Affiliations: [a]
School of Engineering, Faculty of Applied Science, University of British Columbia Okanagan, Kelowna, Canada
| [b]
Faculty of Management, University of British Columbia Okanagan, Kelowna, Canada
Correspondence:
[*]
Corresponding author. Zheng Liu, School of Engineering, Faculty of Applied Science, University of British Columbia Okanagan, Kelowna, Canada. E-mail: [email protected].
Abstract: American Water Works Association has estimated that, by 2050, the total cost of pipeline system management will exceed $1.7 trillion. Thus, it is important to assess the performance of water mains in order to optimize the rehabilitation process. Recently, the use of machine learning methods in pipeline condition prediction has increased. However, existing pipe performance prediction models rely solely on underlying data-generating distributions and do not accommodate different datasets. Hence, a stacking ensemble based method is proposed in this work to overcome the drawbacks of the existing models and improve the predictive power of this mode of analysis. Using soil property data, both a single-model and an ensemble-model were constructed to forecast the pipe condition, and their prediction performance was compared and contrasted. Finally, the superiority of the proposed ensemble method was verified through its lowest value in the root-mean-square error relative to the individual models. The techniques presented in this work can aid in a reliable decision making in infrastructure management of buried pipeline networks.
