A Probabilistic Digital Twin for Leak Localization in Water Distribution Networks Using Generative Deep Learning.

Area of Science:

• Engineering
• Computer Science
• Data Science

Background:

• Leak localization in water distribution systems is complex due to network intricacies, limited sensors, and noisy data.
• Accurate leak detection is crucial for efficient water management and infrastructure maintenance.

Purpose of the Study:

• To develop a robust methodology for leak localization in large water distribution systems.
• To incorporate uncertainty quantification into the leak localization process.
• To leverage generative deep learning and Bayesian inference for improved accuracy and speed.

Main Methods:

• A generative deep learning model using neural networks acts as a probabilistic surrogate for complex system equations.
• Bayesian inference is employed to combine sensor data with the surrogate model's output.
• The methodology quantifies uncertainty in leak location predictions.

Main Results:

• The method demonstrated fast, accurate, and trustworthy leak localization across three test cases of increasing complexity.
• Average topological distances (ATD) ranged from 0.3 to 10 depending on network complexity and noise levels.
• Achieved accuracies of 83%, 72%, and 42% for the respective test cases, with computation times from 0.1 to 13 seconds.

Conclusions:

• The proposed generative deep learning and Bayesian inference approach offers a powerful tool for leak localization in water networks.
• This methodology provides a reliable digital twin solution, integrating advanced mathematical and deep learning techniques.
• The approach effectively addresses the challenges of uncertainty and complexity in water distribution systems.