Prediction of failure frequency of water-pipe network in the selected city

Abstract

The paper presents the modelling results of failure rate of water mains, distribution pipes and house connections in one Polish city. The prediction of failure frequency was performed using artificial neural networks. Multilayer perceptron was chosen as the most suitable for modelling purposes. Neural network architecture contained 11 input signals (sale, production, consumption and losses of water, number of water-meters, length and number of failures of water mains, distribution pipes and house connections). Three neurons (failure rates of three conduits types) were put to the output layer. One hidden layer, with hidden neurons in the range 1-22, was used. Operating data from years 2005-2011 were used for training the network. Optimal model was verified using operational data from 2012. Model MLP 11-10-3 was chosen as the best one for failure rate prediction. In this model hidden and output neurons were activated by exponential function and the learning was done using quasi-Newton approach. During the learning process the correlation (R) and determination (R2) coefficients for water mains, distribution pipes and house connections equaled to 0.9921, 0.9842; 0.8685, 0.7543 and 0.9945, 0.9891, respectively. The convergences between real and predicted values seem to be, from engineering point of view, satisfactory.