Prediction of Pipe Performance with Ensemble Machine Learning Based Approaches

The deterioration of water mains constitutes a major problem in cities' infrastructure management since the replacement and maintenance of water mains need high financial investments. American Water Works Association estimated the total cost for pipeline system management would be more than $1.7 trillion by 2050. Therefore, it is important to assess performance of water mains to optimize the rehabilitation of pipelines. To find an approach to improve the prediction of pipe condition, this paper proposes a stacking ensemble method to reduce the drawbacks of single models and to improve the predictive results. Considering soil property data, both a single-model and an ensemble-model were built to forecast the pipe condition, which is defined as the remaining wall thickness. Compared with the prediction performance of four single models, the superiority of ensemble method was demonstrated with up to 35.7% increase of R-squared value and 13.6% decrease of the root-mean-square error. With the proposed method, reliable decision-making tools can be developed in the infrastructure management system to help manage the large and complicated deteriorating buried pipeline networks.