Explainable ensemble models for predicting wall thickness loss of water pipes

Abstract

Water Distribution Networks (WDNs) are susceptible to pipe failures with significant consequences. Predicting wall-thickness loss in pipes is vital for proactive maintenance and asset management. This study develops optimized, explainable machine learning models for this purpose. Data from four WDNs located in Canada and the USA are collected and preprocessed. Decision Tree, Random Forest (RF), XGBoost, LightGBM, and CatBoost are employed, with optimized hyperparameters via Tree-Structured Parzen Estimator. The proposed framework performance is assessed using dissimilarity-based and similarity-based metrics. Hyperparameter optimization substantially enhances predictive performance such that the mean absolute error of RF improved by 20.51%. Based on the evaluation metrics, the Copeland algorithm was employed to rank the models, and CatBoost emerged as the best-performing model with a Copeland score of 4, followed by XGBoost and RF. The Taylor Diagram offers a visual representation of the linear proportionality between observed and predicted values across various models, with CatBoost and XGBoost showing strong alignment. SHAP analysis identifies age, diameter, and length as key contributors. The optimized models proactively identify potential pipe failures, enhancing maintenance and WDN management.

Keywords

• Water pipelines
• Wall thickness
• Machine learning
• Ensemble learning
• SHAP