Machine-Learning Classification of a Number of Contaminant Sources in an Urban Water Network

Area of Science:

• Environmental Engineering
• Water Resource Management
• Machine Learning Applications

Background:

• Water distribution networks face contamination risks, necessitating accurate scenario identification.
• Determining the number of contaminant injection locations is crucial for effective response strategies.
• Existing methods often have limitations based on the number of injection points.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting the number of contaminant injection locations.
• To differentiate between single and multiple injection events in water distribution systems.
• To assess the impact of sensor network design and data uncertainty on prediction accuracy.

Main Methods:

• Utilized Neural Network and Random Forest classification algorithms.
• Trained models using simulated contamination scenarios with varying injection parameters (time, duration, concentration, number of locations).
• Tested the approach on medium (Net3) and large (Richmond) benchmark water networks.

Main Results:

• Achieved excellent accuracy in classifying single versus multiple contaminant injection locations.
• Demonstrated good accuracy in predicting the exact number of injection locations (1 to 4).
• Investigated the influence of sensor layouts, demand uncertainty, and fuzzy sensors on model performance.

Conclusions:

• Machine learning models, specifically Neural Networks and Random Forests, are effective for identifying the number of contaminant sources in water networks.
• The proposed approach enhances the ability to manage contamination events by providing critical information on injection locations.
• Robustness of the prediction models is influenced by factors such as sensor placement and data quality.