Explainable deep learning models for predicting water pipe failures.

Area of Science:

• Environmental Engineering
• Water Resource Management
• Data Science

Background:

• Water distribution network (WDN) failures cause significant environmental and economic damage.
• Existing pipe failure prediction models lack focus on leak and burst probabilities.

Purpose of the Study:

• To develop and evaluate deep learning models for predicting the probability of leaks and bursts in WDNs.
• To enhance model performance through hyperparameter optimization and data scaling.
• To provide interpretable insights into factors influencing pipe failure predictions.

Main Methods:

• Application of Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), and TabNet for failure prediction.
• Hyperparameter optimization using Bayesian Optimization (BO).
• Model interpretation using Copeland algorithm and SHapley Additive exPlanations (SHAP).

Main Results:

• Bayesian Optimization significantly improved model predictive abilities, with a 36.2% increase in TabNet's F1 score for leak prediction on standardized data.
• The Copeland algorithm identified CNN as the top-performing model for both leak and burst probability prediction.
• Pipe diameter, material, and age were identified as critical features influencing failure predictions via SHAP values.

Conclusions:

• Optimized deep learning models, particularly CNN, offer a robust approach to forecasting WDN pipe failures.
• The developed models provide actionable insights for water utilities to improve network management and mitigate failures.
• User-friendly web applications enable practical, real-time prediction of leaks and bursts.