Data Mining Algorithms for Water Main Condition Prediction—Comparative Analysis

Abstract

Accurate prediction of water mains condition is critical for effective rehabilitation planning. Advances in machine learning techniques can improve condition predictions. This paper compares the capabilities of various data mining techniques in predicting the condition of water mains. Predictive models investigated include generalized linear model, deep learning, decision tree, random forest, XGBoost, AdaBoost, and support vector machines. Models are first constructed leveraging a portion of the City of Waterloo, Canada, database. Genetic algorithm and cross-validation are then employed to optimize the hyperparameter tuning process. Several performance metrics and statistical tests are employed to compare the performance of the developed models utilizing a new set of data not previously used. The XGBoost model yielded the most promising results, with a mean relative error of 1.29%. Water main conditions are numerically represented on a scale from 0 to 10, with 10 indicating the highest condition. Extensive sensitivity analysis is conducted to obtain deeper insights into the most critical attributes for condition prediction. The developed model may help city managers develop optimal rehabilitation and renewal plans, considering the current and expected condition of their pipe inventory.