of Science and Technology, Faculty of Environmental Engineering, Wrocław, PL

REGRESSION TREES AS A TOOL FOR FORECASTING OF FAILURE RATE


This paper presents the possibility of applying regression trees (RT) to predict the failure rate of water pipes. Generally speaking, a regression tree (RT) is a directed graph having a root and nodes (leaves), in which the conditions regarding variables are checked, as well as branches comprising decision rules. An analysis using a tree building algorithm consists in finding a set of logical division conditions, and relations between the predictors (independent variables) and the dependent variable, which leads to prediction results. The failure rate of distribution pipes and house connections was predicted on the basis of operational data for the years 2008-2014 in one selected zone in medium sized Polish city. Independent variables were such parameters as length of conduits and number of damages registered in each year at distribution pipes and house connections. Models for failure rate forecasting of distribution pipes and house connections were created separately. The calculations were carried out using the Statistica 13.1 software. RT models for house connections and distribution pipes have one divided node and two end nodes. In both models number of damages was the variable with the highest significance equalled to 1. The value of resubstitution cost amounted to 0.0056 and 0.00073 for models describing house connections and distribution pipes, respectively. The results of analysis and forecasting investigations show that regression tress are relatively good tool for failure rate prediction, using even such basic predictors as length and number of damages.

Keywords: regression methods, water pipes, prediction, failure intensity