Abstract

The economic, social and environmental impacts of water main failures impose a great pressure on utility managers and municipalities to develop reliable rehabilitation/replacement plans. The Canadian Infrastructure Report Card 2012 stated that 15.4% of Canadian water distribution systems was in a “fair” to “very poor” condition with a replacement cost of CAD 25.9 billion. The “fair”, “poor” and “very poor” conditions represent the beginning of deterioration, nearing the end of useful life and no residual life expectancy, respectively. The majority of municipalities in Canada do not possess complete dataset of water distribution networks. The annual number of breaks or breakage rate of each pipe segment is known as one of the most important criteria in condition assessment of water pipelines. The main objective of this research is to develop a research framework that circumvent the limitations of existing studies by: 1) identifying the most critical factors affecting water pipe failure rates, 2) determining the best mathematical expression for predicting water pipeline failure rate 3) developing deterioration curves, and 4) deploying sensitivity analysis to recognize the effect of each input change on the breakage rate.
The proposed research framework utilizes Best Subset regression to recognize the most effective factors on water pipelines. Best-Subset Algorithm is a procedure to find the best combination of variables to predict the water pipe failure rate among all possible candidates. Once the process of critical factor selection is performed, selected variables are employed to predict the number of breaks of water pipes using Evolutionary Polynomial Regression (EPR). The EPR is an intuitive data mining technique performed in two stages: 1) the search for the best model using Multi-Objective Genetic Algorithm (MOGA), and 2) the parameter estimation for the model using Least Square Method. The predicted number of breaks, computed by EPR, is utilized to develop deterioration curves by applying Weibull distribution function. Finally, sensitivity analysis is performed to: 1) recognize the effect of changing each input on the failure rate, and 2) study the relationship between the selected inputs and the output.
The developed research framework is applied into two case studies to test its effectiveness. The case considers the water distribution networks in the City of Montréal, Canada and the City of Doha, Qatar. Physical factors, such as age, length, diameter and pipe material were identified as the most critical factors to affect the failure rate of pipes. The results indicate that the developed models successfully estimated the number of breaks for the City of Montreal and City of Doha with a maximum R-Squared of 89.35% and 96.27%, respectively. Also, it is tested by using 20% of each dataset and promising results were generated with a maximum R-Squared of 84.86% and 74.39% for dataset of Montreal and Doha respectively. This demonstrates the accuracy and robustness of the developed models in assessing and analyzing water distribution networks. The developed model is useful for municipalities and decision makers to prioritize the maintenance, repair, rehabilitation, and budget allocation of water distribution networks.