Application of K-nearest neighbours method for water pipes failure frequency assessment

Abstract

The paper describes the results of failure rate modeling using K-nearest neighbours method (KNN). This algorithm is one among other regression methods, called machine learning methods. The aim of the presented paper was to check the possibilities of application of such kind of modelling and the comparison between current results and investigations of failure rate prediction in another Polish city. Operational data from 12 years of exploitation, received from water utility, were used to predict dependent variable (failure rate). Data (249 and 294 for distribution pipes and house connections, respectively) from the time span 2001–2012 were used for creating the KNN models. On the basis of other data (one case for each year) the validation of optimal model, based on Euclidean distance metric with the number of nearest neighbours K = 2, was carried out. The realization of the modelling was performed in the software program Statistica 12.0.