Predicting Condition of Sanitary Sewer Pipes with Gradient Boosting Tree

Abstract

Utility managers and owners have challenges when addressing appropriate intervals for inspection of gravity sanitary sewer pipelines and other underground pipeline systems. Closed-circuit television (CCTV) inspection technology is the most common method to identify aging sewer pipes requiring rehabilitation. While these inspections provide essential information on the condition of pipes, assessing all pipes in the network is expensive, and often limited to small portions of an entire sewer system. Therefore, it would be more beneficial to use predictive analytics to leverage existing inspection datasets and then forecast the condition of pipes that have not yet been inspected. The predictive capabilities of machine learning model, namely gradient boosting tree is demonstrated based on data provided by inspection reports from City of Tampa. Three main factors, physical, operational, and environmental were considered during the selection of variables in development of this model. Thirteen independent variables including pipe’s age, material, diameter, flow rate, pipe segment length, depth, slope, soil type, pH, sulfate content, water table, soil hydraulic group, and soil corrosivity were used to build the prediction model. Complications posed by imbalance between condition classes are overcome by changing the classification classes into a binary format (where pipes are in either good or critical structural condition) and then the receiver-operating characteristic (ROC) and confusion matrix were used to measure the performance of the model. The developed model showed 87% accuracy to predict condition of un-inspected sewer pipes. The results can be used by utility companies and municipalities to forecast condition of sanitary sewer pipes, schedule inspection times, and make cost-effective decisions to match budget allocations.